A methodological research on software engineering applied to design of smart grids using a complex system approach

A methodological research on software engineering applied to the design of Smart Grids using a Complex System approach Author: Jos´ e´ Evora-G´ omez Advisor: Francisco Mario Hern´ andez-Tejera Advisor: Jos´ e Juan Hern´ andez-Cabrera Text printed in Las Palmas de Gran Canaria First edition, October 2014 Dedico esta tesis a mis seres m´ as queridos. A mi madre, mi padre y mi hermano Chenko por apoyarme siempre. A Paula por darme ´ animos, escucharme y much´ ısimo m´ as. A mi abuela Mamen y mi t´ ıa Margarita que siempre se han preocupado por mi. CONTENTS 9 Simulation results analysis 101 9.1 Business Intelligence methodologies for analysing data . . . . . . 103 9.2 On-Line Analytical Processing for Smart Grids . . . . . . . . . . 106 10 Demand Side Management policy design 115 10.1 Swarming Intelligence techniques for designing policies . . . . . 116 10.2 Particle Swarm Optimisation . . . . . . . . . . . . . . . . . . . . 116 10.3 Multi Objective Optimisation . . . . . . . . . . . . . . . . . . . . 117 10.4 Multi Objective Particle Swarm Optimisation . . . . . . . . . . . 119 IV Experimentation 121 11 Experimentation considerations 123 11.1 Modelling in Tafat . . . . . . . . . . . . . . . . . . . . . . . . . 124 11.2 Experiments comparing different simulation timings . . . . . . . 132 12 Agent-Based Modelling of Electrical Load at Household Level 137 12.1Description ............................. 138 12.2CaseStudy ............................. 141 12.3Results................................ 142 12.4Discussion.............................. 146 13 Modelling lifestyle aspects influencing the residential load-curve 149 13.1Description ............................. 150 13.2Casestudy.............................. 151 13.3Results................................ 154 13.4Discussion.............................. 158 14 A Multi-Objective Particle Swarm Optimisation method for Direct Load Control in Smart Grid 161 14.1Description ............................. 162 14.2CaseStudy ............................. 169 14.3Results................................ 173 14.4Discussion.............................. 178 III CONTENTS 15 A large-scale electrical grid simulation for massive integration of distributed photovoltaic energy sources 181 15.1Description ............................. 182 15.2Casestudy.............................. 184 15.3 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . 185 15.4Discussion.............................. 186 16 Criticality in complex sociotechnical systems, an empirical approach 187 16.1Description ............................. 188 16.2Casestudy.............................. 191 16.3Results................................ 193 16.4Discussion.............................. 203 17 Vehicle to Grid 205 17.1Description ............................. 205 17.2Casestudy.............................. 210 17.3Results................................ 215 17.4Discussion.............................. 218 18 Frequency Management with Swarm Intelligence: a case study in Smart Grids 219 18.1Description ............................. 219 18.2Casestudy.............................. 220 18.3Results................................ 222 18.4Discussion.............................. 226 19 Agent-based modelling for designing an EV charging distribution systems: a case study in Salvador of Bahia 227 19.1Description ............................. 227 19.2Casestudy.............................. 229 19.3Results................................ 230 19.4Discussion.............................. 233 IV CONTENTS V Conclusions 235 20 Results 237 20.1 About Smart Grid modelling . . . . . . . . . . . . . . . . . . . . 239 20.2 About data analysis . . . . . . . . . . . . . . . . . . . . . . . . . 240 20.3 About strategy design . . . . . . . . . . . . . . . . . . . . . . . . 241 20.4 Empirical hypotheses validation . . . . . . . . . . . . . . . . . . 241 20.5Transference............................. 243 20.6Projects ............................... 244 20.7Publications............................. 244 21 Discussion 249 21.1Contributions ............................ 249 21.2Otherremarks............................ 256 21.3Futurework............................. 259 Bibliography 265 V List of Figures 1.1 DSM policy design method . . . . . . . . . . . . . . . . . . . . . 9 1.2 Model for evaluating DSM policies . . . . . . . . . . . . . . . . . 10 1.3 Addition of layers for implementing the policy . . . . . . . . . . 11 2.1 Power grids’ challenges . . . . . . . . . . . . . . . . . . . . . . . 18 3.1 Power consumption variability . . . . . . . . . . . . . . . . . . . 29 3.2 Demand modification with DSM . . . . . . . . . . . . . . . . . . 34 4.1 Complex system illustration . . . . . . . . . . . . . . . . . . . . 39 5.1 MDE abstraction levels . . . . . . . . . . . . . . . . . . . . . . . 53 5.2 MDEexample............................ 54 6.1 DSS concept evolution . . . . . . . . . . . . . . . . . . . . . . . 58 6.2 Decision making process under a DSS environment . . . . . . . . 59 6.3 BIdataframework ......................... 62 8.1 Abstraction levels for modelling power grids . . . . . . . . . . . . 76 8.2 Tafat framework architecture . . . . . . . . . . . . . . . . . . . . 78 8.3 A Metamodel example for a power grid . . . . . . . . . . . . . . 81 8.4 Dependencies examples between objects . . . . . . . . . . . . . . 88 8.5 Modelexample ........................... 89 8.6 Synchronous vs Asynchronous simulation . . . . . . . . . . . . . 90 8.7 Datarequest............................. 91 8.8 Request gets blocked . . . . . . . . . . . . . . . . . . . . . . . . 92 VII LIST OF FIGURES 8.9 The data is delivered . . . . . . . . . . . . . . . . . . . . . . . . 92 8.10Cyclicdependency ......................... 93 8.11Messagesending .......................... 94 8.12 Message is received and applied . . . . . . . . . . . . . . . . . . 94 8.13 Simulation Time propagation . . . . . . . . . . . . . . . . . . . . 95 8.14 Simulation Time reception . . . . . . . . . . . . . . . . . . . . . 95 8.15Scallingup ............................. 96 8.16 Tafat asynchronous simulation architecture . . . . . . . . . . . . 98 8.17 Interaction example with agents . . . . . . . . . . . . . . . . . . 99 9.1 Structure to export simulation results . . . . . . . . . . . . . . . . 102 9.2 Cube for residential consumption . . . . . . . . . . . . . . . . . . 104 9.3 An OLAP Cube structure . . . . . . . . . . . . . . . . . . . . . . 104 9.4 Description of a fact . . . . . . . . . . . . . . . . . . . . . . . . . 106 9.5 Scenario composition . . . . . . . . . . . . . . . . . . . . . . . . 107 9.6 Householdcube........................... 107 9.7 Household dimension . . . . . . . . . . . . . . . . . . . . . . . . 108 9.8 TVcube............................... 108 9.9 TVdimension............................ 109 9.10Radiatorcube............................ 109 9.11 Radiator dimension . . . . . . . . . . . . . . . . . . . . . . . . . 110 9.12 Radiator data miner . . . . . . . . . . . . . . . . . . . . . . . . . 111 9.13 Information visualisation example . . . . . . . . . . . . . . . . . 112 11.1 Case study’s metamodel . . . . . . . . . . . . . . . . . . . . . . 125 11.2 Averaged curve after 100 simulation runs. . . . . . . . . . . . . . 129 11.3 Curves of 100 simulation runs. . . . . . . . . . . . . . . . . . . . 130 11.4 Curve of two simulation runs with significant differences. . . . . . 130 11.5 Curves of 100 simulations runs with fixed number of agents. . . . 131 11.6 Curves of two simulations runs with fixed number of agents. . . . 132 11.7Couplingdetails1.......................... 133 11.8Couplingdetails2.......................... 133 11.9 Times in which each entity kind finished . . . . . . . . . . . . . . 135 VIII LIST OF FIGURES 12.1 Load curve example . . . . . . . . . . . . . . . . . . . . . . . . . 139 12.2 Agent architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 140 12.3 Socio-demographic groups used in the case study . . . . . . . . . 141 12.4 Simulated load curve . . . . . . . . . . . . . . . . . . . . . . . . 143 12.5 Model construction pattern . . . . . . . . . . . . . . . . . . . . . 144 12.6 Comparison with real data . . . . . . . . . . . . . . . . . . . . . 145 13.1 Absence from home on weekdays . . . . . . . . . . . . . . . . . 155 13.2 Simulated load curve with real distributions . . . . . . . . . . . . 156 13.3 Simulated load curve for conservative well-off . . . . . . . . . . . 157 13.4 Simulated load curve for entertainment seekers . . . . . . . . . . 158 14.1DLCstructure............................ 163 14.2MOPSOparticles.......................... 165 14.3Searchspace............................. 169 14.4 Controllers and consumers . . . . . . . . . . . . . . . . . . . . . 170 14.5 Restrictions sent by the grid operator . . . . . . . . . . . . . . . . 171 14.6 Individual appliances consumption . . . . . . . . . . . . . . . . . 174 14.7 Individual appliances consumption with DLC . . . . . . . . . . . 175 14.8 Neighbourhoods consumption . . . . . . . . . . . . . . . . . . . 176 14.9 Neighbourhoods consumption with DLC . . . . . . . . . . . . . . 176 14.10Total consumption . . . . . . . . . . . . . . . . . . . . . . . . . . 177 14.11Total consumption with DLC . . . . . . . . . . . . . . . . . . . . 177 14.12Non-DLCvsDLC.......................... 177 15.1 GIS layers related of the provided data . . . . . . . . . . . . . . . 182 15.2 DLC strategy flowchart . . . . . . . . . . . . . . . . . . . . . . . 183 15.3 Power offer-demand unbalances . . . . . . . . . . . . . . . . . . 184 15.4 Non-DLC vs DLC for a building . . . . . . . . . . . . . . . . . . 185 15.5 Non-DLC vs DLC for all customers . . . . . . . . . . . . . . . . 186 16.1Stabilityregimes .......................... 192 16.2 Parameters and their classification . . . . . . . . . . . . . . . . . 194 16.3 Stability considering the refrigerator share . . . . . . . . . . . . . 196 16.4 Number of oscillation vs shares . . . . . . . . . . . . . . . . . . . 197 IX LIST OF FIGURES 16.5 Stability considering the door opening rate . . . . . . . . . . . . . 198 16.6 Stability considering the frequency threshold . . . . . . . . . . . 199 16.7 Edge of chaos considering the frequency threshold . . . . . . . . 200 16.8 Stability considering the external temperature (scenario 4-a) . . . 201 16.9 Edge of chaos considering the external temperature (scenario 4-a) 202 16.10Stability considering the external temperature (scenario 4-b) . . . 203 16.11Edge of chaos considering the external temperature (scenario 4-b) 203 17.1 Charging stations consumption for both strategies . . . . . . . . . 217 18.1 Frequency reaction facing a generation unit failure . . . . . . . . 221 18.2 Indicators that are used to evaluate policies . . . . . . . . . . . . 222 18.3 Frequency reaction to a failure with smart refrigerators . . . . . . 223 18.4 Frequency reaction for the policy #2 . . . . . . . . . . . . . . . . 223 18.5 Frequency reaction for the policy #3 . . . . . . . . . . . . . . . . 224 18.6 Best policy vs base case . . . . . . . . . . . . . . . . . . . . . . . 225 19.1 EVs consumption when they start charging as soon as they are plugged ............................... 231 19.2 Comparison of the original load curve for 2014 (grey) with the total consumption including EVs charging for 2030 (black) . . . . . . . 232 19.3 EVs consumption according to the RTP based policy which makes cheaper charging from 1am to 8am . . . . . . . . . . . . . . . . . 232 19.4 Comparison between original load curve for 2014 (grey) with the total consumption including EVs charging (black) using the RTPbasedpolicyfor2030........................ 233 20.1 Policies design problems . . . . . . . . . . . . . . . . . . . . . . 239 21.1Kuhn’scycle ............................ 250 X List of Tables 11.1 Configuration of the agents that uses the shopping centres. . . . . 127 11.2 Timing of every device at each simulation . . . . . . . . . . . . . 134 11.3 Performance comparison between synchronous and asynchronous cases. ................................ 135 13.1 Otte lifestyle groups in Stuttgart . . . . . . . . . . . . . . . . . . 152 14.1 Elements existing in the model scene. . . . . . . . . . . . . . . . 172 14.2 MOPSO parametrisation used. . . . . . . . . . . . . . . . . . . . 173 14.3 MOPSO execution times. . . . . . . . . . . . . . . . . . . . . . . 178 17.1 PSO parametrisation used. . . . . . . . . . . . . . . . . . . . . . 216 18.1 Indicators for both cases. . . . . . . . . . . . . . . . . . . . . . . 225 19.1Vehicles............................... 229 XI 1.1 The future of power grids sufficient electricity [Mas13]. This solution is effective but not efficient since these backup units must be operative even when renewable energy is being produced. On the one hand, infrastructures (buildings, power lines, transformers, generators) should replicate the renewable installed power. This infrastructure must be installed and maintained. On the other hand, backup generators keep burning fossil-fuels while renewable sources are producing energy. Therefore, renewable energy is not zero-emission, since it is indirectly producing GHG. An alternative solution involves the use of batteries or other energy storage technology. This is the case of El Hierro island where an energy storage infrastructure has been built with a bidirectional hydroelectric power plant [BC05]. The main problem with this solution is the investment cost which is very high. However, the cost of energy storage technology could decrease in the coming years and this solution could become more competitive [And09]. However, SGs could provide alternative solutions that could be more efficient. That means, the same results could be achieved with less expense. For example, when the sun is hidden by clouds, PV production may be affected, making production lower than consumption. In this case, some loads can be selectively disconnected in order to rebalance consumption and production. In another case, when clouds disappear and production is recovered, loads can be reconnected. These actions are respectively known as load shedding and load shifting [Str08] and they are central topics in this thesis. Since these policies act directly on the consumption side, they are classified as Direct Load Control within Demand Side Management (DSM) [GMR+03]. From a general point of view, DSM includes all the policies that adapt the demand to grid requirements [GMR+03, NNdGW09]. Some of DSM policies are based on making the people aware of efficient energy use. Other policies are related to the modification of prices, encouraging consumption in off-peak hours and discouraging consumption in peak hours. However, nowadays there is information technology (IT) that could automatically modify the consumption of specific loads in order to rebalance the grid [NNdGW09]. This technology is an opportunity that could benefit future grids. The main prerequisite for this solution is developing an IT network for monitoring and acting over the grid components. In this way, power grids could become 5 1. CONTEXTUALISATION smarter since they would be able to perceive the environment and act accordingly. This is a disruptive conception that introduces a new power grid paradigm, SGs, where the management of both sides, production and consumption, can be carried out. Nowadays, many institutions related to power grids are researching and designing the transition towards SGs. This research and design process is not an easy task since there are many factors that have to be defined: infrastructure, market, device, communication, strategies and concerns, among others. All of these factors need to be well defined and studied since they will be implemented in real power grids which are critical systems that work around the clock. Considering all these concerns mentioned above, the focus of this thesis is made on the research of new DSM policies. IT-based DSM policies need to be designed, evaluated and validated prior to their implementation in a real power grid infrastructure. This is necessary because they may involve risks for the stability and performance of power grids. The system could spin out of control, provoking malfunctioning and in the worst case, power failures. This could happen in the case of a production-consumption unbalance when DSM policies are modifying the consumption. Since it is necessary to study the impact of DSM policies on SGs, this research contributes to the field of computer simulations that address this kind of studies. There are specific challenges that must be faced when simulating DSM policies. 1.2 Simulating future scenarios To illustrate the need for simulations of DSM policies, some examples of DSM in SGs are presented. In all these examples, it is assumed that the grid under study has a communication network which allows the remote control of the appliances on the demand side. Let us think of a policy where the criteria to make decisions is based on the state of the power grid. When there is an unbalance between generation and demand, this policy will react. When there is more generation than demand, the policy will try to increase consumption in appliances (e.g. switching water heaters on). In the 6 1.2 Simulating future scenarios opposite case, when there is more demand than generation, the policy will act on devices to decrease the demand (e.g. switching water heaters off). In the case of a power grid unbalance in which demand is higher than generation, the application of the policy reduces the demand of all refrigerators of the grid. Since every household has a refrigerator, this action has a huge impact on the power demand causing the opposite unbalance: more generation than demand. In this situation, the application of the policy switches all the refrigerators on again, trying to balance the grid. Obviously, going back to the original position, i.e. when all refrigerators are on, unbalances the grid since demand is once again higher than generation. In this case, the application of the policy may lead to an unstable situation where refrigerators are continuously being switched on and off and the consequences could include a power failure or damage to the refrigerators [VKE+13]. An experiment of this situation has been done and more information can be found in chapter 16. Another example of policy is reducing demand when there is an unbalance by controlling the intensity of lighting. Let’s assume lighting intensity can be remotely modified by the policy applier. Then, whenever there is a grid unbalance, this policy reduces the intensity of all lighting by 30%. At first glance, this seems like a good idea and much energy demand can be reduced in the face of a power grid unbalance problem. Nevertheless, what happens if this occurs at night? Would all customers be in favour of having lights at 70% of their normal intensity? Lights may be considered as an essential power usage, especially at night. This policy does not take into account the quality of the service that is offered to customers. These examples point out the need to study DSM policies prior to their implementation. Even though the policy designs are obviously wrong, there may be many other issues (e.g. technological, social, etc.) that are not as easy to infer as these. These issues are not easy to infer due to the fact that power grids are complex systems that contain many different actors whose own self-interested decisions affect the grid [PAB12]. Therefore, these issues must be identified and taken into account before implementing policies in real grids. Simulation is a way to test these policies. Through simulation, policies can be tested on a large-scale making it possible to see effects at different levels of aggregation. Simulations for testing DSM policies may be developed by using 7 1. CONTEXTUALISATION disaggregated models since these policies, such as Direct Load Control, are devoted to acting at the lowest level of consumption (devices on the demand side). Examples of these simulations have been cited in many documentary items as in [CGLP94, PKZK08, EKM+11, RVRJ11]. When many decisions are being made in a local and distributed manner, the aggregated effects of these decisions result in an emergent behaviour that cannot be easily predicted. In general, this is a common characteristic of complex systems and the best way to study them is through simulations. These simulations may be useful for analysing the policies in SGs and refining their design [PFR09]. Complexity in SGs arises when many components are acting and small variations in their behaviour may cause the system to evolve in an unpredictable manner. The challenges that are addressed in this document are related to deal with this complexity when developing simulations to test these policies. Power grids are huge engineering constructions which consist of many different elements linked to each other. In this sense, different approaches have been researched in this document to deal with these levels of complexity. 1.3 Problem definition In the previous sections, the difficulties for simulating DSM policies have been described. Nevertheless, they are also challenges that must be faced. On the one hand, the need for disaggregating the demand in order to study DSM policies has been justified. In addition, the need for running the studies through simulation experiments has been justified as well. The design of complex systems is usually approached through a trial and error method. In [Sim91], this is expressed: “the more difficult and novel the problem, the greater is likely to be the amount of trial and error required to find a solution”. This is the case in the design of DSM policies which are both difficult and novel. This trial and error method is not completely random or blind; it is actually highly selective, since the knowledge acquired in previous experiments is used to design new ones. Following this idea, the way in which DSM policies are designed can be based on an iterative process (Figure: 1.1). This process starts by defining the objectives 8 1.3 Problem definition Figure 1.1: A trial and error method for designing DSM policies. that the application of the policy must achieve. According to these, the behaviour of the individuals that make decisions is defined. This behaviour must be oriented to achieve the objective of the policy. Thus, the power grid needs to be modelled and simulated in order to evaluate the policy. After the simulations are executed, the data they provide is analysed in order to compare it to the objectives previously defined. Depending on the results of this comparison, the iterative process may be completed or not, starting a new iteration. The execution of the stages of this trial and error method may be effort-consuming. The main difficulties that these stages have are discussed in the following paragraphs through the description of a case study carried out in this research. Some experiments developed within the context of the Millener project [EDFb, CSMJ13] were oriented to testing DSM policies on a French island called La R´ eunion [AER]. These experiments required the modelling and simulation of the island’s entire grid involving the demand side. Concretely, the experiment was oriented to simulating the effects of applying DSM policies over detached homes. To this end, every detached home has to be modelled including, at least, the appliances that can be remotely controlled. This involved a total number of elements, including their respective behavioural models, of approximately 2.5 million. Facing this complexity is not a trivial problem. The management of such complex models is an arduous task. Modelling a scenario such as the one presented requires the integration of multiple data coming from multiple sources (Figure: 1.2). 9 1. CONTEXTUALISATION Figure 1.2: Base model enabled for the evaluation of Demand Side policies. First of all, data for representing the power lines of the grid has to be integrated. Later on, a database of buildings throughout the island must be included in the model. At this point, appliances that are required for simulating the demand need to be modelled. Furthermore, the behaviour of the people living in homes must be designed and implemented in order to simulate the way people use the appliances at the residential level. This modelling task described above is just one part of the whole process since it only provides the base scenario on which DSM policies can be tested. The inclusion of a policy like this may involve the implementation of a communication network which enables the transmission of messages among different actors in the grid to command the demand side (Figure: 1.3). In this network a set of devices that are responsible for making decisions, transmitting messages, applying commands, monitoring the power grid, etc. has to be placed. This is, the addition of the elements that are necessary depends on how the policy is designed and its requirements. In this kind of simulation in which there are millions of elements involved the quantity of results provided by the simulator is huge. Every single element of 10 1.3 Problem definition Figure 1.3: Addition of a communication and a decision making layer to the base model. the simulation is providing its own variable parameters at each time step. Having 2.5 million elements that are outputting three different parameters each for one simulation day with a time step of minutes (so 1440 minutes / day), the total number of results reaches 1010. With so many results, modellers usually focus only on certain features of the simulation since it is almost impossible to check all the output. Therefore, many important conclusions that may be drawn from the results may be missed since the whole data output is normally not being reviewed. Another issue that has been observed is related to the conception of DSM policies. The root problem, and the reason why these policies are created, is that there are limited resources (energy) and many different actors that want to have access to those resources (customers). Furthermore, there is an environment which has limited resources and heterogeneous actors that have their own particular interests. Therefore, the challenge is how to efficiently distribute these resources according to certain objectives and constraints. In conclusion, three major problems have been described in this section: • SGs modelling and simulation: the complexity of dealing with a huge number of elements. • Simulation results analysis: the complexity of dealing with a huge number of results from simulations. • DSM policies design: the complexity of dealing with so many actors that want access to the limited energy resources. 11 1. CONTEXTUALISATION 1.4 Research questions Complexity is a common issue in the problems described in the previous section. The questions presented in this section address the problems that have been presented. In figure 1.1, the way in which DSM policies are designed is presented. First, after defining the policy’s objectives, the way in which decision makers will act must be defined. Second, to test the policy, it is necessary to develop “in silico” experiments. Nevertheless, carrying out these experiments is an arduous task if the power grid represented is huge. Third, these simulations can output huge quantities of data which must be analysed. Although the selective trial and error method is assumed to be the base for designing DSM policies, there is a general question that arises: how to improve the execution of this method? That is to say: how the effort to design a policy can be lessened. In other words, this is a question of efficiency and productivity. This can be addressed by making less iterations or reducing the effort to carry out each iteration. This research is focused on the reduction of the effort to carry out the iterations as this issue can be dealt with methodological and technological approaches. The problem of reducing the number of iterations has to do with defining heuristics that suggest the paths that should be tried first [Sim91]. This has not been addressed in this research, although it could be further explored in future research. Thus, this research is oriented to exploring methods and technology for improving efficiency and productivity in the design of DSM policies. To be more specific, this general question is elaborated through questions that address the problem of efficiency and productivity at each stage: definition of the individual’s behaviour; modelling and simulation of complex systems; and analysis of simulation results. Regarding the definition of the individual’s behaviour: the approach researched is based on natural computing, regardless of other ideas that could be explored. “Natural computing” methods take inspiration from nature for the development of problem-solving techniques [RBK11]. These methods have been previously 12 1.4 Research questions applied in other domains by our research group 1. For example, Swarm Intelligence methods have been used in optimisation problems. For this reason, the question “can Swarm Intelligence help to deal with the complexity of designing DSM policies?” has been used to study these specific methods in the field of DSM. In addition, this question is also related to the emergent behaviour of the power grid as the idea consists in using Swarm Intelligence to obtain the power grid objectives of consumption. In other words, Swarm Intelligence techniques could be used to obtain a desired emergent behaviour by acting on the individuals. Therefore, another question could be: “can Swarm Intelligence help obtain a desired emergent behaviour in a SGs?” There may be many different techniques to adress this issue. However, apart from the fact that we have used these techniques in the past, we have observed that other researchers have used them in the power grids field (Section: 4.4). Concerning the modelling and simulation of complex systems, several questions can be defined. The most general one would be, how can the complexity be addressed? However, this question is too general and it must be more specific. A common strategy for solving a problem is using methods. So, a different question could be: “which method would be appropriate for defining complex models?” This question assumes as an axiom that it is better to apply a method rather than nothing. Nevertheless, this question focuses on searching for a method. Since this research has been developed by a group 1with research lines in Model Driven Engineering and Business Intelligence methodologies, it is logical to explore how these methodologies could be applied to the problem of complexity in SGs. Therefore, the above question can be precisely expressed: “could Model Driven Engineering help to deal with the complexity of modelling and simulating SGs?” and “could Business Intelligence methodology help to deal with the complexity of the data analysis from simulations?” The nature of these questions defines an exploratory research, where these methodologies can be adapted to this problem. It is not possible to explore how these methodologies are applied to all the problems related to complexity in SGs. 1CES (Calidad, Eficiencia y Sostenibilidad - Quality, Efficiency and Sustainability) division of SIANI (Sistemas Inteligentes y Aplicaciones Num´ ericas en la Ingenier´ ıa - Intelligent Systems and Numerical Applications in the Engineering) institute, University of Las Palmas de Gran Canaria. 13 1. CONTEXTUALISATION However, it is possible to explore them in certain cases. The goal that has been defined for this research is to provide evidence for or against these methodologies instead of verifying them. Furthermore, the problem cannot be expressed as the discovery of the best methodology for dealing with complexity. Such questions are not worthwhile because it is not possible to define suitable experimentation. 1.5 Structure of the document The rest of this document is organised as follows: after this part that was intended to contextualise the research, the “state of the art” of topics related to this thesis are reviewed: SGs, DSM, modelling and simulation, model driven engineering and data analysis. This part aims to expand upon the information provided in the introduction. It also shows how other researchers are carrying out the DSM experimentation. Furthermore, it describes modelling approaches, applications of model driven engineering and technology for data analysis. The third part presents the hypotheses. These hypotheses are related to exploring the questions previously defined. Therefore, this part is devoted to describing the hypotheses and their development in terms of how the research process has been addressed and the description of the tools that have been developed as a consequence of the research. The fourth part is devoted to validating the research process described in the third part by applying it to case studies. In this part of the document, two synthetic case studies are presented to show features related to the modelling and simulation of SGs. The following case studies described here are based on real data. They are sorted from case studies in which only the demand is simulated to those in which there are policies that act over the demand. The last part of the document, the conclusion, shows the results and discussions of this research. This is to say, in the “results” chapter, the outcomes of this research are described in terms of what has been done, frameworks, projects, papers, etc. The “discussion” chapter provides some interesting reflections and future considerations related to this research. 14 2.3 Scope Technology Platform and the other from the USA Department of Energy. They are cited below: In [tpftenotf], the European Technology Platform for Smart Grids stated that SGs are “electricity networks that can intelligently integrate the behaviour and actions of all users connected to it – generators, consumers and those that do both – in order to efficiently deliver sustainable, economic and secure electricity supplies”. In [Dep], the Department of Energy of US points out that SGs must include the following features: “self-healing from power disturbance events; enabling active participation by consumers in demand response; operating resiliently against physical and cyber attack; providing power quality for 21st century needs; accommodating all generation and storage options; enabling new products, services, and markets; optimizing assets and operating efficiently”. As it can be seen, the Department of Energy is more precise when defining the aims assigned to a SG, highlighting the importance of addressing safety issues [Fre08]. 2.3 Scope SGs were initially conceived to address the improvement of the DSM, energy efficiency and safety through the construction of grids that are robust against sabotage and natural disasters [RPT07]. However, new requirements have expanded the initial scope of the SGs to something broader which includes the creation of frameworks to achieve the interoperability of all the actors within the SG [FMXY12]. An interesting way to analyse what SGs are supposed to address is through the proposal of frameworks. These frameworks are usually developed by institutions and one of the main points is the definition of the scope of SGs. The next paragraphs will summarise the scope that some relevant frameworks have defined for SGs. a) NIST Framework proposal. According to a report from NIST (National Institute of Standards and Technology of US) [JG12] developed in 2010, SGs scope must be focused on: • Improving power reliability and quality 21 2. SMART GRIDS • Optimizing facility utilization and averting construction of back-up plants • Enhancing capacity and efficiency of existing electric power networks • Improving resilience to disruption • Enabling predictive maintenance and self-healing responses to disturbances • Facilitating expanded deployment of RES • Accommodating distributed power sources • Automating maintenance and operation • Reducing GHG emissions by enabling EVs and new power sources • Oil usage by reducing the need for inefficient generation in demand peaks • Presenting opportunities to improve grid security • Enabling transition to plug-in EVs and new energy storage options • Increasing consumer choice • Enabling new products, services, and markets. b) ETP framework proposal. The European Technology Platform (ETP) considers that SGs have been conceived to meet the challenges and opportunities of the 21st century [C+06]. The use of revolutionary new technologies, products and services is considered as the main pillar for SGs. In particular, the SGs scope must be oriented to reduce peaks and waste, encourage manufacturers to develop more energy-efficient appliances and sense and prevent blackouts by isolating disturbances on the grid. c) EEGI framework proposal. The European Electricity Grid Initiative (EEGI) considers that SGs scope must include increasing hosting capacity for renewable and distributed generation, integration of national networks into market-based networks, active participation of users in markets and energy efficiency, and opening business opportunities and markets [EE10] including the standardisation and interoperability. d) IEA DSM Task XVII framework proposal. The International Energy Agency DSM task XVII considers that SGs must be oriented to face integration of DSM, Distributed Generation, RES, energy storages [Int09]. This integration is considered as important since Distributed Generation, distributed energy storages and Demand Response can be seen as distributed energy resources [HHM11] that may help to integrate intermittent RES. 22 2.4 Challenges Integrating all views and definitions from the literature that has been reviewed, SGs can be in extenso defined by listing their main observed features: • SGs are devoted to improve energy efficiency, reliability and security. • SGs will incorporate capacities for monitoring and controlling devices providing flexibility. • SGs will redesign the way in which energy markets are working nowadays, including the participation of many new actors in the decision making process over the grid. 2.4 Challenges In the literature review, different strategies that SGs may be addressing in the future have been shown. This last section of the review is intended to summarise SG strategies following the literature review made in [XABO+14]. a) Smart meters and demand flexibility. Smart meters appear to be a genuinely cross cutting component of SGs, although they are not universally perceived as necessary for a SG [Eur10]. This technology is expected to help reduce demand based on better information and by shifting load consumption to off-peak times [LC11]. Nowadays, there is a massive introduction of thermal loads in the power grid such as heating systems, air conditioning systems (HVAC) and refrigerators. In a near future the massive introduction of Electrical Vehicles (EVs) is predicted. Both thermal loads and EVs may become a major driver for SGs [XABO+14]. This is due to the fact that both may help satisfy demand peaks by either injecting stored energy or stopping consumption [CMG11]. However, their introduction in the grid involves a considerable increment in the overall energy consumption. In this situation, the challenge is divided into two sub-challenges: improving the capacity of power grids to support their introduction and adding mechanisms to control these kinds of devices to support demand peaks. The second one is related to a concept that is further developed in chapter 3 called DSM. b) Security of supply. As mentioned before, this is an important topic since the introduction of RES may involve security of supply weaknesses due to their 23 2. SMART GRIDS intermittent nature. As said in [Spe10], we are starting to become aware of the threat of supply disruption which may frame the development of SG around energy security. c) Cyber security, privacy, and control. The introduction of an infrastructure to control and monitor makes it possible to capture data that may be sensitive. For this reason, data security becomes an important concern both from a data governance and a cyber-security point of view [Hor11]. The first one concerns the privileges to access this data, whereas the second one concerns how that data must be handled to prevent if from being accessed by intruders [TM11]. The introduction of Smart Meters makes this issue even more urgent and must be approached by SG strategies. d) System defragmentation. The non-coordinated operation of companies that operate in the energy sector involves several different standard technologies and protocols. This leads to different business models that require to be merged in order to overcome such differences [Jac11]. At the same time, this makes a transition to a decentralised system more difficult, which is core for SGs [XABO+14]. This fragmentation problem must be addressed by SGs. e) Microgeneration and decentralization. Micro-generation is generally very important for low carbon electricity systems, e.g. to alleviate system congestion [BMW10]. Microgeneration offers benefits for reducing the demand, especially in a wider decentralised context. Individuals would be responsible for their energy production and consumption and, therefore, would be aware of how they have to use energy in order to optimise both variables [DW07]. However, apart from the high costs these technology have, there are two main factors which are preventing the development of these technologies [XABO+14]. On the one hand, Distribution Network Operators discourage these investments on innovation. On the other hand, many energy markets are based on central generation. The inclusion of these technologies in the grid would open this market to small energy sellers which, on the aggregate, may become important players of the market. An interesting idea [PRS08] to address this kind of regulatory issues is the use of Virtual Power Plants. Virtual Power Plants can emerge given the commercial and regulatory support. These plants will depend on the cooperation of those who are in control [Wol12]. 24 2.4 Challenges f) Interoperability. Previously presented frameworks are not only intended to define SGs scope, but many more things. One of the most important is the definition of interoperability. For instance, NIST has defined a SG Interoperability Panel (SGIP). This panel is oriented to provide a forum that supports stakeholder participation and representation with the aim of developing and evolving interoperability standards [JG12]. SGIP has three primary functions: • To oversee activities intended to expedite the development of interoperability and cyber-security specifications. • To provide technical guidance to facilitate the development of standards for a secure and interoperable SG, and • To specify testing and certification requirements necessary to assess the interoperability of SG-related equipment. 25 CHAPTER 3 Demand side management In the literature, there are many sources that review DSM. From all of them, [Str08] is the one that has been followed, due to its clarity, to conduct the structure of this chapter. After reviewing DSM, different approaches to simulate DSM policies are presented as this is a core concept in this thesis. Within the SG concept, monitoring and controlling the demand is one of the key aspects to improve power system efficiency. Then, the approach of new power grids may consist in, besides acting on the production, acting on the demand too. Power grids may act upon both the production and demand side, therefore they both become flexible grid elements. The objectives of DSM are, among others, the minimisation of peak demand and the improvement at the system operation and planning level [GMR+03]. 3.1 Opportunities for Demand Side Management There are different sectors where opportunities for the DSM can be found: generation, transmission, distribution and demand. These sectors are reviewed below. 27 3. DEMAND SIDE MANAGEMENT 3.1.1 Generation Power grids are designed to support the maximum peak of demand. This peak of demand varies over time on a daily and seasonal basis. Apart from this, power grids are built with a 20% margin of generation capacity to deal with the uncertainty of generation capacity and unpredicted demand increases [Str08]. The average utilisation of the generation capacity in a year is below 55%. There is another issue arising with the introduction of intermittent renewable sources as wind or photovoltaic. These sources of energy are not predictable and this causes the introduction of uncertainty in the power grid generation [BI04]. Nowadays, this is dealt by using backup generation units that are available to substitute intermittent renewable generation in case it is needed. This requires a high investment [Mas13]. This is an opportunity for DSM to apply load shifting (move some usages of energy) from peak to off-peak periods [GMR+03]. Load shifting can help, on the one hand, to reduce the required generation capacity since peaks are smaller and, on the other hand, improve the efficiency by increasing the average utilisation of generation capacity. Concerning the introduction of intermittent renewable generation, DSM can help to absorb energy in cases in which there is a high production of intermittent renewable sources and a low consumption. 3.1.2 Transmission and distribution From the point of view of the transmission and distribution, power grids are designed to be robust in case a circuit is lost. When this happens, remaining circuits take over the load of the faulty one. However, these circuits cannot become overloaded. That is, under peak-load conditions, these circuits are usually loaded below 50% [Str08]. An opportunity emerges for the DSM in this field when power grids are designed to have a disaggregated generation. Nowadays, most of the generation capacity is centralised. However, the distribution of generation avoids the need of having low-used circuits just in case a circuit fails [Str08]. DSM may help in this task to make an active control of the distributed production according to the realtime needs of the grid. 28 3.1 Opportunities for Demand Side Management Figure 3.1: Left: the consumption of a winter weekday. Right: the consumption of a summer weekday. (Source: Red El´ ectrica de Espa˜ na). 3.1.3 Demand Demand side is largely uncontrollable in current power grids and varies with time of the day and season [Str08]. For instance, in Great Britain, the minimum consumption occurs in summer nights and is about a 30% of the winter peak. The variability of the consumption of a power grid depends on regional environmental conditions. In Spain, as in Great Britain and many other power grids, there is a high variability in the amount of energy consumed during the day and between seasons. In figure 3.1, on the left side, the consumption of a winter weekday in the Spanish mainland is presented and, on the right side, a curve of a summer weekday (Source: Red El´ ectrica de Espa˜ na [REE]). On the one hand, variability concerning the daytime can be observed since there are visible valleys and peaks. On the other hand, variability between seasons can be observed since not only the amount of energy changes but also the shape of the curve. DSM can have an opportunity to balance this unbalanced condition that exists between peaks and off-peaks [PD11]. This can be made by shifting loads from peak periods to consume when the grid is in an off-peak period. A way to address a load shifting is through the use of electric energy storages (batteries) so that they can inject energy in peak-periods and consume energy in off-peak periods. Since the massive introduction of batteries is nowadays very expensive, other energy storages that are found in the demand side can be used, such as thermal loads [PD11]. These loads can be shifted from peak-periods to off-peak periods. 29 3. DEMAND SIDE MANAGEMENT The energy consumption of these loads is normally not reduced but postponed. 3.2 Review of Demand Side Management techniques DSM is divided in several kinds of interventions at the customer level including Direct Load Control (DLC) and Demand Response (DR) [NNdGW09]. DLC is an intervention at the device level in order to shift customers’ consumption regarding grid state objectives. DR consists in modifying the customers’ energy usage from their normal consumption patterns in response to changes (e.g. price of electricity over time) [AES07]. The next subsections introduce different DSM techniques. 3.2.1 Direct load control The next sections present several techniques in which a direct-load control approach is being used. In these techniques, appliances are directly addressed by remote devices in order to modify their consumption according to certain criteria. a) Night-time heating with load switching. The idea of this technique is to set heaters to consume energy at night in order to balance the consumption of the grid. This technique requires the use of additional devices which allows to switch heating devices remotely by using radio tele-switching. This technique can fit into what is known as Direct-Load Control. b) Commercial and industrial programmes. There are some programmes that are aimed at commercial and industrial purposes. Particularly popular are loadinterruptible programmes which are oriented to provide reserve services enhancing the system reliability. Other programmes that are normally available for commercial customer consist in controlling loads by using building control systems for HVAC. These devices can be connected to power grid aggregators to command to either reduce or increase consumption according to the state of the grid. 3.2.2 Demand response The next sections present several techniques in which a demand response approach is being used. These techniques are characterized for using external stimulus, such 30 4.1 Power grid simulators TEFTS [Uni00]: “program has been designed to do transient stability and energy function analyses of reduced dynamic models of ac/dc power systems, with additional capabilities for voltage stability (bifurcation) studies based on continuation methods”. MATPOWER [ZG97]: “is a package of MATLAB M-files for solving power flow and optimal power flow problems. It is intended as a simulation tool for researchers and educators that is easy to use and modify”. Voltage Stability Toolbox (VST) [Nwa02]: “developed at the Center for Electric Power Engineering, Drexel University combines proven computational and analytical capabilities of bifurcation theory and symbolic implementation and graphical representation capabilities of MATLAB and its Toolboxes. It can be used to analyze voltage stability problem and provide intuitive information for power system planning, operation, and control”. Power System Analysis Toolbox (PSAT) [Mil05]: “is a Matlab toolbox for electric power system analysis and simulation. The main features of PSAT are: Power Flow; Continuation Power Flow; Optimal Power Flow; Small Signal Stability Analysis; Time Domain Simulation; Complete Graphical User Interface; User Defined Models; FACTS Models; Wind Turbine Models; Conversion of Data”. InterPSS (Internet technology based Power System Simulator) [Zho]: “is a free and open software development project. Simulation is key to enhancing power system design, analysis, diagnosis, and operation”. AMES Market Package [Tesa]: “is our software implementation, in Java, of the AMES Wholesale Power Market Test Bed. Our objective is the facilitation of research, teaching, and training, not commercial-grade application”. DCOPFJ [Tesb]: “is a free open-source Java solver for bid/offer-based DC optimal power flow (DC-OPF) problems suitable for research, teaching, and training applications”. OpenDSS [ADHM]: “is a simulator specifically designed to represent electric power distribution circuits. OpenDSS is designed to support most types of power distribution planning analysis associated with the interconnection of distributed generation (DG) to utility systems”. TSAT [Pow]: “is a leading-edge full time-domain simulation tool designed for comprehensive assessment of dynamic behavior of complex power systems”. 37 4. MODELLING AND SIMULATION FOR SMART GRIDS GridLAB-D [Pac]: “is a new power distribution system simulation and analysis tool that provides valuable information to users who design and operate distribution systems, and to utilities that wish to take advantage of the latest energy technologies”. Some of these simulators are able to simulate the demand side at the level of disaggregation that is necessary. However, they present some limitations that are discussed in the third part of this document. 4.2 Power grids as complex systems Power grids are composed of many elements at different levels connected to each other, at a physical level, through a network infrastructure [Kre13]. The paradigm shift in the energy sector which is characterised by new market rules, together with the introduction of renewable energies and distributed systems have increased its degree of complexity. The implementation of SG technologies may involve the introduction of a network layer which enhances the capability of communication among the different actors of the grid. SG concepts, such as DSM, distributed generation and energy efficiency usage, encourage new studies that require new approaches. Previous studies on power grids were mainly focused on how to be more efficient from the point of view of scheduling power generation units according to a demand which was considered as an aggregated load. New studies that are focused on demand side, distribution and local effects of applying demand side strategies require the use of approaches where the power grid must be represented as a complex system [PKZK08]. A complex system is comprised of a (usually large) number of (usually strongly) interacting entities, processes or agents, the understanding of which requires the development, or the use, of new scientific tools, nonlinear models, out-of equilibrium descriptions and computer simulations [Wor10] (Figure: 4.1). An interesting concern is that the complex system refers to the way in which reality is represented. That is, every real system is complex on its own. For example, a system composed by a car going through a path is a complex system. However, its behaviour can be modelled as a simple equation which considers 38 4.2 Power grids as complex systems Figure 4.1: Complex system illustration. Network of different types of entities and links (colours) as example for a complex system. some few factors. In this case, the real complex system is simplified into a equation. Nevertheless, this system can be represented as a complex system in which all main components are separately modelled and interconnected with each other. For example, the car can be modelled by its components (wheels, engine, structure, etc) and each of them will have an influence in the way the car behaves driving along the path. Models and simulators are normally developed to attend the requirements of the experiment hypotheses. Similarly, this may be applied to power grids. Initially, they were simply modelled since the manageable section of the power grid was the production side. To this end, many models and simulations of these power grids were represented by a system of equations. Nowadays, the introduction of new technologies which enhance the capability of having a disaggregated control on the grid where many different actors can influence it boost the study of power grids to a higher level of complexity [KdH09]. This way, we can understand better the way in which SG concepts can impact on power grids concerning a wide variety of ranges: production, demand, markets, distributed generation, on-line microgrids, etc. The next paragraphs are intended to provide some insights on the most important properties of complex systems in order to fully understand the interest of incorporating this approach to the simulation of power grids. These properties are heterogeneity, networks and emergence [BY97]. One of the main properties of complex systems is the heterogeneity of the elements that conform the system. This property is one of the factors that determine 39 4. MODELLING AND SIMULATION FOR SMART GRIDS how complex a system is. In this sense, referring back to the example of the car driving along a path, we can appreciate the participation of different and heterogeneous entities: four wheels, an engine, a car structure (which can be decomposed) and a path. Let’s assume that the experiment is aimed at measuring how long it takes the car to go from one point of the path to another. In this sense, these elements which have concrete behaviours will influence on the total amount of time that it will take the car to do this trajectory. Power grids are also composed of heterogeneous entities such as generators, consumers, distribution technologies, etc. Moreover, each category of them contains more heterogeneous elements inside. Each element within the power grid has an influence on the system behaviour (which is an emergent behaviour. Emergence in complex systems is further developed in this section). Networks are another important property within complex systems. This property also concerns the heterogeneity property as connections within complex systems may be heterogeneous too. This heterogeneity at the network level is mainly due to two factors: type of connectivity and kind of connections. Connectivity types in networks may be any of the list below [Kre13]: • Fully connected networks: every element of the system is connected to all other nodes of the system. • Distance based networks: elements are connected among them according to distance criteria. For instance, every element is connected to its two closest elements. • Random networks: in this kind of networks, elements are connected to each other following a random criteria which is based on a certain probability. • Scale free networks [B+09]: all system networks that cannot be modelled following any of the approaches above. For example, social networks do not work following the kind of networks exposed before. In this sense, scale free networks have been identified in order to categorise all those system networks which do not fit in the other categories. In social networks, friendship among people does not follow the patterns described above. As they follow different shapes, they are considered to be scale free networks. Under 40 4.2 Power grids as complex systems these networks, elements are connected freely to each other according to the system that is being modelled. In the same way that there is heterogeneity among components within a complex system, connections can be heterogeneous from the point of view of the relation they express. Referring back to the example of the social network, links among people within the social network may be of different nature: friendship, relationship, family, etc. In this sense, links are heterogeneous as they express different kinds of relation among entities and, therefore, they define the way in which components will interact with each other. Especially in the case of SGs, we can find many different kinds of links among components as, for instance, power lines, communication links or containment relations. The cooperation among entities through these links leads us to the last property: emergence. Emergence is an important key issue within complex systems. Complex systems often behave in unexpected ways that cannot be inferred directly from the behaviour of their components; this is known as emergent behaviour [NN12]. Emergent phenomena are a consequence of complexity. The operating mode of a system that is being analysed following a traditional approach can be predicted from the system model as this operating mode has had to be programmed [SPT06]. In the case of complex systems, emergent behaviours are not predictable through tools devoted only to the data analysis, which is the reason why simulation becomes an essential tool for the analysis. In the case of a complex system approach, every single component behaviour of the system is individually modelled according to the behaviour that each unit must have. However, the results at macro-scale are the consequence of all of these individual behaviours running and interacting through their links to each other. In order to explain the main difference, a case based on the demand side of a power grid is used. Demand side on power grids can be modelled in many different ways that are discussed in the next paragraphs. A first approach can consist in taking into account profiles of consumption on aggregated levels of the grid to be simulated (e.g. profiles modelled through an equation). Even though this equation 41 4. MODELLING AND SIMULATION FOR SMART GRIDS may provide different results with respect to the profiles of consumption, the behaviour of the system can be directly inferred. Another approach can be based on collecting data from appliances and consumers in order to fully represent the demand of a region. On the one hand, data from appliances can be used to, according to this data, model each individual appliance that is in the grid. On the other hand, data from customers can be used to model the way in which users interact with the appliance within a household. In this sense, every single component of this complex system has been modelled from their own point of view. Nevertheless, the cooperation of all of them within the system provides different results in the overall consumption of the grid. This consumption cannot be inferred directly from the way in which appliances and customers can be modelled. Then, the demand in power grids is the result of the aggregation of all the consumptions of the power grid. Should distributed generation be incorporated to the power grid, this generation can be considered as negative consumption. Taking into account that demand side policies are oriented to play a role at the highest level of detail (appliances), it is necessary to represent not only the model of each different kind of device, but also the social behaviour of the people living in the household [PKZK08]. Since people living in the household are self-interested, they are considered as agents whose acts cause an effect at the macro-scale. This way of conceiving the modelling of a power grid as a complex system in which there are intelligent agents is known as agent-based modelling (ABM). 4.3 Agent-based modelling and simulation ABM is the computational study of social agents as evolving systems of autonomous interacting agents [Jan05]. ABM is a tool that allows for the study of social systems from the point of view of an adaptive complex system. Therefore, the researcher is interested in the way in which macro phenomena are emerging as a consequence of the heterogeneous individual behaviours that are taking place at the micro level [Hol92]. This approach makes it possible to systematically test different hypotheses that are related to the attributes of the agents, their behavioural rules, their types of interactions and the way in which they affect the system. 42 4.3 Agent-based modelling and simulation An interesting discussion which concerns the usefulness of this approach for running experiments is given in [Jan05]. There are researchers that wonder why an approach based on ABM may be needed for running experiments. Is it not possible to approach the experiments based on equations? The author’s answer is that this decision may depend on the types of issue that are addressed. Many problems may be faced by using equation-based models. However, problems that concern coordination or strategy interaction where multiple agents are participating need to be addressed in a different way using ABM. One of the main issues in ABM is the possibility to represent the complex structures of social interactions. In some systems (e.g. a power grid), the macroscale properties (e.g. overall demand of energy) are sensitive to the structure of interactions among agents and social networks (e.g. customers making decisions that affect energy consumption). However, using an equation-based approach, these agents are supposed to be implicit in these equation models making it impossible to research on the sensitivities of the structure of interactions [Jan05]. The research developed on multiagent systems in Artificial Intelligence (AI) has influenced much the architecture of agents present in ABM. Multi-agent research studies the adaptive behaviour of autonomous agents in a concrete environment [Jan05]. Intelligent agents are able to act flexibly and autonomously [Woo02]. This means that agents are goal-directed (satisfying or maximizing their utility), reactive (adapting themselves to environmental changes) and capable of interacting with other agents. The use of ABM for research and management is growing rapidly in a number of fields [RLJ06]. This growth is due to the ability of these models to address problems by the theory of evolution [GRB+05] and strategies [GR05]. However, this kind of modelling is still an obstacle for researchers since it requires an intensive software development in order to be carried out. For this reason, simulation platforms have been developed to make experiments using an ABM approach. These libraries follow a framework and library paradigm, providing a framework 1 along with a library of software which implements the framework and providing simulation tools. These platforms have succeeded since they provide standardised software designs and tools enabling the simulation of different kinds of models. 1a set of standard concepts for designing and describing ABM 43 4. MODELLING AND SIMULATION FOR SMART GRIDS However, these platforms have well-known limitations. From [RLJ06], five different platforms are reviewed in the next paragraphs: NetLogo, Mason, Repast, Swarm for Objective-C, Swarm for Java. Furthermore, Anylogic and Flame are also reviewed as they are also popular solutions for agent-based simulation. a) NetLogo [TW04]. It is suggested to develop models that are compatible with its paradigm of short-term, local interaction of agents and a grid environment which are not extremely complex. Highly recommended as a tool for prototyping models that can be, later on, implemented in lower-level platforms. However, its simplified programming environment may make experienced programmers feel uncomfortable since all code must be placed in just one file (which is contrary to good practices in object-orientation programming) and the lack of a stepwise debugger. b) Mason [LCRPS04]. This is a good alternative for experienced programmers who work on models that are computationally intensive since they provide a good performance and the best execution time of the five platforms tested. Things that can be improved are its non-standardised terminology, its incompatible classes with the scheduler and its lack of a terminal window for debugging purposes. c) Objective-C Swarm [MBLA96]. This version of Swarm is stable providing a fairly complete set of tools, a clear conceptual basis and clever design, where the model can be separated from the interfaces. This tool is oriented to help the model organisation by allowing the design and implementation of modules in separated swarms, each of one owns objects and schedules of their actions. This helps to manage the complexity of the models which is interesting when dealing with high complex systems. The drawbacks of this platform are the lack of friendly development tools, lack of garbage collection, weak error handling and low availability of documentation. d) Java Swarm [MBLA96]. They provide the Swarm implementation for Java users but it contains significant drawbacks: clumsy workarounds to implement Swarm features in Java, difficulty on debugging errors that happen on Objective-C libraries and slow execution speed, among others. e) Repast [Col03]. Apart from implementing most of Swarm’s functions, they have added capabilities like reset and restart models from the graphical interface and the multi-run experiment manager. The execution speed is good when compared to other platforms. They also provide geographical and network support. 44 4.4 Swarm Intelligence However, it presents weaknesses like the difficulty to get started on it, especially for amateur developers, and poor documentation. f) Anylogic [BF04]. AnyLogic is a multi-paradigm simulator supporting ABM as well as Discrete Event modeling. It provides support to develop flowcharts, System Dynamics and stock-and-flow descriptions. AnyLogic can capture arbitrary complex logic, intelligent behaviour, spatial awareness and dynamically changing structures. AnyLogic is object oriented and based on the Java programming language. To a certain degree this ensures a compatibility and reusability of the resulting models [ZLB07]. g) Flame [HCS06]. FLAME (Flexible Large-scale Agent-based Modelling Environment) is an ABM framework which allows modellers from various disciplines like economics, biology and social sciences to easily write agent-based models and simulate them on parallel hardware architectures. The environment allows to create agent-based models that can be run in high performance computers and graphical processing units. The simulation code is generated by processing a model definition [KRH+10]. 4.4 Swarm Intelligence The Swarm intelligence (SI) concept comes from the fields of AI, Distributed Intelligence and Robotics, where one of the main challenges is coordinating several robots. This concept of SI was firstly introduced in [BW89]. The author was interested in how robots programmed with simplistic behaviours may output intelligence as a result of the emergence coming from their collective behaviour. The next paragraph presents how the author defined these systems in which several robots were present: “Systems of non-intelligent robots exhibiting collectively intelligent behaviour evident in the ability to produce unpredictably ‘specific’ ([e.g.] not in a statistical sense) ordered patterns of matter in the external environment.” [BW93] Based on these ideas, the author introduced a new concept: SI. Then, Swarm intelligent systems refer to those systems in which each individual, or agent, has been implemented with simple rules. The main characteristic of these systems 45 4. MODELLING AND SIMULATION FOR SMART GRIDS is that their emergence is not predictable. Once more, the author’s definition is presented for SI: “Basically swarm intelligent systems are unpredictable for their definition of unpredictable (which is thorough) and they produce results that are improbable so are in some way surprising or unexpected.” [BW93] Another important author in the field of SI is Eric Bonabeau. In [BM01], he explains SI by exposing main characteristics of living systems as for instance an insect colony: • Flexible: the colony can respond to internal perturbations and external challenges • Robust: tasks are completed even if some individuals fail • Decentralised: there is no central control(ler) in the colony • Self-organised: paths to solutions are emergent rather than predefined In [BM01], SI is considered a mindset rather than a technology that uses a bottom-up approach to control and optimise distributed systems. This bottom-up approach that provides an emergent behaviour is supported by the use of resilient, decentralised and self-organised techniques. In this thesis, the evaluation of this kind of intelligence is made as a way to design and perform DSM policies. It can be considered that every single load of the demand side is an ant. Then, the system has a huge amount of ants which can be controlled at any time to balance the power grid. In the literature review there are different authors that have approached the DSM policy design and implementation through SI algorithms. SI have been previously used to control different aspects in power grids. In [YKF+00], the control of reactive power and voltage is made using Particle Swarm Optimisation. In [Cao04], the power grid is balanced through a collection of local interactions using Ant Colony Optimisation. An ecosystem of intelligent, autonomous and cooperative ants is presented in [SD06]. These ants make decisions when the system is unbalanced. An improved version of the Ant Colony Optimisation is used in [SS03]. In this version, multiple colonies are used for optimising the performance of a congested network by routing energy via several alternative paths. 46 5.3 Model Driven Engineering abstraction levels Figure 5.1: Proposal of MDE abstraction levels based on four layers. In order to better understand these abstract concepts, an example related with films is explained (Figure: 5.2). In this example, the film Casablanca is in the lowest-level of abstraction (M0). In this level, there are objects of the reality, so this film, Casablanca, refers to a concrete instance of the film (e.g. a concrete DVD). This DVD is an instance of a film which is defined in M1. M1 may be devoted to represent the domain of the films. To do so, the film concept is defined there. However, this film is an instance of a concept which is coming from the level M2. In this level, a metamodel is defined in order to allow defining concepts of the film domain. Furthermore, in M2, a class called Attribute is defined. This class is oriented to provide the possibility of giving attributes in the level M1. This is the case of the attribute year in the concept film. Finally, attribute and concept are instances of classes, element that is provided from M3 allowing to describe and define the metamodel in M2. Having these four levels is powerful from the point of view of the development. M0 and M1 levels do not need justifications as they are the common levels that are supported by all object-oriented languages in which the class definition would be similar to the concepts present in M1 and the object instances are in M0. In this example, M2 contains classes that are necessary to define elements in M1. Thanks to these classes, the film concept can be defined. This level is neces53 5. MODEL DRIVEN ENGINEERING Figure 5.2: Example of the use of MDE abstraction levels (using UML notation). 54 5.4 Model Driven Engineering application cases sary not only for this reason, but also for having a formal description that will be used for generators and translators so that they can interpret M1 models. Therefore, based on this M2, generators and translators will be able to process all M1 models that are M2 compliance, generating as many software solutions as different M1 models are developed. At the same time, M2 requires a set of mechanisms to define all its elements. To this end, the metametamodel, M3, provides one mechanism which allows defining classes in the M2 level. This M3 level is normally called metametamodel but it is usually a language that allows defining classes in M2. Another example can be information systems. Under this context, M3 may provide a set of mechanisms to define classes in M2. These classes defined in M2 may describe all needed elements for building information systems. Then, M1 may describe a specific definition of an information system (e.g. information system of a specific city council) being supported by M2. M0 may represent the information system in exploitation. Therefore, many M1 models representing information systems of several organisations can be developed based on classes that are in M2. Generators and translators can build a software solution for each of these M1 models. 5.4 Model Driven Engineering application cases There are many successful MDE application cases. For instance, Motorola has been working using MDE [BLW05]. In one way or another, they have been using MDE for nearly two decades. They have found that through the coordinated and controlled introduction of MDE techniques, significant quality and productivity gains can be consistently achieved, and the issues encountered can be handled in a systematic way. Another example is the application of MDE in eBusinesses [Her09]. This research is oriented to help small organisations that have limited resources to define their organisational and technological projection. A conclusion of this research is that this approach is not only useful for this kind of organisations but also for modelling service structures of public administrations. 55 5. MODEL DRIVEN ENGINEERING MDE has also been tried in the field of mobile applications development. In this field there are new challenges that must be faced, particularly, the prediction of performance of a given design [TWDS09]. The authors have seen in MDE a promising approach to address these challenges. Using MDE it is possible for developers to quickly understand the consequences of architectural decisions. A research made in [RMMG08] evaluates the use of MDE in designing adaptive multi-agent systems. This is a topic that is really close to what this thesis is addressing. They conclude that it worked as expected for an ad-hoc case but that there is still some work to be done for generalising. In [SCF+06], the authors tackle the main problems of auto-generated user interfaces. To this end, MDE techniques are evaluated. Their conclusion states that reusing MDE technologies may be promising. 56 CHAPTER 6 Data analysis and decision making support Decision making processes take place in every company or organisation. In these institutions, many decisions that must be made are oriented to solve the problems of the domain in which they are working. Since these decisions have repercussions in each institution, they must be made with maximum accuracy so that benefits can be maximized. One of the main keys for succeeding in decision making processes consists in having as much information as possible which concern the domain of the decisions to be made. However, getting this information may frequently be impossible or too costly. For this reason, the use of strategies to gather information or to reduce the uncertainty is important in order to improve decision making processes. There are two key concept that must be defined: decision and decision making. A decision is defined as the choice of one among a number of alternatives. Decision Making refers to the whole process of making the choice [Boh03] Over time, different concepts have emerged to support decision making processes [HW05]. There are many other classifications of all these concepts that are 57 6. DATA ANALYSIS AND DECISION MAKING SUPPORT CONCEPT EVOLUTION MIS Management Information Systems DSS Decision Support Systems EIS Executive Information System DW Data Warehouse BI Business Intelligence 1960 1970 1980 1990 2000 2010 Figure 6.1: Evolution of concepts that have addressed the decision making support [HW05]. related to support decision making processes. However, checking the literature review, it can be observed that most relevant literature for each concept is mainly concentrated in the years that figure 6.1 presents. In this chapter, two of the concepts are reviewed in a high level of detail: Decision Support Systems (DSS) and Business Intelligence (BI). The first has been included because it is one of the research topics that has had more relevance in the past for decision making processes. The second one has been included too because it is the one that is nowadays in vogue. Nevertheless, the other three are briefly described in the next paragraphs: a) Management Information System [Swa74]. The main idea is to have a data bank containing all the relevant information concerning the company. To this end, each executive of the company will be equipped to remotely connect through terminal to a large scale computer that contains this data bank. b) Executive Information System[Spr80]. EIS provides a set of capabilities including report preparation, inquiry capability, a modelling language, graphic display commands, and a set of financial and statistical analysis subroutines. c) Data Warehouse[Gup97]. A data warehouse is a repository of integrated information available for querying and analysis [IK93, Wid95]. The information is stored in sets of views derived from the data retrieved from the sources. 58 6.1 Decision Support Systems Figure 6.2: Decision making process under a DSS environment. 6.1 Decision Support Systems DSS are computer technology solutions that are used to support complex decision making and problem solving [SWC+02]. Classic DSS tool design is comprised of components for (i) sophisticated database management capabilities with access to internal and external data, information, and knowledge, (ii) powerful modelling functions accessed by a model management system, and (iii) powerful, yet simple, user interface designs that enable interactive queries, reporting, and graphing functions. Much research and practical design effort has been conducted in each of these domains. The process that is usually followed for making decisions under a DSS environment is shown in figure 6.2 [SWC+02]. The main focus is made on the model development and problem analysis. Once the problem is recognised, this is described in terms that make the creation of models easier. Models are created representing alternative solutions to the problem raised. Then, one of these solutions is selected and implemented. This process is iterative and has looping backs to previous stages according to the better recognition of the problem, solutions that fail, etc. This research field has evolved for the last 50 years providing different approaches to DSS. The different approaches to DSS developed during this period can 59 6. DATA ANALYSIS AND DECISION MAKING SUPPORT be classified in one of the following groups: model-driven, data-driven, communicationdriven, document-driven and knowledge-driven [Pow07]. a) Model-driven DSS. The first approaches to DSS were model-driven. Examples of them can be seen in [MS84] and [FJ69]. Model-driven DSS is focused on the accessibility and manipulation of financial, optimisation and simulation models. This is, simple models that provide the most elementary levels of functionality. Model-driven DSS use limited data and parameters provided by decision makers in order to allow the analysis of a situation by the decision makers, not being necessary large data bases [Pow02]. IFPS (Interactive Financial Planning System) is the first commercial tool that allows to develop model-driven DSS based on financial and quantitative models. This tool, IFPS, was developed by Gerarld R. Wagner and his students at the University of Texas in the 1970s. Another DSS tool based on a model-driven approach was Expert Choice [exp]. VisiCalc [vis], a tool developed by Dan Brickling and Bob Frankston, provided the opportunity for the analysis and decision support at a reasonably low cost. This tool was the first killer application for personal computers and made possible the development of many model-oriented, personal DSS for managers to use. b) Data-driven DSS. Data-driven DSS is focused on the access and manipulation of temporal data. This data can be either internal or external from the point of view of the company and sometimes it is required to deal with real-time data. In [CCS93], the authors consider that Data-Driven DSS with On-Line Analytical Processing (OLAP) provide the highest level of functionality and decision support to analyse large collections of data. In [Nyl99], the development of BI related to Procter & Gamble’s effort was considered in 1985. They built a DSS that linked information of sales and retail scanner data. BI became popular as a term that was coined and promoted by Howard Dresner of the Gartner Group in 1989. BI was described as a set of concepts and methods that are oriented to improve the decision making of businesses by using fact-based support systems. c) Communication-driven DSS. The use of the network and communication technologies in order to make easier the decision-relevant collaboration and communication is known as Communication-driven DSS. The dominant element of the architecture in these systems is the communication technologies such as groupware, video conferencing and based bulletin boards [Pow02]. Apart from these 60 6.2 Business Intelligence and On-Line Analytical Processing primary technologies, the massive expansion of the internet has made possible the increment of technologies for synchronous communication-driven DSS such as voice and video delivered through the internet. d) Document-driven DSS. A document-driven DSS provides document retrieval and analysis using a computer storage in conjunction with processing technologies. These computer storages may contain scanned documents, hypertext documents, images, sounds and videos. This material can be accessed by document-driven DSS. An example of this kind of DSS is a search engine [Pow02]. e) Knowledge-driven DSS. Knowledge-driven DSS solutions provide recommended actions to managers. These systems are person-computer systems with specialized problem-solving expertise. Such expertise consists of knowledge about a particular domain, understanding of problems within that domain and skills for solving some of these problems [Pow07]. These systems are normally implemented based on AI techniques. 6.2 Business Intelligence and On-Line Analytical Processing BI is a set of methodologies, processes, architectures and technologies that combine data gathering, data storage, and knowledge management with analytical tools to present complex internal and competitive information to planners and decision makers [EN08, Neg04]. The objective is, therefore, to improve the timeliness and quality of inputs to the decision process by making the access to the information easier. These systems are considered to be proactive and are composed by the following essential components [LV03]: • real-time datawarehousing, • data mining, • automated anomaly and exception detection, • proactive alerting with automatic recipient determination, • seamless follow-through workflow, • automatic learning and refinement, • geographic information systems 61 6. DATA ANALYSIS AND DECISION MAKING SUPPORT Figure 6.3: BI data framework. • data visualisation The way in which BI tools work is through converting data into useful information which, using human analysis, is then converted into knowledge [Neg04]. A main task consists in creating forecasts based on historical data, past and current performance, and estimates regarding future trends. Based on an analysis of the future, alternative scenarios can be designed in order to evaluate their impact (What if analysis). Another necessary task is accessing the data in order to answer specific questions. This is, that the data warehouse must be flexible enough for answering this kind of questions. Furthermore, BI tools must provide strategy insights as results of the queries that are launched on the tool. BI is devoted to assist in strategical and operational decision making. An interesting summary of this kind of decision makings is provided in [Wil02]: • Corporate performance management • Optimizing customer relations, monitoring business activity, and traditional decision support • Packaged standalone BI applications for specific operations or strategies • Management reporting of BI One of the main issues within BI is the management of semi-structured data. The tasks that are developed by the analysts in order to deal with both structured and semi-structured data is required [RC03]. Semi-structured data do not fit into relational or flat files, which is called structured data. A survey performed in [BA03] indicated that 60% of Chief Information Officer and Chief Technology Officer consider semi-structured data as critical for improving operations and creating new business models. Examples of 62 7.1 Hypothesis #1 all system. However, the way in which the overall system will behave cannot be inferred from these individual behaviours. Therefore, as said before, this system needs to be studied as a complex system. Furthermore, among all the elements, there are also agents which are characterised by being able to perceive the environment and act autonomously according to their particular goals. Thus, power grids may be represented using an ABM approach. This approach uses the same principles as complex systems, considering the inclusion of agents as an important part of the modelling and simulation process. In the state of the art, several agent-based platforms were studied in order to perform SG simulations. As a result, we have found some limitations. Initially, the lack of an explicit semantic representation in models does not allow the reuse, sharing or combination. This is due to the fact that each model has its own semantic, so different developers can produce models that cannot be reused. Therefore, working in a modular way is not facilitated by these platforms due to their lack of semantic support. Another limitation is the lack of support for developing large scale models. It is possible to build large scale models of these platforms, but the development process could resemble the complexity of the models. These platforms do not provide a methodology that supports the development process in large scale models. This is to say, we find there is a lack of architectural support for developing large-scale models. Furthermore, these platforms provide languages for describing that are isolated from the problem domain. This means a semantic gap between the concepts of platform and modellers. Ideally, modellers would be more comfortable expressing their models in a language closer to the problem domain. The approach of this thesis focuses on complexity. MDE has been explored as a methodology for developing an agent-based complex system for SGs. The application of this methodology in this field has been researched in order to represent large scenarios through models. This hypothesis is stated as follows: “MDE helps to develop models for large power grids using an agent-based approach with a high level of disaggregation and can also generate simulators according to these models.” 69 7. MANAGING COMPLEXITY 7.2 Hypothesis #2 Large-scale simulations provide a large amount of data that needs to be analysed. This analysis is necessary since problems in the modelling process can be detected and corrected, and consequences of DSM policies can be analysed. These consequences will refute or validate the ideas defined for a DSM policy and based on this, new ideas can emerge to improve its design. New approaches to data analysis and management define the problem of dealing with data as the challenge of the 3 Vs: volume, variety and velocity [Lan01]. This “3 Vs” challenge was created by Doug Laney, a Gartner analyst. “Volume” refers to the large amount of data, “variety” to the data heterogeneity and “velocity” to the need to extract relevant information from the data as fast as possible. The application of BI methodologies has been researched to address these problems of dealing with simulation data output. The use of these methodologies to analyse data coming from large-scale simulations may be helpful, from the point of view of the design of DSM policies, to identify problems in the design of the policy or “bugs” in its implementation or in the base scenario in which the policy is tested. This hypothesis is stated as follows: “BI helps to analyse large amount of heterogeneous data coming from simulations.” 7.3 Hypothesis #3 DSM policies aim to modify the way in which energy is consumed in order to improve the efficiency of power grids. Nevertheless, this modification of demand may affect the consumers by not meeting their needs. Therefore, these policies, especially those that directly act on demand appliances, must be designed to consider consumer needs in order to ensure a minimum quality of service. Thus, the environment in which policies are operating consists of many different self-interested agents. This problem can be formulated as an optimisation problem in which resources must be distributed efficiently. In this sense, policies based on optimisation methods can be designed in order to deal with real problems on the demand side such 70 7.4 Research concerns as the massive introduction of EVs into the grid or the modification of the demand side according to grid states. SI offers a promising approach to dealing with this kind of optimisation problems. In this document, the application of this approach in the design of DSM policies is explored and evaluated. This hypothesis is stated as follows: “SI helps to design DSM policies by dealing with the interests of many actors involved in the process.” 7.4 Research concerns The hypotheses presented involve the application of methodologies. Thus, the main research concern is how to validate that these methodologies are going to be helpful in solving the problems stated. It is important to distinguish the concept of “methodology” from the concept of “method”. In this section, both concepts are introduced and related to each other to clarify the semantic of the terms. On the one hand, a “methodology” in engineering can be considered as a guideline for solving a certain kind of problem. Methodologies are generally comprised of the following four elements: description of the problem that needs to be solved, definition of which techniques are to be used and when they will be used, giving advice on product quality management as well as providing tools to facilitate the process [RH97]. On the other hand, a method is a procedure that defines a regular and systematic way of accomplishing something [How12]. Generally speaking, methodologies do not describe specific methods. A method thoroughly defines the steps that have to be performed in accordance to a methodology guideline. Therefore, many methods can be based on one methodology provided they follow the methodology directives [Vac12]. In general, methodologies cannot be validated through the scientific method. This means, the certainty of the hypotheses cannot be demonstrated with experiments since it is not possible to execute the same project with two or more different methodologies in order to compare them. Moreover, this concern becomes more determinant since the execution of the project depends on many other variables: 71 7. MANAGING COMPLEXITY human knowledge, the nature of the system, budget, time, etc. Therefore, it is almost impossible to isolate the methodology variable in the success of a project. Nevertheless, this work is an experimental research with the goal of providing evidence so as to the validity of these methodologies for solving the problems of engineering SGs through case studies. “Case study research design” is proposed as a valuable and important empirical approach for validating a methodology [LR04, ZW97]. This is a useful method for validating methodologies by using them in real world situations. Generally, in experimental research, it is necessary to establish evidence for causality through internal validity [SAA+02]. However, this is not sufficient when conducting experiments using engineering methodologies. These experiments may also study the external validity of hypotheses, which ensures their application in a broader spectrum of cases rather than only in experimental situations. An experiment has internal validity if it demonstrates a causal relation between two or more variables [Bre00]. A causal inference can be based on a relation when three criteria are satisfied [SCC02]: temporal precedence (cause precedes effect); covariation (cause and effect are related) and nonspuriousness (there are no more explanations for the observed covariation). Internal validity is easy to demonstrate through the results. It is also important to demonstrate external validity. External validity is a generalisation of causal inferences in scientific studies [MJ12]. That is to say, results can be extrapolated from an experiment to other situations. At this point, it is necessary to define what can be understood by “external validity” in the context of this research. In this sense, two different concepts of external validity can be considered. In the first, it can be said that the hypotheses of this research have external validity if they can be applied to more than one experiment related to SGs. The second, it could be that the hypotheses of this research have external validity if they can be applied to the engineering of other fields apart from SGs. In this research, a set of experiments has been defined to validate the hypotheses. Some of these experiments have been jointly defined with our partners and colleagues from EIFER [eif] and EDF [edfa]. These experiments analyse real problems in real power grids. These contributions to the experimental part of this 72 7.4 Research concerns work are valuable because the validation has been made using experiments based on reality instead of only synthetic problems. Furthermore, these experiments are important because they are real cases which helps to demonstrate both internal validity and external validity. The internal validity is demonstrated through the results of the execution of experiments. Concerning external validity, the first definition presented is also demonstrated since several experiments have been conducted using the same methodologies. The second definition is not demonstrated in this research but it could be proposed as future work. 73 CHAPTER 8 Smart Grid modelling As said before, a simulator is required for each experimental study. For small experiments, the construction of a simulator is usually simple and fast. However, large experiments require the construction of a more complex simulator since different behaviours must be implemented. Furthermore, in large experiments, the initial experiment conditions are changing constantly, so it is necessary to modify the simulator in a fast and easy way [PHS+08]. As the simulator construction may be an arduous task during the experimental study, gaining productivity in this task is very important. From the point of view of Software Engineering, this simulator must be supported by a good architecture. This architecture should facilitate the simulator construction and allow the execution of continuous changing of requirements that take place during the experimental study. Since the software development of these simulations is time-consuming, it is necessary to find a way to improve development performance. In the next sections, a framework to develop simulators based on MDE approach is presented. This framework, known as Tafat [EKM+11, EHHK12, EHH13c], is intended to speed up the creation of simulators for domains in which a complex system approach and a discrete timing can be applied. Among other advantages, the use of MDE enhances 75 8. SMART GRID MODELLING Figure 8.1: Abstraction levels for modelling power grids. the capability of component reuse and hides implementation details providing a high level language to develop simulators. The adaptation of MDE to the field of complex system simulations is the base that supports the Tafat framework. The design of the abstraction applied to this case consists of three levels of abstraction. The first level, considered the lowest abstraction level, consists of the concrete elements that can be found in a specific domain. The second and third levels propose a way to abstract the definition of this concrete element. In figure 8.1 these three levels are represented for a concrete case which is related to the world of power grids. In a first level, which is related to the model description, concrete elements are located. This is, a concrete household, power line and customer. However, these concrete elements can be abstracted into a second level. The second level, known as Metamodel, the concept of a household, a power line and a customer is represented without considering specific details of a concrete instance. Based on this concept specification, the concrete elements can be defined by parametrising the elements of the second level. A third level of abstraction simplifies the complex system world into three different types of elements: entities, connections and agents. Therefore, in this case, a building is considered as a entity, a power line as a connection and a customer as an agent of the power grid system. 76 8.1 Model Driven Engineering for modelling complex systems 8.1 Model Driven Engineering for modelling complex systems Tafat allows building models of complex systems where the overall scene is decomposed into different components. Each component of the model is statically represented by means of attributes and variables. In this way, it is modeled a structural view of the system. In order to model the behavioural view of the system, each component may include several behaviours that describe how this component changes over time and the way in which it interacts with other components. That is, the dynamic of the system. This separation between structural and behavioural view is a major design concern in Tafat since it allows modeling a complex system in a modular approach. At the same time, this design technique is oriented to face the challenge of building large scale simulations. Basically, when a simulation is running, all the behaviours are concurrently executing and modifying the variables of their components. 8.1.1 Architecture The core component of Tafat architecture is the Metamodel (Figure 8.2) which describes the types of instances that could exist in the complex system which shall be represented. The Metamodel is oriented to a specific domain and thus establishes a common semantic which allows modelers to share simulations and models. The Metamodel allows a representation in terms of entities, connections, agents, attributes, variables and context, among others. The Metamodel can be translated into several formats: HTML, XSD and Java. The first one, HTML, allows the observation of the Metamodel elements, as well as their properties, in an easy and comfortable way by using a browser. The XSD translation returns a XML scheme model that helps to validate the model construction from a semantic point of view and provide valid next tokens to add when writing models. Finally, the most important one, the Java classes that implements the Metamodel elements and their features. These Java classes are used in combination with the Simulator Engine and behaviours (repository) in order to build simulators. 77 8. SMART GRID MODELLING Figure 8.2: Architectural diagram of the framework. 78 8.1 Model Driven Engineering for modelling complex systems can widely vary from one to another. Therefore, a main task for developing simulations consists in preparing data in order to use them in the model. Depending on the data format, coherence, completeness, complexity, etc. the effort to develop this task may vary. The final goal is to have the data with a format that allows it to be parsed by the Profiler. However, some small experiments may not have a complicated process for preparing the data. • Model creation. This step is divided in two main parts: creation of the simulation model and creation/modification of model elements or behaviours. The first one is mandatory for developing a simulation since in this simulation model the scenario of the experiment is described. The second one depends on the experiment requirements (e.g. a new element that is not in the Metamodel may be needed or a new behaviour, etc.). Based on the data formatted in the step before, the Profiler could help to generate the scenario, especially, if it is a large one. • Model simulation and calibration. The model simulation step finalises with the results gathering. However, this does not only consist of running the model in the simulator and waiting for the results, but also verifying and validating that this process is correct (known as calibration). This calibration process concerns the verification of the correct working of each simulation element by checking that the outputs they have are the expected ones. This can be approached in serveral ways: using testing frameworks, output checking by hand, etc. Furthermore, the experiment requirements must be confronted with the simulation features, so that, it is checked whether the simulation features match the expected requirements or not. • Result analysis. According to the experiment goals, the results must be evaluated in order to obtain conclusions. Sometimes, this evaluation may involve the simulation of other simulation experiments in order to check how the system works under different conditions. Even if the experiment planning is thoroughly detailed, many conditions can be interesting to be modified after watching the simulation results so that new simulation experiments may be carried out. 85 8. SMART GRID MODELLING 8.2 Simulation performance Most of the tools presented in this document execute the simulation with a synchronised approach. Synchronous simulations have the advantage of simple time management as all objects of the modelled system are running in the same time instant. It forces objects to always perform calculations, in every time step. Sometimes these calculations are unnecessary due to the fact they cannot provide new results. For example, a washing machine is usually waiting for an agent to be turned on, considering this as an event. Later on, it develops some washing cycles where the power may vary along the time. Whenever the washing machine state does not change, calculations could be avoided. In this section, it is proposed that an asynchronous simulation approach is included in Tafat which would allow objects to develop their own time as desired. They could behave both event and time-based according to their nature. Furthermore, they could use variable steps from one calculation to another. For example, in an asynchronous simulation where the power consumption of a washing machine is analysed, calculations would be done only when the washing machine state changes. The advantage with respect to a synchronous simulation is clear since in the synchronised case, calculations are done every time step. In the context of discrete event simulation the asynchronous concept has dual connotation. One of them consists in variable time-increment procedures as opposed to a “synchronous” or fixed time-increment procedures for simulation control. This connotation is related to the known concept Distributed Discrete Event Simulations [Kau87, Mis86]. For instance, Simula [Poo87], a simulation-oriented programming language, is based on this asynchrony concept where the time management is mainly event-based. This kind of asynchrony was already considered in Tafat through using different time steps for each mode of behaviour [EKM+11]. On the other hand, the asynchrony can be understood as a non-sequential processing where simulation parts may not be executed in the proper temporal order. That is to say, later parts of the simulation may be executed before previous ones [Gho84]. The last connotation is the one to which we subscribe in this document. The objective is to apply the time-management to each model element allowing them to be in different time instants. 86 8.2 Simulation performance 8.2.1 Tafat asynchronous simulation This section examines a new approach to achieve asynchronous simulations with Tafat. This section introduces the concepts and constructions that Tafat architecture includes to model power grids. Theses constructions are focused on dependencies between objects that are massive and very relevant in a complex system simulation. In order to properly handle an asynchronous simulation, it is important to understand the dynamics of coupled objects. For the sake of clarity, a traced execution of objects interaction during an asynchronous simulation is demonstrated. 8.2.2 Tafat system modelling Normally, a Tafat behaviour is coupled with other objects, both for querying their states or sending messages in order to change their states. In the Tafat model representation, defining behaviour which interacts with other objects is allowed. This representation approach consists of interfaces that should be defined in the object which could be externally accessed. In Tafat, there are two types of interfaces: 1. event interfaces that handle messages and are responsible for modifying the object internal variables as requested, and 2. data interfaces that handle queries and provide the value of requested attributes An example of these types of interfaces is shown in the figure 8.4. On the one hand, the thermal behaviour within a household has a data dependence with the temperature of the surrounding Outdoor. In this case, the Outdoor temperature data is requested by the associated object through the outdoor data interface. On the other hand, an sociological agent behaviour wants to turn on the washing machine. Then, this sociological agent must use the washing machine event interface to achieve this task. The washing machine event interface would change the washing machine mode to ``ON´´. The washing machine operational behaviour would calculate the proper power consumption based on this mode. Later on, when the cycle ends, the operational behaviour turns off the washing machine. 87 8. SMART GRID MODELLING Outdoor Household Washing Machine Operational Behaviour Temperature Behaviour Agent Activity Behaviour Activity Behaviour turn on Event Interface Data Interface get(Temperature) Figure 8.4: Dependencies examples between objects. 8.2.3 A power grid simulation case In order to consider the main issues that involve asynchronous simulation a simulation case is proposed to show how objects interact when working in different times (Figure: 8.5). The objects within this simulation case are an Outdoor, a Household, a Washing Machine and a Radiator. • The Outdoor is the object that represents environmental conditions, in this case, the temperature. The Outdoor temperature behaviour is responsible for setting the temperature which can be loaded from an external database. • The Household works as a container of the appliances of a household, a Washing Machine and Radiator in this case. The Household Behaviour is concerned with the thermal dynamics inside the household. • The Electrical devices inside the Household are a Radiator and a Washing Machine. These devices are handled by an Agent. 88 8.2 Simulation performance Outdoor Household Washing Machine Operational Behaviour Thermal Behaviour Agent Activity Behaviour Temperature Behaviour turn on Event Interface Data Interface get(Temperature) Radiator Operational Behaviour Data Interface Data Interface get(Temperature) get(Power) Figure 8.5: Model composition. • Finally, the Agent represents the people living in the Household and the associated behaviour defines the actions that these people are performing. For example: a person turning on the Washing Machine. The coupling in this model is represented by the dotted lines in the figure 8.5. This coupling is always defined from behaviours to interfaces. The Agent depends on the Washing Machine to change the operation mode of this device. The Radiator depends on the Household temperature, since the heat radiation is calculated based on the gap between the Radiator reference temperature and the Household temperature. The Household has two dependencies: with the Outdoor temperature and with the Radiator power, since the Household temperature is calculated by a numerical solution of a differential equation which includes these two variables. Note that, in this case, there is a cyclic dependence between the Household and the Radiator. 89 8. SMART GRID MODELLING 8.2.4 Asynchronous simulation dynamics A system simulation requires time-management to ensure that temporal aspects are correctly represented and emulated. This temporal representation only exists during the simulation process and is referred to as “Simulation Time”. Simulation Time is represented as a timestamp, a long integer where a unit corresponds to a millisecond of real time. The time-management in a synchronous simulation is centralised while the time-management in an asynchronous simulation is distributed. That is, an asynchronous simulation involves that every object manages its time, so they could have different timestamps (Figure: 8.6). Outdoor Household Agent Washing machine Radiator Simulation Time Outdoor Household Agent Washing machine Radiator Simulation Time Simulation Time Simulation Time Simulation Time Simulation Time Figure 8.6: Synchronous vs Asynchronous simulation. In this simulation paradigm, when an object is not coupled with other objects, its Simulation Time develops without considering other object Simulation Times. In this simulation case, the Outdoor is completely independent of other objects. However, when objects are coupled, the challenge consists of correctly reproducing temporal relationships. The identified temporal relationships are as follows: 1. Coupling with a data interface 2. Cyclic coupling with data interfaces 3. Coupling with an event interface 90 8.2 Simulation performance In the following sections these relationships are discussed. 8.2.4.1 Coupling with a data interface Since an object could access a variable of an external object which may be in a different time instant, every object must keep the different states that have been calculated during the simulation execution. So, when a variable is modified, a state snapshot is created in order to keep the object state in this time instant. If an object is querying for a variable value in a time instant ti, there are two cases: the object Simulation Time is delayed or ahead with respect to the external object Simulation Time. In the first case, the external object is able to provide the value by retrieving the last snapshot previous to this time instant (ti). In the second case, the dependent object must wait until the external object reaches this time instant (ti). Figure 8.7: The Household asks the Outdoor. Time is vertically represented. In the figure 8.7, the first case is shown. The Household Simulation Time is tiand the Outdoor Simulation Time is tj. Whenever tiis lesser or equal than tj, the requested data can be delivered since the data has already been calculated and stored. However, when the Household Simulation Time (ti) is greater than the Outdoor Simulation Time (tj), the Household behaviour is blocked (Figure: 8.8) until tjis greater or equal than ti(Figure: 8.9) delivering the last Outdoor Temperature value stored in the last calculated snapshot. 91 8. SMART GRID MODELLING Figure 8.8: The Household behaviour request gets blocked. Figure 8.9: The data is delivered. 8.2.4.2 Cyclic coupling with data interfaces The cyclic dependence is a concrete case of the data dependence. Two objects depending on each other whose Simulation Times are different, is handled with the following rules: the most delayed one will always retrieve the required data while the most advanced will be blocked until the delayed reaches its Simulation Time (Figure: 8.10). The mutual blocking is not possible since objects retrieve the value for the current Simulation Time to calculate the next Simulation Time value. In the example shown in the figure 8.10, the Household requires the power consumption of the Radiator in order to calculate the new temperature value. On the other hand, the Radiator behaviour needs the Household temperature value to 92 8.2 Simulation performance Figure 8.10: Radiator and Household cyclic dependence resolution. modify the Radiator state, since the reference temperature at the Radiator thermostat serves as a control mechanism. 8.2.4.3 Coupling with an event interface The event coupling means that an object receives external messages that contain orders for changing its internal variables. This is the case of objects which are managed by people that are represented as Agents in the model. The Agent interacts with these objects by sending a message using the object event interface. When the message is received by the object interface, the object Simulation Time is developed and then, a new snapshot state is created. It could happen that the Agent develops its Simulation Time without the intention of sending an order to any object. In this case, the Agent behaviour must send a ``Notification Time Message´´to the object. In fact, when the Agent Simulation Time develops, the Agent behaviour must send a Notification Time Message to all objects the Agent is controlling. This notification determines how long an object can develop its Simulation Time. This type of relationship means that object’s Simulation Time that is controlled by an Agent, will never exceed the Agent Simulation Time. Figures 8.11-8.14 shows an event relationship between a social Agent that turns on the Washing Machine. In this example, the Washing Machine Simulation Time 93 8. SMART GRID MODELLING Figure 8.11: The Agent sends a message to turn on the Washing Machine. Figure 8.12: The Washing Machine event interface changes the object mode to on. is always behind the Agent Simulation Time. In other words, the Agent Simulation Time sets a restriction for the Washing Machine Simulation Time. In the case of the Washing Machine, its power consumption would be 0 at the beginning of the simulation as it’s off. Therefore, a new snapshot is created when the Agent turns on the Washing Machine. From that moment, the Washing Machine behaviour will calculate the new power consumption with the restriction that the calculations development should not exceed the Agent Simulation Time, in case the Agent turns off the Washing Machine. 94 CHAPTER 9 Simulation results analysis As said before, simulations play a crucial role in the design of SG policies since they are a way to test them before their launch. However, the output provided by the simulations must be managed in a way that allows the policy designers to make decisions. This section explains the main concerns when analysing results obtained in a SG simulation. When facing a simulation of SGs based on a complex system approach, the results analysis becomes a difficult stage since the amount of entities is large. All systems containing a large amount of entities and relations in simulation processes provide a large amount of results. The way in which these data are normally exported is through data files. These data files are usually designed according to the data that will be managed thus avoiding the possibility of querying this data beyond what was decided to export. Therefore, whenever we deem it convenient to extract data, which was not considered to export at the design phase, a new simulation must be configured and executed. In order to exemplify this issue, a disaggregated model of a power grid system is used. This system only consists of the demand side, which is disaggregated at the device level. It is precisely at this level where we can find a layer consisting of heterogeneous elements, since the characteristics to extract from a radiator are not the same as the ones from a television (TV). If we want to preserve all variables that 101 9. SIMULATION RESULTS ANALYSIS Figure 9.1: Structure example to export simulation results. are not common to every device, it will be necessary to export each device type into a different data sheet (Figure: 9.1). At this point, once the data exportation process has been defined, we can start thinking about querying it. The list below states some query examples and how they should be dealt with according to this data exportation structure: •Querying the consumption of all devices. This query is very likely to be required. According to our data structure, firstly we calculate the total consumption at each device type. This would involve opening as many files as device types and making the calculations to obtain the total consumption per device type. Secondly, those columns which have the aggregated value at each device type must be moved into a new sheet where the final calculation would be performed obtaining the query result. The more device types there are, the trickier this process becomes. •Querying the consumption of all devices in a specific household. This process would consist in gathering the columns belonging to all the devices contained in the household from the data files. Once they are all together in a new sheet, the query result can be obtained by adding up. •Querying the consumption of all devices in a specific district. The process to obtain this query is really tricky. Firstly, all the devices belonging to a specific district must be listed. Next, all the columns which refer to the devices consumption must be gathered from the device type sheets following this list. Finally, all gathered columns can be moved to a new sheet where the query can be obtained. Taking these examples into account, it is possible to imagine how tricky the results management of more complicated queries can get. Probably, some of these 102 9.1 Business Intelligence methodologies for analysing data queries are easier to obtain by redefining the simulation results format and running it again. However, it would also be really tedious, and depending on the simulation kind, the results may differ from the previous simulation and in the end it would be necessary to start the result analysis from the beginning. All these difficulties in querying the output of a simulation could involve that many other queries are not made due to the fact that they involve a strong and time consuming effort to perform them. Unfortunately, this usually leads to focus on a small subset of variables of the simulation neglecting much information and wasting too much time in performing simple queries. The root of the problem behind the result analysis is that such results have a multi-dimensional and a multi-scale (namely temporal and spatial) nature which cannot be managed by using conventional data sheets. The example of the demand disaggregation is multi-dimensional and multi-scale. Multi-dimensional, since each dataset (for example, a power measure) is related to a specific device, location (household, building, district, etc.) and time. Multi-scale, since the information can be aggregated at different time scales (per hour, per day, per month, etc.) and at different spatial levels (device, household, etc.). 9.1 Business Intelligence methodologies for analysing data A framework, named Sumus, for applying BI methodologies has been developed. This framework for data analysis is based on OLAP. OLAP is a solution used in BI, the aim of which is to accelerate querying large amount of data. OLAP is based on cubes [CD97](Figure: 9.2), a multi-dimensional structure where data is stored. These cubes enable the insertion of data, namely facts, which are referred to several dimensions. For example, the measure of power taken from a washing machine can be referred to the device, the household where the device is and the time. Therefore, in this case, there would be three dimensions: devices, households and time. The structure of a multi-dimensional cube which addresses our problem is 103 9. SIMULATION RESULTS ANALYSIS Figure 9.2: A multi-dimensional cube for residential consumption. Figure 9.3: An OLAP Cube structure. presented in the figure 9.3. Every cube consists of dimensions, measures and indicators. The list below describes every cube component. •Dimension: it establishes a way to access the data inside the cube. Every single data is related to some elements such as when and where it happened. For example, a data of power consumption of a household would be related to the dimensions household and time. – Component: it is an element which is related to a dimension. For example, a dimension which concerns households would be filled by components which are households. 104 9.1 Business Intelligence methodologies for analysing data *Feature: it is a property of the component. In case the components are households, a possible feature could be the number of square meters there is in each household. – Taxonomy: it is a way of categorizing a dimension. There are different ways to categorize the components inside a dimension. Each of these ways is known as taxonomy. In the example of the household dimension, a taxonomy could be the size or the orientation of the facade. *Category: it is a set of components that satisfy some specific conditions. For instance, possible categories for the size taxonomy could be small, medium or big. Therefore, each of these categories would contain a set of household components the relationship of which is having a similar size. ·Rule: it establishes the condition that a component must meet in order to fall into the category that owns the rule. In the case of the small category, a possible rule could be: all the household components the feature of which number of square meters is below 80m2 •Measure: it provides a semantic to the data inserted in the cube, e.g. the power of the household mentioned above is just a number. However, the power measure is what provides the semantic to this number. A measure is usually related to a metric which enables the comparison among measures that are in different cubes. In this case, the metric of the power measure would be Watts. •Indicator: it designates the way in which a measure or a set of measures are aggregated. For example, the power measure could be aggregated using an average function. This way of aggregating measures is known as indicator. It is possible to have several indicators for one measure, i.e. the integral operator over the power measure would provide a second indicator over this measure which could be designated as energy indicator. •Fact: it relates the measures of a cube with the dimensions. A fact indicates that a certain combination of values (measures) took place for a specific combination of elements (components). In other words, a fact can be understood 105 9. SIMULATION RESULTS ANALYSIS Figure 9.4: A fact consists of context and state. as a relation of a state to a context. The state is a set of measures and the context consists of components including time. In figure 9.4, the state contains 20 (centigrades) and 135 (Watts) as measures. These measures are related to a context which indicates the time and household where those measures were taken. 9.2 On-Line Analytical Processing for Smart Grids In this section, all concepts exposed previously will be used in a practical case. Assuming that a new SG policy is to be tuned, several simulations of power grid demand will be performed. To make decisions, these simulations must focus on the power demand and the temperature at the residential sector. Therefore, the scenario for those simulations consists of several districts with households (Figure: 9.5). Each household contains several devices and calculates the internal temperature. To this end, several cubes have been designed so as to analyse the data coming from the simulation: first of all, the household cube which contains the facts regarding the temperature and, secondly, one cube per device type (TVs, Radiators and Washing Machines, among others) which contain facts about the devices. Since there are many kinds of devices in a household, in this example we are going to focus on two of them: TVs and radiators. There are two dimensions in the household (HH) cube: one measure and one indicator (Figure: 9.6). The Time dimension is common to all cubes and configures a standard way of categorizing the timeline. Household dimension contains the 106 9.2 On-Line Analytical Processing for Smart Grids Figure 9.5: Scenario composition. Figure 9.6: Household cube. households transformed into components which are described by features. The temperature of the household is the only measure that this cube is going to store and it will be aggregated using an average criteria according to the designation of the indicator. The household dimension contains a taxonomy which concerns the locations (Figure: 9.7). This taxonomy is categorized following several levels: country, city and district. For instance, two household components have been included, both of which contain a feature which is their location using UTM coordinates. Therefore, these location features allow the dimension to identify which district each household is located in. The TV and Radiator cases are exposed in order to demonstrate why devices must be disaggregated into separated cubes. The main reason for this separation is due to the fact that both devices do not share the same features and, therefore, their classification methods are different. This separation enhances the capacity of making queries since it is possible to filter components by features that are only present in a specific kind of device. 107 9. SIMULATION RESULTS ANALYSIS Figure 9.7: Household dimension. Figure 9.8: TV cube. The TV cube registers data about power consumption as well as the TV mode (off, standby and on) (Figure: 9.8). Every set of measures (power and mode) is related to three dimensions: time, household and TV. Time and household dimensions are exactly the same dimensions as the ones detailed above. The TV dimension contains information about the TVs in a component format. Furthermore, there are two indicators which are responsible for aggregating measures: the mode indicator, which performs a calculation that provides the percentage of TVs that are turned on, and the power indicator, which aggregates the power measures registered using an average formula. The TV dimension, like the household dimension, focuses on specific features related to TV components (Figure: 9.9). In this case, the possibility of filtering TVs using a technological criteria is considered relevant. Therefore, two categories 108 9.2 On-Line Analytical Processing for Smart Grids Figure 9.9: TV dimension. Figure 9.10: Radiator cube. have been created so as to separate LED televisions from LCD televisions. This information will allow us to compare the consumption among the different TV technologies. Hence, TV components contain the technology feature which will be used to calculate whether a TV belongs to the LED or LCD category by using the rules that are related to these categories. The radiator cube stores measures related to both the power consumption and the thermostat level (Figure: 9.10). These measures are related to three dimensions, as in the case of the TV cube. In this case, apart from time and household dimensions, a new dimension has been designed: radiator dimension. This dimension contains components that represent radiators and their features. In addition, there are two indicators which aggregate the measures. On the one hand, the thermostat indicator aggregates the measures stored using a gradient function which shows big changes in the thermostat level in short periods of time. On the other hand, the power indicator aggregates the power measures using an average formula like in the TV cube. 109 9. SIMULATION RESULTS ANALYSIS Figure 9.11: Radiator dimension. The radiator dimension focuses on specific features which concern radiator components (Figure: 9.11). Since radiators are usually considered big consumers, a taxonomy to classify them into two groups has been designed. Indeed, this taxonomy will allow us to find out the amount of radiator components which are in what we consider a small consumer category (under 1kW installed power) or a big consumer category (over 1kW). Two components belong to this dimension and contain the feature installed power which is used to perform the classification in the installed power taxonomy. 9.2.1 N-Level indicators and Data mining So far, some mechanisms which allow us to extract information based on the measures have been presented: indicators. These indicators are regarded as first level indicators since they are just based on measures. However, it is possible to define, a second level of indicators which are computations carried out based on previous level indicators. This idea can be extended to the concept of N-Level indicators. Through introducing this concept, data mining[RKU11] procedures can be used in order to find out patterns. An example of this is presented in figure 9.12. In this case, a data miner has been designed in order to identify consumption habits which concern radiators. Using the thermostat indicator, which calculates the gradient based on the thermostat 110 10.3 Multi Objective Optimisation Where vid(t)is the velocity of the particle iin the dimension dat time t.xid(t) is the position of the particle iin the dimension dat time t.c1and c2are weight factors that adjust the movement. pbestiis the best position achieved by the particle i.pbestgis the best position of the neighbours of the particle i.u1and u2are random factors in an interval of [0,1] and wis the inertial weight. This optimisation method is initialised with a randomly-generated population of particles in the decision space which try to converge into an optimal one by following the best particles at every iteration. The algorithm of this method is presented below: I n i t i a l i s e swarm . I n i t i a l i s e p b e s t . Update g b e s t . I n i t i a l i s e v e l o c i t y . iterationCounter = 0 while (iterationCounter <maxIteration ){ for ( each p a r t i c l e ) { Pick random u1 and u2 for ( each dimension ) { Update p a r t i c l e v e l o c i t y and p o s i t i o n } Update p b e s t } Update g b e s t iterationCounter++ } Report r e s u l t s i n a r c h i v e Listing 10.1: PSO code. For further information please check the bibliography 10.3 Multi Objective Optimisation A multi-objective optimisation problem can be expressed as follows: min x ~ F(x) = [F1(x), F2(x)...Fk(x)]T(10.3) Subject to: gj(x)≤0, j = 1,2...m (10.4) hl(x)=0, l = 1,2...e (10.5) Where K is the number of objective functions and j and l are, respectively, the number of inequalities and equality constraints. x∈Enis the vector of design 117 10. DEMAND SIDE MANAGEMENT POLICY DESIGN variables and ~ F(x)∈Ekis the vector of objectives, criteria, fitness or cost functions to be optimised. Any comparison (≤,≥, etc.) among vectors applies to vector components [MA04]. Other primary concepts in multi-objective optimisation are non-dominated and dominated points [Ste89]. A vector of objective functions ~ F(x∗)∈Zis non-dominated iff there is not another vector, ~ F(x)∈Z, such as ~ F(x)≤~ F(x∗)with at least one Fi(x)< Fi(x∗). Otherwise, ~ F(x∗)is dominated. [LTDZ02] proposed a relaxed form of dominance named -dominance. This acts as an archiving strategy to ensure both properties of convergence towards the Pareto-optimal set and properties of diversity among the solutions found. - dominance is proposed as an extension of the Pareto-dominance relation so that a point x∗not only dominates those points xlower or equal in all their objectives ~ F(x)≤~ F(x∗)and strictly lower in at least one objective Fi(x)< Fi(x∗), but also all points close enough to x∗(i.e., those with a distance to x∗that is less than an ). This value, , can be provided by the decision maker to control the size of the solution set [HDSQCCM07]. In multi-objective optimisation problems, the optimum solution is, in general, not as easy to formulate as in single objective cases. Since typically there is no single global solution, it is often necessary to determine a set of points that all match a predetermined definition of an optimum for every single problem [MA04]. In this sense, the predominant concept in defining an optimal point is that of Pareto optimality [Par06]. A point ~ x∗∈~ Xis Pareto optimal iff there does not exist another point, x∈X such as ~ F(x)≤~ F(x∗)and Fi(x)< Fi(x∗)for at least one function. All Pareto optimal points lie in the boundary of the feasible criterion space Z[AP96]. A Pareto-optima hardly ever provides a single solution, but rather a set of solutions called non-inferior or non-dominated solutions. The minima in the Pareto sense are going to be in the boundary of the objective region, or in the locus of the tangent points of the objective functions, that is in the region in the space of values of the objective functions vector [Coe98]. Often, algorithms provide solutions that may not be Pareto-optimal but may fulfil other criteria, which can be useful for practical applications. It is the case of weakly Pareto optimal. A point ~ x∗∈~ Xis weakly Pareto-optimal iff there is not another point, ~ x∗∈~ Xsuch as ~ F(x)≤~ F(x∗). That is, a point is weakly Pareto optimal if there is no other point that improves all the objective functions 118 10.4 Multi Objective Particle Swarm Optimisation simultaneously. 10.4 Multi Objective Particle Swarm Optimisation The implementation of the MOPSO used in these studies is described in [SC05]. This optimisation method is based on PSO. In this method, the fitness of a particle is calculated considering several objectives which are defined through fitness functions. This method uses the Pareto-optimality [SC05] approach to address the multi-objective problem. A set containing the best particles raised during the iterations is stored in what is named as archive, which retrieves the elements that reach the -dominance. This dominance relaxes the weak-dominance constraints. The next pseudo-code explains how the MOPSO works: I n i t i a l i s e swarm . I n i t i a l i s e l e a d e r s . Send l e a d e r s t o a r ch i ve . crowding ( l e a d e r s ) . iterationCounter = 0 while (iterationCounter <maxIteration ){ for ( each p a r t i c l e ) { S e l e c t l e a d e r . F l i g h t . Mutation . E va l u a t i o n . Update p be st . } Update l e a de rs , Send l e a d e r s t o a r c h i v e . crowding ( l e a d e r s ) . iterationCounter++ } Report r e s u l t s i n a r c h i v e Listing 10.2: MOPSO code. For further information please check the bibliography Initially the swarm is populated through the creation of the particles or individuals. A particle contains a system configuration which is set up randomly. This configuration is the particle genetic code. Before starting the iterative process, the leaders are calculated using the fitness functions which evaluate the optimality of the particles. Those that reach the -dominance criteria will be stored in the archive. That is, the algorithm includes a crowding process that is used to establish a second discrimination criterion (additional to Pareto dominance) [CL02, SC05]. Along the execution this archive may be filled up and, in these cases, the archive deletes particles based on the crowding factor criteria. At every method iteration, each particle is moved according to its previous position and speed which are calculated as follows: 119 10. DEMAND SIDE MANAGEMENT POLICY DESIGN • The speed is calculated based on the previous particle speed, the distance to its best position and the distance to its leader. This method splits the search space in hypercubes, so that the particle leader will be the one that has the best fitness within the hypercube. • The position is calculated adding the previous particle position to the speed calculated above. After the particle movement, the particle could be mutated with a probability calculated as 1 / genetic code size (mutation rate). If a particle is going to be mutated, this mutation may be uniform or non-uniform. The uniform mutation allows the particle to explore the search space. However, a non-uniform mutation allows the particle to exploit the search space wherever it is. Later on, the particle fitness is evaluated and the best position of the particle is changed whenever the new position is better than the best achieved in the past. Once all the particles have been processed, the whole population is evaluated in order to update the leaders of the archive. If this archive reaches the size limit, some of the particles inside it will be deleted following the same criteria as described above. After this step, the next iteration will be executed. The number of iterations must be set beforehand. The reason why this method was selected instead of others such as Genetics Algorithms is that MOPSO has a good result quality and time response relation. The time response is important since the orders could arise regarding the current production-consumption balance in a real-time system where the speediness of applying the measurements is critical. 120 Part IV Experimentation 121 CHAPTER 11 Experimentation considerations This part of the document is devoted to show case studies that have been carried out in this research. As previously said in chapter 7, the hypotheses that have been stated in this document cannot be verified, but validated. That is, it is not possible to verify them as to do so it would be necessary to carry out all SG studies. Nevertheless, they can be validated through the experimentation with case studies. In this part, these case studies are presented and explained, providing evidence that validates the stated hypotheses. In this chapter, some synthetic case studies are presented. The studies have been designed to show how the tools described in the hypotheses part work. That is, how a metamodel is defined, how a simulation model is defined, how to create large-scale scenarios, etc. Each of the chapters in this part refers to a specific case study. These case studies are oriented to solve real problems and, therefore, they can be considered as important evidences that validate the hypotheses. 123 11. EXPERIMENTATION CONSIDERATIONS 11.1 Modelling in Tafat The work, published in [EHH13c] 1, presented in this section concerns prospective experiments that were developed for studying the human flows on a shopping centre from the point of view of the amount of people. In order to run these experiments, the proposed framework will be customised. For this the development of the metamodel, the behaviors and models in which the experiments will be described is necessary. However, as the aim of this case study is not to provide accurate results, but to test the framework for complex system simulation, a non-contrasted information has been used for simulating the human flows. The information we have designed behind the human flows going shopping is expressed in the list below: • From 9.30 to 10.00, around 60-70 workers arrive to the Shopping Centre, leaving it between 16:00 and 16:50. • From 15.30 to 16.00, around 80-90 workers arrive to the Shopping Centre, leaving it between 22:00 and 22:50. • From 10.00 to 16.00, around 1000-2000 customers will arrive to the Shopping Centre, leaving it between 20 and 180 minutes later. • From 16.00 to 22.00 around 2000-4000 customers will arrive to the Shopping Centre, leaving it between 20 and 180 minutes later. If such leaving time reaches the 22.00 limit, the leaving time will be 22.00 The next sub-sections focus on the steps for developing a simulator from scratch for running this kind of experiment. The first step is the metamodel development where the elements that are going to be used in the simulation models are described. Later on, the behaviours which address the way of acting of each model element are developed. Once the metamodel and behaviours are implemented, the Profiler tool is used for generating a huge scenario which contains many agents and two shopping centres. Once the model is ready, the simulation is executed and the results coming from it are analysed. 1People that participated in this experiment: Jos´ e´ Evora, Jos´ e Juan Hern´ andez and Mario Hern´ andez. 124 11.1 Modelling in Tafat Figure 11.1: Case study’s metamodel. 11.1.1 Metamodel development For this experiment, the metamodel development is simple since it is only composed by three elements. Note that the metamodel design is one of the most critical processes since its good design will allow flexibility for adding or modifying elements without affecting the whole structure. In figure 11.1, the metamodel for this complex system is presented. From the scene part, there is only one element: the Shopping centre; the topology is empty in this case and the population has two elements: seller and buyer agents. From the point of view of the experiment context, both do the same: get in and get out from the Shopping centre. However, they are separated in order to provide conceptual clarity. When running experiments, it is important to keep the concepts which exist in reality in the metamodel. The next XML codes represent the description of every metamodel element: <class name=” Sho ppingCen tre ” p a r e n t =” L oc at io n ”> <feature name=” a d d r e s s ” ty pe =” s t r i n g ” /> <v a r i a b l e name=” personCount ” type =” i n t ” i n i t i a l −va lue =”0” /> </class> Listing 11.1: Description of the Shopping Centre metamodel class. This description contains the address where the Shopping Centre is located. Furthermore, it contains a variable which works as a counter of the number of people inside. <class name=” S e l l e r ” p a r e nt =” Bus ine ss ”> <v a r i a b l e name=” s h o p p i n g C e n t r e I d ” t yp e =” s t r i n g ” r e q u i r e d =” t r u e ” >Id of t he shopping c e n t r e where th e s e l l e r works </ v a r i a b l e> <context name=” s ho p pi ngC ent re ” t yp e =” ShoppingCentre ”>Shopping Ce nt re where t h e s e l l e r works </context> 125 11. EXPERIMENTATION CONSIDERATIONS </class> Listing 11.2: Description of the Seller agent metamodel class. When the Seller is instantiated in the simulation model, the shopping centre id is required in order to relate the agent to where it works. The init/configure code will look for this id in order to set in the context attribute the shopping centre provided. <class name=” Buyer ” p a r e n t =” B us in es s ”> <context name=” s h o p p i n g C e n t r e L i s t ” t yp e =” Sho pp in gC en tr e ” replicated=”true”>A l i s t c o n t a i n i n g t he Shopping C ent re s where t he buyer u s u a l l y go </context> </class> Listing 11.3: Description of the Buyer agent metamodel class. In this case, the buyer context is a list of shopping centres where the buyer usually goes. That’s the reason why an id is not necessary to be provided since this context list will be filled by a discovery process where the shopping centres will be associated to this agent following a criteria (e.g. proximity, prices, etc) 11.1.2 Simulation development The simulation development is guided by the presented Simulation Life Cycle which concerned four main steps. At this moment, all the ideas about how the experiment should be developed must be clear in order to decide how to proceed with the data, which elements are necessary to model, etc. 11.1.2.1 Data processing Since the proposed experiments involve the simulation of many elements, the use of a tool for automating and generating simulation models is necessary. In this case, the Profiler is fed by a data store where the statistical information concerning the human flows on the shopping centre is stored. Based on this information, a model that represents a concrete scenario of the people flows is generated. At this point, and through the use of the Profiler tool, the scenario can be modified as needed according to the experiments design in order to test how the people flows would react concerning variables variation. The table into which the hypotheses were translated contains the information about the population of the simulation model (Table: 11.1). Getting into the details 126 19.2 Case study Table 19.1: Vehicles Brand Model Capacity Range Mega e-City 9kWh 100km Reva L-Ion 11kWh 120km Think City 25kWh 200km Mitsubishi i-Miev 16kWh 130km Citroen C-Zero 16kWh 130km Renault Fluence-ZE 22kWh 160km Nissan Leaf 24kWh 160km Tesla Roadster 42 42kWh 257km Tesla Roadster 70 70kWh 483km 19.2 Case study This section is devoted to explain how models have been developed. The EVs have been modelled to represent any EV in the market. It is easily parameterised by simply providing the battery capacity (in kWh), the autonomy (in km) and the starting state of charge (from 0% to 100%). In this simulation, vehicles in table 19.1 have been taken into account. These vehicles are discharged based on the mileage they cover. This means that the energy that has to be discharged from the battery is calculated based on the kilometres that have been driven and the autonomy of the vehicle. Vehicles are charged according to the standard which is at 3,700 watts. This electrical vehicle model needs to be commanded by an agent model in order to perform the trips (e.g. going to work). Therefore, an agent has been developed to implement the behaviour of the driver who needs transportation. This agent has been implemented in accordance to ordinary lifestyles, since it was not possible to obtain information of usage patterns for vehicles. These agents execute four activities which concern the use of the vehicle: going to work, returning from work, going to shops, returning from shops. An example of how these agents are scheduled is presented below: • Activity 1. Id: go work. Time: 7:00. Deviation: 500s. Pattern: Monday, Tuesday, Wednesday, Thursday, Friday. 229 19. AGENT-BASED MODELLING FOR DESIGNING AN EV CHARGING DISTRIBUTION SYSTEMS: A CASE STUDY IN SALVADOR OF BAHIA • Activity 2. Id: back from work. Time: 15:00. Deviation: 500s. Pattern: Monday, Tuesday, Wednesday, Thursday, Friday. • Activity 3. Id: go shop. Time: 18:00. Deviation: 500s. Pattern: Tuesday, Thursday, Sunday. • Activity 4. Id: back from shop. Time: 21:00. Deviation: 500s. Pattern: Tuesday, Thursday, Sunday. In addition to this information, agents are also provided with the distance to work and to the shops. These distances are calculated based on a normal distribution centred in the average mileage driven by people in Salvador de Bahia. After each activity, the agents will plug in the EVs so that they can start charging their batteries. However, depending on the strategy, the connection of the vehicles to the grid will not necessarily involve its charging, as this will be decided by the strategy. This means that in the simulation in which there is not a strategy, EVs will start to charge as soon as they are plugged in. In the simulation in which there is a strategy, EV charging will start in accordance to the policies of the strategy. 19.3 Results Keeping the same distribution infrastructures, the charging potential of the substation for electrical vehicles can be defined as the difference between the maximum deliverable power to vehicles connected to charging stations and the current energy demand: Pp(t) = Pm(t)−Pd(t). Obviously, this potential is not the same in all power grid substations, since both factors, Pm and Pd, may vary from one substation to another as well as seasonally. In the proposed charging model that this paper explores, one or more substation feeders will be enabled to allow the management of all vehicles based on these two factors. Ideally, the energy used for their charging should be the cheapest. Moreover, system operation should be facilitated so that generators are switched on and off as little as possible. To this end, the objective should pursue the flattening of the demand curve. 230 19.3 Results 0 2 4 6 8 10 12 14 16 18 0:00 2:00 4:00 6:00 8:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 MW Figure 19.1: EVs consumption when they start charging as soon as they are plugged 19.3.1 No DSM policy In this experiment, the model previously presented will be simulated considering that EVs start charging as soon as they are plugged in. In the figure 19.1, the consumption of EVs for a weekday is presented. The highest peak of this load curve can be found at 7pm approximately. This peak is due to the fact that almost everyone is at home at this time, with their EV plugged in, with the result that that their charging reaches the greatest level of overlapping. This overlapping effect is also high in the morning and afternoon, coinciding with commuter “rush hours”. In the figure 19.2, two load curves are presented: the original load curve for 2014 (grey) and for 2030 in which EVs are included (black). In this chart, it can be seen the increment that involves the EVs in the total consumption. As a remainder, this significant impact is caused by 5,253 EVs in a power grid serving an area of 142,000 (in 2014) and 149,000 (predicted for 2030) inhabitants. If any of the three predictions that are used in this study increases (population, vehicle ownership rate or EVs penetration rate), would result in the consumption increment being higher. It can be observed that the impact of introducing 5,253 EVs can reach 8 MW, which is significant considering that the maximum consumption of this area for the analysed day is 103 MW. This approximately means an 8% increment. Along with this increment, we must also consider how the rest of the energy demand not provoked by the EVs introduction has increased as well. This leads to a total increment of approximately 15MW. This would mean that this substation would have to be re-sized. 19.3.2 RTP-based DSM policy Knowing the valley existent in the early morning, RTP policy has been programmed to output a very cheap price for the energy that is consumed from 1am to 8am. In these experiments, all vehicles have been tuned to only accept this price. In the 231 19. AGENT-BASED MODELLING FOR DESIGNING AN EV CHARGING DISTRIBUTION SYSTEMS: A CASE STUDY IN SALVADOR OF BAHIA 0 20 40 60 80 100 120 0:00 2:00 4:00 6:00 8:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 MW Figure 19.2: Comparison of the original load curve for 2014 (grey) with the total consumption including EVs charging for 2030 (black) 0 2 4 6 8 10 12 14 16 18 0:00 2:00 4:00 6:00 8:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 MW Figure 19.3: EVs consumption according to the RTP based policy which makes cheaper charging from 1am to 8am figure 19.3, the consumption of the EVs is presented under this new policy. As it can be observed, since many vehicles have not been charged during the previous days, most of them start charging their batteries at 1am, thus producing a high demand peak of 16MW. This consumption becomes 0 at 7am approximately. In the figure 19.4, again, two load curves are presented: the original load curve for 2014 (grey) and for 2030 in which EVs are included (black). This time, the peak that already existed in the load curve for 2014 has only increased as consequence of the population increase for 2030. This time, EV charging has been shifted to an off-peak period which is in the early morning. In this new interval, EVs charging does not increase at all the overall maximum peak of the demand allowing for a greater introduction of EVs if it was necessary. Notwithstanding, there is an abrupt increase of the consumption at 1am since almost all EVs start charging at that time due to the low prices. This is a non-desired effect for the grid stability and should be further studied in order to smooth their charging starting process. 232 19.4 Discussion 0 20 40 60 80 100 120 0:00 2:00 4:00 6:00 8:00 10:00 12:00 14:00 16:00 18:00 20:00 22:00 MW Figure 19.4: Comparison between original load curve for 2014 (grey) with the total consumption including EVs charging (black) using the RTP-based policy for 2030 19.4 Discussion This work has been oriented to develop the model of Salvador of Bahia’s power grid using an Agent-based modelling approach. This work has allowed to validate the ability of this modelling approach to: 1. Model individual decision making. 2. Consider local constraints. 3. Model power grids as social technical system. 4. Study emergent synchronisation and coupling effects when many agents coincide. The major advantage is that this model is able to simulate human decisions and actions that would affect the functioning of the power grid. In addition, such changes in the grid would in turn influence human decisions and actions. Once this model is developed, future projects will be oriented to design and assess DSM policies that would help reduce investments on grid infrastructure. On the other hand, the agent decision model could also be improved. For example, agents could make decisions based on responses to changes in the system, which will in turn change the context for future decisions; or agents could behave in a heterogeneous way, thus maximising a certain profit, either a full charging of the vehicle or save money. 233 Part V Conclusions 235 CHAPTER 20 Results Research work has been conducted to deal with the problems of designing and evaluating management policies related to the SG paradigm. These policies are aimed at different stakeholders to benefit their interests. For instance, final customer’s interests is the reduction of the electricity costs and the improvement of the energy efficiency. The interests of operators and retailers is to improve their profit by reducing operation and production costs. Globally, people want to have a more efficient power grid in which CO2emissions are reduced with a higher RES penetration. The motivation for conducting research in the field of Power Grids is due to its proposed evolution to SGs. In this evolution, IT seems to have an important role. This evolution is proposed because there is a significant need for the improvement of power grids, such as: reduction of the dependency on fossil-fuel based energy production; market concerns such as fuel prices volatility; and reduction of GHG emissions. SGs aim to face the next challenges: higher introduction of RES; more efficiency and flexibility; and less dependency on fossil fuels. Moreover, they are characterised by the introduction of automation at distribution level. It is expected that this evolution will overcome the challenge of the massive introduction of electrical vehicles, distributed generation and RES [HHM11]. The SG paradigm 237 20. RESULTS is an even more important challenge for isolated power grids as they have limited resources on the production side to balance offer and demand. This is the case of the Canary Islands. In this territory, there are excellent environmental conditions for increasing the exploitation of RES. The Canary Islands have 2,500-3,000 hours/year of solar radiation producing an average of 56kWh/m2per day (source: Canary Institute of Technology). Furthermore, there are 3,000-4,500 hours/year of wind with a speed average of 7-8m/s. This produces 625MWh per day having installed 75MW of wind energy. In this sense, the results of this research can be directly applied to improving power grids in the Canary Islands. The study of these smart management policies requires models representing the system at the lowest level of detail allowing to analyse side-effects. Aggregated models are not able to represent events on the demand side that should be considered in designing these policies. Therefore, Power Grid models should be disaggregated to represent all loads that can be managed. In these models, complexity arises since there are many heterogeneous components which are interacting at different scales. The emergent behaviour of the system cannot be inferred with traditional models, because complexity cannot be represented. SGs cannot be studied using traditional approaches since these are not able to capture all of the system’s properties. The best way to study complex models is through Software tools. However, if SGs are not supported by the required formalisms and methodologies, software technologies by themselves are not sufficient to perform these studies. This means, the problem is not choosing the software tool, but applying the right formalisms and methodologies. For these reasons, several authors have suggested the use of complex system based formalisms to represent SGs. A Complex Systems approach facilitates the representation of SGs at the lowest level of detail. Under the complex system approach, only software tools that support this formalism can be considered to study SGs. Nevertheless, there are structural concerns that must be taken into account to support the engineering of large-scale complex systems. In the case of SGs, models may have millions of components. To model them, several complex system simulation software tools have been reviewed in this document. In all these tools, a lack of a methodological orientation to face the modelling of large-scale complex systems was found. 238 20.7 Publications • Towards an interdisciplinary approach for the simulation of future SG architectures from a complex systems science point of view –Authors: Viejo, P., Kremers, E., Hern´ andez, M., Hern´ andez, J., ´ Evora, J., Langlois, P., Daude, E., Gonz´ alez de Durana, J.M., & Barambones, O. –Congress: European Conference on Complex Systems 2011 –Place: Vienna, Austria –Date: September, 2011 • A large-scale electrical grid simulation for massive integration of distributed photovoltaic energy sources –Authors: ´ Evora, J., Kremers, E., Hern´ andez, M., & Hern´ andez, J. J. –Congress: 6th European Conference on PV-Hybrids and Mini-Grids (OTTI) –Place: Chambery, France –Date: April, 2012 • Modelling lifestyle aspects influencing the residential load-curve –Authors: Hauser, W., ´ Evora, J., & Kremers, E. –Congress: 26th European Conference on Modelling and simulation (ECMS 2012) –Place: Koblenz, Germany –Date: May, 2012 • Asynchronous Smart Grid Simulations –Authors: ´ Evora, J., Hern´ andez, J. J., & Hern´ andez, M. –Congress: UCNC’12 (Unconventional Computation and Natural Computation): CoSMoS - Proceedings of the 2012 Workshop on Complex Systems Modelling and Simulation –Place: Orleans, France –Date: September, 2012 • Decision support for Complex Systems: a Smart Grid case –Authors: ´ Evora, J., Hern´ andez, J. J., & Hern´ andez, M. 245 20. RESULTS –Congress: UCNC’13 (Unconventional Computation and Natural Computation): CoSMoS - Proceedings of the 2013 Workshop on Complex Systems Modelling and Simulation –Place: Milan, Italy –Date: July, 2013 • Criticality in complex sociotechnical systems: an empirical approach –Authors: Viejo, P., Kremers, E., ´ Evora, J., Hern´ andez, J. J., Hern´ andez, M., Barambones, O., & Gonz´ alez de Durana, J. M. –Congress: European Conference on Complex Systems 2013 (ECCS’13) –Place: Barcelona, Spain –Date: September, 2013 • Asynchronous approach to simulations in Smart Grid –Authors: ´ Evora, J., Hern´ andez, J. J., & Hern´ andez, M. –Congress: European Simulation and Modelling Conference 2013 (ESM’13) –Place: Lancaster, England –Date: October, 2013 • Tafat: A framework for developing simulators based on Model Driven Engineering –Authors: ´ Evora, J., Hern´ andez, J. J., & Hern´ andez, M. –Congress: European Simulation and Modelling Conference 2013 (ESM’13) –Place: Lancaster, England –Date: October, 2013 • Model Driven Engineering for data miners simulation –Authors: ´ Evora, J., Hern´ andez, J. J., & Hern´ andez, M. –Congress: IEEE International Conference on Data Mining 2013 (ICDM’13) - International Workshop on Domain Driven Data Mining –Place: Dallas, United States of America –Date: December, 2013 • Agent-based modelling for designing an EV charging distribution systems: a case study in Salvador of Bahia –Authors: ´ Evora, J., Hern´ andez, J. J., Barbosa, D. & Nazareno, P. 246 20.7 Publications –Congress: European Simulation and Modelling Conference 2014 (ESM’14) –Place: Porto, Portugal –Date: October, 2014 Next list contains the publications released in journals: • Advantages of Model Driven Engineering for studying complex systems –Authors: ´ Evora, J., Hern´ andez, J. J., & Hern´ andez, M. –Journal: Natural Computing –DOI: 10.1007/s11047-014-9469-y • Criticality in complex sociotechnical systems: an empirical approach to electrical grids 1 –Authors: Viejo, P., Kremers, E., ´ Evora, J., Hern´ andez, J.J., Hern´ andez, M., Barambones, O, Gonz´ alez de Durana, J.M. –Journal: Applied energy 1This paper has been submitted and now it is awaiting for the acceptance 247 CHAPTER 21 Discussion In this chapter the scientific and technical contributions of this research are described. “Scientific contributions” refer to the creation of knowledge whereas technical contributions refer to the creation of new processes to deal with specific problems. The purpose of this research is to explore and describe the application of already existing methodologies to identify problems in SGs. This exploration has also allowed an outlook on new research targets that could be studied in future research. 21.1 Contributions 21.1.1 On Applying the complex system approach in Smart Grids According to the well-honoured philosopher of science Thomas Kuhn, in the scientific development of a discipline there are three main stages [Kuh12]. These three main stages are: “prescience”, “normal science” and “revolutionary science” (Figure: 21.1). The “prescience” stage is characterised by having numerous incompatible and incomplete theories. The consensus of a prescientific community in terms of methods, terminologies, experiments, etc. leads to the “normal science”. In this stage, new paradigms can be conceived to deal with anomalies reaching the “revolutionary science” stage. 249 21. DISCUSSION Figure 21.1: Kuhn’s cycle. In the case of the SG topic, it can be considered to be currently in the second stage: “normal science”. Within the study of this normal science, several limitations have been found when SGs are being modelled (appearance of anomalies). The complex system approach has been proposed to overcome them (new paradigms). There are several studies in the documentation that assert that power grids cannot be studied using traditional formalisms. In these studies, the hypothesis states that complex system approach could overcome the limitation of the traditional formalisms. This hypothesis cannot be absolutely verified since it cannot be proved to carry out all possible case studies. However, what the hypothesis states can be reproduced and refuted. In this research, this hypothesis has been reproduced through the experimentation of case studies. These case studies provide additional evidence that this formalism is valid [CGLP94, PKZK08, EKM+11]. Several case studies have been carried out to analyse SGs following complex system approach. It has been found that this formalism is useful for studying SG policies with a maximum level of disaggregation. When a new policy is being proposed, the behaviour of a power grid is not known. In these cases, the system can be modelled through its components’ behaviours and simulated to obtain the emergent behaviour. A maximum level of disaggregation also allows to re-aggregate the components’ behaviours and analyse 250 21.1 Contributions them at different levels (e.g. appliance, household, district, country level). Nevertheless, in this research, it has been recognised that complex system approach by itself may not be enough to deal with the study of SGs. Other issues appear when using this formalism: complexity for modelling systems, complexity for analysing data and complexity for designing strategies. An important contribution is the identification of complementary methodologies or formalisms that could be useful for overcoming these issues. In particular, the application of MDE, BI and SI should be considered. The validity of this contribution has been demonstrated through the execution of case studies. As in the case of complex system approach, these hypotheses cannot be verified. In order to verify them absolutely, these hypotheses should be applied to perform all SG case studies, which is impossible. Nevertheless, in this research, several case studies have been carried out to provide evidence of its validity. 21.1.2 On Modelling Smart Grids This research has found that the modelling of SGs requires the coexistence of different natures in their components. When SGs are being modelled, it is necessary to describe components that may have different natures such as: electrical, thermodynamical, meteorological or sociological, among others. Obviously, electrical components must be represented since their behaviour is closely related to the human behaviours (e.g. when these components are switched on and off). Actually, the way in which the humans use these electrical components will determine the system’s emergent behaviour. The interaction they do over this components involve consumptions that affect how the system works. Thus, many electrical models must be modelled in order to represent production, distribution, demand, etc. with the degree of detail that is necessary for testing the SG implementation. In addition, thermodynamical components must also be modelled since it is necessary to represent the thermal transferences between thermal electrical components (e.g. refrigerators, domestic water heaters, etc.). The modelling of these electrical components must include thermodynamical behaviour that calculates the thermal gains derivative from their consumptions. Moreover, infrastructure as households may include a thermal behaviour as well. This behaviour would cal251 21. DISCUSSION culate the infrastructure temperature based on the external temperature and the thermal gains derivative from the thermal electrical loads. In this manner, thermal electrical loads can calculate their internal temperature based on their consumption and the household temperature. The calculation of the internal temperature of a household depends on the external temperature, as previously stated. Therefore, it is necessary to represent the meteorological conditions of the environment where the household is located. Thus, another kind of behaviour which is necessary to model power grids is identified. Furthermore, meteorological conditions must be represented if there are RES in the power grid under study. Renewable energy technologies, such as PV cells and wind turbines, require the representation of environmental conditions. These conditions are used by the models of these types of technologies to calculate the energy produced. Furthermore, the representation of SGs also use the modelling of human interactions. Human beings are users of Power Grids and the way they use energy must be represented. This is a very important consideration because of the significant impact human interaction has on the power grid. As mentioned previously, in state of the art, power demand varies geographically, seasonally, culturally, sociologically, etc. All these variations have to do with the way in which humans interact with the power grid. In conclusion, the study of the application of SG policies requires the modelling of many components with heterogeneous natures. The behaviours that have been previously mentioned are not only different because they have different natures: electrical, thermal, social, etc. but also because they require different execution paradigms: continuous (e.g. thermal) or discrete (e.g. washing machine). 21.1.3 On Developing Complex Models In this research, MDE has been found to be a complementary methodology for modelling SGs with a complex system approach. When it is necessary to integrate different modelling techniques, structural guidelines are needed to deal with this complexity. In these cases, MDE is useful for managing the model definition and the simulator construction. In this research, the application of MDE has been validated by carrying out the 252 21.1 Contributions case studies. Moreover, this approach is being used by other researchers. Currently, EIFER and EDF are developing their own case studies with this methodology. From a scientific point of view, it is known that other laboratories are researching in the hypothesis. To this day, the hypothesis has not been refuted. The most important implication of applying MDE in this field is technical: MDE helps to build up and maintain large-scale simulation models. Modellers are assisted in the construction process by separating concerns: the scenario description (simulation model) and the description of individual components (behaviours). In this way, a set of steps are established to guide the modeller to the goal: • Developing the simulation model. • Describing a component in the metamodel if it has not been previously defined there. • Developing necessary behaviours not yet modelled. • Building a simulator based on this simulation model and simulate it. Another MDE implication is its adaptability to changes. A single component or behaviour can be easily replaced or added to an existing model scenario. Flexibility is an important feature for engineering SGs. In the design of a new policy, it is necessary to test it in different scenarios. The flexibility provided by MDE makes this possible. These scenarios may not only use different behaviours for the components but may also organise components in a different way. In addition, the same scenario can be used for testing different policies. Since MDE requires the definition of a metamodel, a technical contribution of this research is the development of a metamodel for SGs. In this sense, a metamodel can be understood as the creation of an agreement contract between many modellers in order to make collaboration easier. The metamodel has several important implications: • Modellers can share their models whenever they have been built under the same metamodel. • Modellers can contribute to the metamodel by extending its definition or creating new behaviours that can be used by others. • Modellers do not need to describe the components of the reality every time they start a new simulation since they take advantage of the already defined metamodel. 253 21. DISCUSSION • Modellers can use previous work developed by other modellers. • Modellers can develop their models faster thanks to all of these above-mentioned factors. In this way, large simulation models can be built by developing the simulation model; and, in case it is needed, developing new components and behaviours. In this sense, modellers are guided completely so they know what they have to do to develop simulation models. Therefore, the complexity of the scenario only affects the simulation model and not the overall framework. Furthermore, tools have been developed to deal with the complexity of the representation of large-scale system in this simulation model. The tool named Profiler was a response to the need of many modellers had for the construction of largescale scenarios. In this sense, this tool handles the complexity of building these simulation models. Based on input data provided by the modeller, this tool creates the scenario. Another important tool is the Simulator Builder which builds a simulator based on a model that describes the scenario. 21.1.4 On Analysing Simulation Results This research demonstrates that the application of BI methodologies is helpful. These methodologies have been used to analyse data coming from the simulation of a complex system. When SGs are studied as complex systems, the huge quantity of results coming from these simulations can be managed using BI methodologies. BI methodologies can be used to discover laws and regularities that provide a better knowledge of the emergent behaviour in SGs. These methodologies allow science to be done over simulated experiments. Based on the output available from the simulations, it is possible to find rules, laws, regularities, etc. automatically. In this research, BI methodologies are used on the output of the case studies presented. The use of these methodologies is really important in order to know how the emergent behaviour in SGs is reached; and how it can be influenced. After performing all of the case studies presented in this document, it has been shown that BI can be used in this context with naturalness. The mechanisms provided by BI are sufficient to represent data coming from SG simulations and allow its exploitation. 254 21.3 Future work are progressively introduced; or RES quota increases. Furthermore, case studies in which both scenarios and policies can evolve can be considered. These kind of scenarios may be important for testing the applicability of SG policies in the long-term. In order to support them, new mechanisms must be designed within this ecosystem of methodologies and formalisms. 21.3.3 On Discovering Patterns As said before, BI methodologies can be used to find patterns in the simulations of SGs. The discovery of patterns may have many different usages: obtaining a better knowledge of how emergence is reached, simplifying a complex model, feeding back into the design of SG policies, etc. In this field, much work can be carried out as a continuation of this research. Data mining techniques can be explored in order to automatically detect patterns in series of data. However, this is not an easy task, as there is a main underlying concern that must be considered when exploring these techniques: the huge quantity of data. As stated in the hypotheses part of the document, an important concern of data management is the velocity. Huge quantity of data is contradictory to the velocity. This is especially tricky when dealing with data mining techniques that require a thorough analysis of the data. For this reason, this research should consist in: investigating data mining procedures to find patterns; and finding procedures to accelerate the process. One of the applications that this research may have is the simplification of complex models. This would open another interesting future project that could consist in automating the model simplification process. That means, the use of methodologies, formalisms or techniques to identify behavioural features in complex models that can characterise simplified models. To this end, results coming from the discovery of patterns can be used in this research. After, software tools could be able to parse data coming from complex models in order to create simplified models. The behaviours provided by the simplified model must be certified to be working at certain intervals that ensure measurable errors. For instance, these simplified models could be used as submodels within bigger simulations. Therefore, smaller complex simulations can be used to 261 21. DISCUSSION engineer bigger ones in which the computational costs can be reduced. 21.3.4 On Other Fields In the “hypotheses part” of this document, the definition of external validity in the context of this research was discussed. However, a hypothesis has external validity if it applies to several case studies of SG policies. This external validity has been checked in this document as these methodologies have been successfully applied in all presented case studies. Nevertheless, the second definition provided for the external validity is not dealt within this research. This definition concerned the use of these methodologies in other complex systems apart from SGs. This is actually an extension of the scope of the hypotheses defined in this work. It is reasonable to believe that if their application in the field of SGs has been useful, their application in other fields can be helpful as well. However, this should be checked and is proposed for future research. In this sense, the hypothesis to be researched could be: “MDE, BI and SI can be helpful methodologies for engineering complex systems of any nature”. As in the previous hypotheses, this cannot be fully verified because it is not possible to perform all of the case studies in all complex systems. However, a subset of them can be carried out, which could provide new ideas or requirements. 21.3.5 On Experimenting with Overview, Design concepts and Details protocol Another line that can be explored is the integration of the Overview, Design concepts and Detail (ODD) protocol in our approach to model agent-based models with MDE. This protocol is a proposed standard for describing agent-based models [GBB+06]. Therefore, our approach could include the description of models that are compliant with ODD. This protocol consists of three blocks (Overview, Design concepts and Details), which are subdivided into seven elements: Purpose, State Variables and Scales, Process Overview and scheduling, Design concepts, Initialization, Input, and Submodels. Therefore, agent-based models can be described according to these seven 262 21.3 Future work elements. The next list summarises the contents to place at each of these elements: • Purpose: description of the scope of the model. Clear, concise and specific formulation of the model’s purpose. Explanation of the need for building a complex model. • State variables and scales: state variables refer to low-level variables that characterise low-level entities of the model. The scales refer to the simulation configuration (e.g. length of time steps, size of habitat cells). • Process Overview and scheduling: the process overview describes environmental and individual processes that are built into the model (e.g. food production, growth, movement). The schedule has to do with how these processes occur and their order. • Design concepts: these concepts provide a common framework for designing and communicating individual-based models. • Initialization: defines how the environment and the individuals have to be created. • Input: conditions of the environment that change over space and time. • Submodel: presentation and explanation of all the processes listed above in a detailed manner, including the parameterisation of the model. 263 Bibliography [ADG08] A Arenas and A Diaz-Guilera. Synchornization in complex networks. Physics Reports, 469:93–153, 2008. 191 [ADHM] Robert Arritt, Roger Dugan, Robert Henry, and Tom McDermott. OpenDSS. 35, 37 [AEE08] ADEME, ENERTECH, and EDF. Campagne de mesures des appareils de production de froid et des appareils de lavage dans 100 logements. Technical report, ADEME and ENERTECH and EDF, 2008. 150 [AER] ADEME, EDF, and R´ egion R´ eunion. Millener, une longueur d’avance pour la R´ eunion. 9 [AES07] M.H. Albadi and EF El-Saadany. Demand response in electricity markets: An overview. In Power Engineering Society General Meeting, 2007. IEEE, pages 1–5. IEEE, 2007. 30 [AG12] Hazem Ahmed and Janice Glasgow. Swarm intelligence: Concepts, models and applications. School Of Computing, Queens University Technical Report, 2012. 47 [All09] P Allen. Complexity, Innovation and Organizational Evolution. 2009. 189 [Ami01] Massoud Amin. Toward self-healing energy infrastructure systems. Computer Applications in Power, IEEE, 14(1):20–28, 2001. 4, 19 265 BIBLIOGRAPHY [And09] David L Anderson. An evaluation of current and future costs for lithium-ion batteries for use in electrified vehicle powertrains. PhD thesis, Duke University, 2009. 5 [AP96] Timothy Ward Athan and Panos Y Papalambros. A note on weighted criteria methods for compromise solutions in multiobjective optimization. Engineering Optimization, 27(2):155– 176, 1996. 118 [B+09] Albert-L´ aszl´ o Barab´ asi et al. Scale-free networks: a decade and beyond. science, 325(5939):412, 2009. 40 [BA03] Robert Blumberg and Shaku Atre. The problem with unstructured data. DM REVIEW, 13:42–49, 2003. 62 [BB02] Tim M Blackwell and Peter Bentley. Improvised music with swarms. In Computational Intelligence, Proceedings of the World on Congress on, volume 2, pages 1462–1467. IEEE, 2002. 48 [BBS+09] Prasanna Balaprakash, Mauro Birattari, Thomas St¨ utzle, Zhi Yuan, and Marco Dorigo. Estimation-based ant colony optimization and local search for the probabilistic traveling salesman problem. Swarm Intelligence, 3(3):223–242, 2009. 47 [BC05] C Bueno and J A Carta. Technical–economic analysis of windpowered pumped hydrostorage systems. Part II: model application to the island of El Hierro. Solar energy, 78(3):396–405, 2005. 5 [BF04] A Borshchev and A Filippov. Anylogic-multi-paradigm simulation for business, engineering and research. In The 6th IIE Annual Simulation Solutions Conference, volume 150, 2004. 45 [BI04] J.P. Barton and D.G. Infield. Energy storage and its use with intermittent renewable energy. Energy Conversion, IEEE Transactions on, 19(2):441–448, June 2004. 28 266 BIBLIOGRAPHY [Blu05] Christian Blum. Beam-aco-hybridizing ant colony optimization with beam search: An application to open shop scheduling. Computers & Operations Research, 32(6):1565–1591, 2005. 47 [Blu08] Christian Blum. Beam-aco for simple assembly line balancing. INFORMS Journal on Computing, 20(4):618–627, 2008. 47 [BLW05] Paul Baker, Shiou Loh, and Frank Weil. Model-driven engineering in a large industrial context-motorola case study. In Model Driven Engineering Languages and Systems, pages 476–491. Springer, 2005. 55 [BM01] Eric Bonabeau and Christopher Meyer. Swarm intelligence. Harvard Business Review, 79(5):106–114, 2001. 46 [BMW10] PE Baker, C Mitchell, and B Woodman. Electricity market design for a low-carbon future. London, UKERC, 2010. 24 [Boh03] Marko Bohanec. Decision support. Springer, 2003. 57 [BPSK02] Michael Butler, Luigia Petre, Kaisa Sere, and Stuart Kent. Model Driven Engineering. In Integrated formal methods, volume 2335 of Lecture Notes in Computer Science, pages 286–298. Springer Berlin / Heidelberg, 2002. 50, 52 [Bre00] Marilynn B Brewer. Research design and issues of validity. Handbook of research methods in social and personality psychology, pages 3–16, 2000. 72 [BRS+10] Steve Bou Ghosn, Prakash Ranganathan, Saeed Salem, Jingpeng Tang, Davin Loegering, and Kendall E. Nygard. Agent-Oriented Designs for a Self Healing Smart Grid. In 2010 First IEEE International Conference on Smart Grid Communications, pages 461–466. IEEE, October 2010. 36 [Bun11] Bundesverband der Energieund Wasserwirtschaft. BDEW - Standardlastprofile (SLP), 2011. 143 267 BIBLIOGRAPHY [BVB08] Christian Blum, Mateu Y´ abar Vall` es, and Maria J Blesa. An ant colony optimization algorithm for dna sequencing by hybridization. Computers & Operations Research, 35(11):3620–3635, 2008. 47 [BW89] Gerardo Beni and Jing Wang. Swarm Intelligence in Cellular Robotic Systems, Proceed. In NATO Advanced Workshop on Robots and Biological Systems, Tuscany, Italy, 1989. 45 [BW93] Gerardo Beni and Jing Wang. Swarm intelligence in cellular robotic systems. In Robots and Biological Systems: Towards a New Bionics?, pages 703–712. Springer, 1993. 45, 46 [BY97] Yaneer Bar-Yam. Dynamics of complex systems, volume 213. Addison-Wesley Reading, MA, 1997. 39 [C+94] CDIF Technical Committee et al. Cdif framework for modeling and extensibility. Electronic Industries Assocaiation, EIA/IS-107, 1994. 52 [C+06] European Commission et al. European technology platform smart grids: vision and strategy for europe’s electricity networks of the future. Office for Official Publications of the European Communities, 2006. 22 [Cao04] Junwei Cao. Self-organizing agents for grid load balancing. In Proceedings of the 5th IEEE/ACM International Workshop on Grid Computing, pages 388–395. IEEE Computer Society, 2004. 46 [CCS93] Edgar F Codd, Sharon B Codd, and Clynch T Salley. Providing olap (on-line analytical processing) to user-analysts: An it mandate. Codd and Date, 32, 1993. 60 [CD97] Surajit Chaudhuri and Umeshwar Dayal. An Overview of Data Warehousing and OLAP Technology. ACM Sigmod Record, 26(1), 1997. 103 268 BIBLIOGRAPHY [CD10] E Chappin and G.P. Dijkema. Agent-based modelling of energy infraestructure transitions. International journal of Critical Infraestructures, 6:106–130, 2010. 190 [CDMT94] Alberto Colorni, Marco Dorigo, Vittorio Maniezzo, and Marco Trubian. Ant system for job-shop scheduling. Belgian Journal of Operations Research, Statistics and Computer Science, 34(1):39–53, 1994. 47 [CGLP94] A. Capasso, W. Grattieri, R. Lamedica, and A. Prudenzi. A bottom-up approach to residential load modeling. Power Systems, IEEE Transactions on, 9(2):957–964, 1994. 8, 32, 33, 250 [CL02] C A Coello Coello and M S Lechuga. MOPSO: a proposal for multiple objective particle swarm optimization. In Evolutionary Computation, 2002. CEC ’02. Proceedings of the 2002 Congress on, volume 2, pages 1051–1056, 2002. 119, 164, 168 [Cla11] C´ edric Clastres. Smart grids: Another step towards competition, energy security and climate change objectives. Energy Policy, 39(9):5399–5408, September 2011. 3, 20 [CMB10] A.J. J. Conejo, J.M. M. Morales, and L. Baringo. Real-Time Demand Response Model. Smart Grid, IEEE Transactions on, 1(3):236–242, 2010. 33 [CMG11] Timothy Cooper, Gillian Mead, and Phil Gr¨ unewald. The Future of Energy Storage: Stakeholder Perspectives and Policy Implications. Technical report, UK Energy Research Centre, Imperial College, 2011. 23 [CnA00] C.A. Ca˜ nizares and F.L. Alvarado. UWPFLOW Program, University of Waterloo, 2000. 35, 36 [Coa99] Meta Data Coalition. Open information model. At URL: http://www. mdcinfo. com/OIM/MDCOIM10. pdf, 1999. 52 269 BIBLIOGRAPHY [Coe98] Carlos A. Coello Coello. A comprehensive survey of evolutionary-based multiobjective optimization techniques. Knowledge and Information Systems, 1:269–308, 1998. 118 [Col03] Nick Collier. Repast: An extensible framework for agent simulation. The University of Chicago’s Social Science Research, 36:2003, 2003. 44 [CSMJ13] Trinidad Chardin-Segui, Sebastien Mathiou, and Yves Jeanjean. The Millener project: Five hundred RES & one thousand fifty DSM installations for households in French islands, aggregated for grid support. In 2013 IEEE Grenoble Conference, pages 1–6. IEEE, June 2013. 9 [Dep] Deparment of Energy, USA. http://energy.gov/oe/technologydevelopment/smart-grid. 21 [DGS07] Joyce Dargay, Dermot Gately, and Martin Sommer. Vehicle ownership and income growth, worldwide: 1960-2030. The Energy Journal, pages 143–170, 2007. 228 [DMCM91] Marco Dorigo, Vittorio Maniezzo, Alberto Colorni, and Vittorio Maniezzo. Positive feedback as a search strategy, 1991. 47 [Dor92] Marco Dorigo. Optimization, learning and natural algorithms. Ph. D. Thesis, Politecnico di Milano, Italy, 1992. 47 [Dou06] Douglas C. Schmidt. Guest Editor’s Introduction: Model-Driven Engineering. Computer, 39:25–31, February 2006. 49, 50, 51 [Duf10] D Duffert. Lastmodellierung zur Untersuchung von Lastmanagement - Maßnahmen in Elektrizit¨ atsnetzen. PhD thesis, Final Project Degree, Karlsruhe Institute of Technology, 2010. 182 [DW07] Patrick Devine-Wright. Energy citizenship: psychological aspects of evolution in sustainable energy technologies. People centred eco-design: consumer adoption of low and zero carbon products and systems, page 63, 2007. 24 270 BIBLIOGRAPHY National Institute of Standards and Technology, page 33, 2012. 21, 25 [Joh07] Neil F Johnson. Two’s company, three is complexity: a simple guide to the science of all sciences. Oneworld Pubns Ltd, 2007. 67 [Kau87] Fred J Kaudel. A literature survey on distributed discrete event simulation. SIGSIM Simul. Dig., 18(2):11–21, 1987. 86 [KC00] Joshua D Knowles and David W Corne. Approximating the nondominated front using the pareto archived evolution strategy. Evolutionary computation, 8(2):149–172, 2000. 168 [KdD12a] Enrique Kremers and J.M. Gonzalez de Durana. Emergent synchronisation properties of a refrigerator demand side management system. Applied Energy, 2012. 191, 220, 222 [KdD12b] Enrique Kremers and J.M. Gonzalez de Durana. Towards complex system design and management in the engineering domain - The Smart grid Challenge. In European Conference on Complex Systems 2012, 2012. 191, 194, 220, 223 [KdH09] S. Karnouskos and T.N. de Holanda. Simulation of a smart grid city with software agents. In Computer Modeling and Simulation, 2009. EMS ’09. Third UKSim European Symposium on, pages 424–429, Nov 2009. 39 [KE95] J Kennedy and R Eberhart. Particle swarm optimization. In Neural Networks, 1995. Proceedings., IEEE International Conference on, volume 4, pages 1942 –1948 vol.4, 1995. 47, 116, 164 [KHTS09] P. Kordik, M. Hasaj, M. Tvrdy, and M. Snorek. Building Automation Simulator and Control Strategy for Intelligent and Energy Efficient Home. In Computer Modeling and Simulation, 2009. EMS’09. Third UKSim European Symposium on, pages 154–159, 2009. 33 277 BIBLIOGRAPHY [kit] KIT web. http://www.kit.edu/. 139 [KKE01] James F Kennedy, James Kennedy, and Russell C Eberhart. Swarm intelligence. Morgan Kaufmann, 2001. 162 [KR02] Ralph Kimball and Margy Ross. The Data Warehouse Toolkit: The Complete Guide to Dimensional Modeling. John Wiley & Sons, Inc., New York, NY, USA, 2nd edition, 2002. 63 [Kre13] Enrique Alberto Kremers. Modelling and Simulation of Electrical Energy Systems through a Complex Systems Approach using Agent-Based Models. KIT Scientific Publishing, Karlsruhe, 2013. 38, 40, 192 [KRH+10] Mariam Kiran, Paul Richmond, Mike Holcombe, Lee Shawn Chin, David Worth, and Chris Greenough. Flame: simulating large populations of agents on parallel hardware architectures. In Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems: volume 1-Volume 1, pages 1633–1636. International Foundation for Autonomous Agents and Multiagent Systems, 2010. 45 [KSE07] Oliver Korb, Thomas St¨ utzle, and Thomas E Exner. An ant colony optimization approach to flexible protein–ligand docking. Swarm Intelligence, 1(2):115–134, 2007. 47 [Kuh12] Thomas S Kuhn. The structure of scientific revolutions. University of Chicago press, 2012. 249 [KV10] E Kremers and Pablo Viejo. A complex system modelling approach for descentralized simulation of electrical microgirds. In 15th IEEE International Conference on Enginerring of Complex Computer Systems, 2010. 190 [Lan90] C.G. Langton. Computation at the edge of chaos: Pahse transitions and emergent computation. Physica D: Nonlinear Phrenomena, 42:12–37, 1990. 188 278 BIBLIOGRAPHY [Lan01] Doug Laney. 3d data management: Controlling data volume, velocity and variety. META Group Research Note, 6, 2001. 70 [LB08] Lorenz Lutzenhiser and Sylvia Bender. The “Average American” unmasked: Social structure and differences in household energy use and carbon emissions. In 2008 ACEEE Summer Study on Energy Efficiency in Buildings, 2008. 150 [LC11] Ferry Lienert and Matthew Carson. Smart meter roll-out for the domestic sector (GB), 2011. 23 [LCRPS04] Sean Luke, Claudio Cioffi-Revilla, Liviu Panait, and Keith Sullivan. Mason: A new multi-agent simulation toolkit. In Proceedings of the 2004 SwarmFest Workshop, volume 8, 2004. 44 [Lev98] Simon A Levin. Ecosystems and the biosphere as complex adaptive systems. Ecosystems, 1(5):431–436, 1998. 67 [LG03] Ya’an Liu and Xiaohong Guan. Purchase allocation and demand bidding in electric power markets. Power Systems, IEEE Transactions on, 18(1):106–112, Feb 2003. 31 [LGRT11] Martin Lukasiewycz, Michael Glaß, Felix Reimann, and J¨ urgen Teich. Opt4J - A Modular Framework for Meta-heuristic Optimization. In Proceedings of the Genetic and Evolutionary Computing Conference (GECCO 2011), pages 1723–1730, Dublin, Ireland, 2011. 164 [LR04] Seok Won Lee and David C Rine. Case study methodology designed research in software engineering methodology validation. In SEKE, pages 117–122, 2004. 72 [LTDZ02] Marco Laumanns, Lothar Thiele, Kalyanmoy Deb, and Eckart Zitzler. Combining convergence and diversity in evolutionary multiobjective optimization. Evolutionary computation, 10(3):263–282, 2002. 118 279 BIBLIOGRAPHY [LV03] Justin Langseth and Nithi Vivatrat. Why proactive business intelligence is a hallmark of the real-time enterprise: Outward bound. Intelligent Enterprise, 5(18):34–41, 2003. 61 [LVC09] MD Lemmon, G Venkataramanan, and P Chapman. Using microgrids as a path towards smart grids. In Position paper presented at NSF workshop in cyber physical energy systems, Baltimore, MD, 2009. 36 [MA04] R.T. Marler and J.S. Arora. Survey of multi-objective optimization methods for engineering. Struct Multidisc Optim, 26, 2004. 118 [Mas13] Gilbert M Masters. Renewable and efficient electric power systems. John Wiley & Sons, 2013. 5, 28 [MBLA96] Nelson Minar, Roger Burkhart, Chris Langton, and Manor Askenazi. The swarm simulation system: A toolkit for building multi-agent simulations. Technical report, Santa Fe Institute Santa Fe, 1996. 44 [MCH+94] Melanie Mitchell, James P. Crutchfield, Peter T. Hraber, In G. Cowan, D. Pines, D. Melzner (editors, and Complexity Metaphors. Dynamics, computation, and the “edge of chaos”: A re-examination. In Complexity:Metaphors, Models, and Reality, pages 497–513. Addison-Wesley, 1994. 188 [ME04] Leon Messerschmidt and Andries Petrus Engelbrecht. Learning to play games using a pso-based competitive learning approach. Evolutionary Computation, IEEE Transactions on, 8(3):280– 288, 2004. 48 [ME10] Anthony R Metke and Randy L Ekl. Security technology for smart grid networks. Smart Grid, IEEE Transactions on, 1(1):99–107, 2010. 19 [Mey03] Niels I Meyer. European schemes for promoting renewables in liberalised markets. Energy policy, 31(7):665–676, 2003. 18 280 BIBLIOGRAPHY [MH11] Janine Morley and Mike Hazas. The significance of difference: Understanding variation in household energy consumption. In ECEEE 2011 summer study. Energy efficiency first: the foundation of a low carbon society, 2011. 150 [MHC93] Melanie Mitchell, Peter T. Hraber, and James P. Crutchfield. Revisiting the edge of chaos: Evolving cellular automata to perform computations. Complex Systems, 7:89–130, 1993. 188 [Mik01] Donald C Mikulecky. The emergence of complexity: science coming of age or science growing old? Computers & chemistry, 25(4):341–348, 2001. 68 [Mil05] F. Milano. An Open Source Power System Analysis Toolbox. IEEE Transaction on Power System, vol. 20, no.3, 2005. 35, 37 [Mis86] Jayadev Misra. Distributed discrete-event simulation. ACM Comput. Surv., 18(1):39–65, 1986. 86 [MJ12] Mark L Mitchell and Janina M Jolley. Research design explained. Cengage Learning, 2012. 72 [MOF02] OMG MOF. Omg meta object facility (mof) specification v1. 4, 2002. 52 [Mom09] James A Momoh. Smart grid design for efficient and flexible power networks operation and control. In Power Systems Conference and Exposition, 2009. PSCE’09. IEEE/PES, pages 1–8. IEEE, 2009. 18, 19 [MP10] M Moglia and P Perez. Modelling an urban water system on the edge of chaos. Environmental Modelling and Software, 25:1528– 1538, 2010. 189 [MPBL10] A. Monti, F. Ponci, A. Benigni, and J. Liu. Distributed intelligence for smart grid control. In Nonsinusoidal Currents and Compensation (ISNCC), 2010 International School on, pages 46–58, 2010. 32 281 BIBLIOGRAPHY [MS84] James L McKenney and Morton M Scott. Management decision systems: computer-based support for decision making. Harvard Business School Press, 1984. 60 [Neg04] Solomon Negash. Business intelligence. Communications of the Association for Information Systems, 13(1):177–195, 2004. 61, 62 [NN12] Gregoire Nicolis and Cathy Nicolis. Foundations of Complex Systems: Emergence, Information and Predicition. World Scientific, 2012. 41 [NNdGW09] DT Nguyen, M. Negnevitsky, M. de Groot, and C. Wang. Demand response in the retail market: Benefits and challenges. In Power Engineering Conference, 2009. AUPEC 2009. Australasian Universities, pages 1–6. IEEE, 2009. 5, 30 [Nwa02] C. Nwankpa. Voltage Stability Toolbox, version 2, Center for Electric Power Engineering, Drexel University, 2002. 35, 37 [Nyl99] A Nylund. Tracing the bi family tree. Knowledge Management, 1999. 60 [OEC] OECD. OECD Green Growth Studies: Energy. http: //www.oecd.org/greengrowth/greening-energy/ 49157219.pdf. 4, 20 [Off] Office of Electricity Delivery and Energy Reliability. Smart Grid. http://energy.gov/oe/services/ technology-development/smart-grid. 4, 20 [OLA97] OLAP Council. OLAP Definitions, 1997. 63 [Org11] Organisation for Economic Co-operation and Development. Green Growth Studies: Energy, 2011. 3 [Ott05] Gunnar Otte. Entwicklung und Test einer integrativen Typologie der Lebensf¨ uhrung f¨ ur die Bundesrepublik Deutschland. Con282 BIBLIOGRAPHY struction and test of an integrative lifestyle-typology for germany. Zeitschrift f¨ ur Soziologie, 34(6):442–467, December 2005. 150 [PAB12] Coralie Petermann, Soufian Ben Amor, and Alain Bui. A complex system approach for a reliable smart grid modeling. In KES, pages 149–158, 2012. 7 [Pac] Pacific Northwest National Laboratory. GridLAB-D. 38 [Par06] V Pareto. Manuale di economica politica, societa editrice libraria. Manual of political economy, 1971, 1906. 118 [PD11] P. Palensky and D. Dietrich. Demand side management: Demand response, intelligent energy systems, and smart loads. Industrial Informatics, IEEE Transactions on, 7(3):381–388, Aug 2011. 29, 31, 33 [PFR09] M. Pipattanasomporn, H. Feroze, and S. Rahman. Multi-agent systems in a distributed smart grid: Design and implementation. In 2009 IEEE/PES Power Systems Conference and Exposition, pages 1–8. IEEE, March 2009. 8, 191 [PHS+08] Fiona A. C. Polack, Tim Hoverd, Adam T. Sampson, Susan Stepney, and Jon Timmis. Complex systems models: Engineering simulations. In ALife XI, Winchester, UK, August 2008, pages 482–489. MIT Press, 2008. 33, 75 [PKZK08] P. Palensky, F. Kupzog, A. A. Zaidi, and Zhou Kai. Modeling domestic housing loads for demand response. In Industrial Electronics, 2008. IECON 2008. 34th Annual Conference of IEEE, pages 2742–2747, 2008. 8, 32, 38, 42, 145, 250 [PL06] Jukka Paatero and Peter Lund. A model for generating household electricity load profiles. International Journal of Energy Research, 30(5):273–290, 2006. 150 [PM98] Katherine N Probst and Michael H McGovern. Long-term stewardship and the nuclear weapons complex: The challenge ahead. Resources for the Future, 1998. 19 283 BIBLIOGRAPHY [Pol08] Michael G Pollitt. The future of electricity (and gas) regulation in a low-carbon policy world. Energy Journal, 2008. 18 [Poo87] R J Pooley. An introduction to programming in SIMULA. Blackwell Scientific Publications, Ltd., Oxford, UK, UK, 1987. 86 [Poo01] John D. Poole. Model-Driven Architecture: Vision, Standards and Emerging Technologies. In Workshop on Metamodeling and Adaptive Object Models, ECOOP, 2001. 50 [Pow] Powertech. TSAT - Transient Security Assessment Tool. 37 [Pow99] D.J. Power. Decision Support Systems Glossary. DSS Resources, 1999. 63 [Pow02] Daniel Power. Decision support systems: concepts and resources for managers. Greenwood Publishing Group, 2002. 60, 61 [Pow07] Daniel J Power. A brief history of decision support systems. DSSResources. COM, World Wide Web, http://DSSResources. COM/history/dsshistory. html, version, 4, 2007. 60, 61 [PRS08] D Pudjianto, C Ramsay, and G Strbac. Microgrids and virtual power plants: concepts to support the integration of distributed energy resources. Proceedings of the Institution of Mechanical Engineers, Part A: Journal of Power and Energy, 222(7):731– 741, 2008. 24 [RBK11] Grzegorz Rozenberg, Thomas Bck, and Joost N Kok. Handbook of natural computing. Springer Publishing Company, Incorporated, 2011. 12 [RC03] Ken Rudin and David Cressy. Will the real analytic application please stand up? DM REVIEW, 13:30–41, 2003. 62 [RCO09] N. Ruiz, I. Cobelo, and J. Oyarzabal. A direct load control model for virtual power plant management. Power Systems, IEEE Transactions on, 24(2):959–966, May 2009. 33 284 BIBLIOGRAPHY [RCV07] M Rosas-Casals and S Valverde. Topological vulnerability of the European power grid under errors and attacks. International Journal of Bifurcations and Chaos, 17:2465–2475, 2007. 191 [REE] Red El´ ectrica de Espa˜ na. http://www.ree.es/es/. 29 [RH97] Wendy Robson and Prentice Hall. Strategic management and information systems: an integrated approach, 1997. 71 [RKU11] Anand Rajamaran, Jure Keskovec, and Jeffrey D. Ullman. Mining of Massive Datasets. Cambridge University Press, 2011. 110 [RLJ06] S. F. Railsback, S. L. Lytinen, and S. K. Jackson. Agent-based Simulation Platforms: Review and Development Recommendations. SIMULATION, 82(9):609–623, September 2006. 43, 44 [RMMG08] Sylvain Rougemaille, Fr´ ed´ eric Migeon, Christine Maurel, and Marie-Pierre Gleizes. Model driven engineering for designing adaptive multi-agents systems. In Engineering Societies in the Agents World VIII, pages 318–332. Springer, 2008. 56 [Ros77] Robert Rosen. Complexity as a system property. International Journal Of General System, 3(4):227–232, 1977. 68 [RPT07] Saifur Rahman, Manisa Pipattanasomporn, and Yonael Teklu. Intelligent distributed autonomous power systems (idaps). In Power Engineering Society General Meeting, 2007. IEEE, pages 1–8. IEEE, 2007. 21 [RVRJ11] Sarvapali Ramchurn, Perukrishnen Vytelingum, Alex Rogers, and Nick Jennings. Agent-Based Control for Decentralised Demand Side Management in the Smart Grid. In The 10th International Conference on Autonomous Agents and Multiagent Systems-Volume 1, pages 5–12, Taipei, Taiwan, 2011. 8, 32, 33 [RWE11] RWE Rhein-Ruhr Verteilnetz. Lastprofile, 2011. 143 [SAA+02] Dag IK Sjoberg, Bente Anda, Erik Arisholm, Tore Dyba, Magne Jorgensen, Amela Karahasanovic, Espen Frimann Koren, and 285 BIBLIOGRAPHY Marek Vok´ ac. Conducting realistic experiments in software engineering. In Empirical Software Engineering, 2002. Proceedings. 2002 International Symposium n, pages 17–26. IEEE, 2002. 72, 243 [SC05] Margarita Reyes Sierra and Carlos A Coello Coello. Improving PSO-based multi-objective optimization using crowding, mutation and ? -dominance. In In EMO?2005, pages 505?519. LNCS 3410, pages 505–519. Springer-Verlag, 2005. 119, 164, 168 [SCC02] William R.. Shadish, Thomas D Cook, and Donald Thomas Campbell. Experimental and quasi-experimental designs for generalized causal inference. Wadsworth Cengage learning, 2002. 72 [SCF+06] Jean-S´ ebastien Sottet, Ga¨ elle Calvary, Jean-Marie Favre, Jo¨ elle Coutaz, Alexandre Demeure, and Lionel Balme. Towards model driven engineering of plastic user interfaces. In Satellite Events at the MoDELS 2005 Conference, pages 191–200. Springer, 2006. 56 [Sch10] Thomas Schwarz. Lebensstile und W¨ ahlerverhalten in Stuttgart. Ergebnisse der Stuttgarter Lebensstilbefragung 2008. Statistik und Informationsmanagement, July 2010. 151, 152 [SD06] Mohsen Amini Salehi and Hossein Deldari. Grid load balancing using an echo system of intelligent ants. In Parallel and Distributed Computing and Networks, pages 47–52, 2006. 46 [Sei03] Edwin Seidewitz. What models mean. IEEE software, 20(5):26– 32, 2003. 51 [SGRM07] Frank Sensuß, Massimo Genoese, Mario Ragwitz, and Dominik M¨ ost. Agent-based simulation of electricity markets-a literature review. Energy Studies Review, 15(2):2, 2007. 36, 191 286