D3.3: Methodologies and frameworks to optimize demand response and peak load shaving for 4th-5th DHC in three typical European climates

HYPERGRYD. This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 101036656 WP3 – ICT Modules and Simulation Tools Task 3.2 - Development of demand response management, prediction, and optimisation methods for 4th and 5th DHC networks D3.3 - Methodologies and frameworks to optimize demand response and peak load shaving for 4th-5th DHC in three typical European climates. Ref. Ares(2024)6909400 - 30/09/2024 D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 2 DISCLAIMER The opinion stated in this report reflects the opinion of the authors and not the opinion of the European Commission. All intellectual property rights are owned by HYPERGRYD consortium members and are protected by the applicable laws. Reproduction is not authorised without prior written agreement. The commercial use of any information contained in this document may require a license from the owner of that information. ACKNOWLEDGEMENT This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement Nº 101036656. D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 3 Project Project Acronym HYPERGRYD Project Title Hybrid coupled networks for thermal-electric integrated Smart Energy Districts Grant Agreement number 101036656 Call identifier H2020-LC-GD-2020 Topic identifier LC-GD-2-1-2020 Innovative land-based and offshore renewable energy technologies and their integration into the energy system Funding Scheme Research and Innovation Action Project duration 42 months (From 1 October 2021) Coordinator ARCbcn Website http://hypergryd.eu Deliverable Deliverable No. 3.3 Deliverable title Methodologies and frameworks to optimize demand response and peak load shaving for 4th-5th DHC in three typical European climates Description D3.3 (Methodologies and frameworks to optimize demand response and peak load shaving for 4th-5th DHC in three typical European climates) summarize the main activities in Task 3.2 (Development of demand response, prediction, and optimisation methods for 4th 5th DHC networks) Most power-to-heat solutions operate independently, so this task aims to integrate demand response at a higher level, optimizing storage (buildings, substations, piping networks) using surplus or low-exergy heat/cooling sources. It will develop techniques for energy forecasting and dynamic control strategies, using machine learning methods like Artificial Neural Networks and Genetic Algorithms. KTH will apply advanced ML techniques for RES models (waste heat, P2H, CHP, etc.), with DHC models using weather data from partners (SONNE, ENVI, EURAC). ML methods will be evaluated for computational cost, privacy, and accuracy. GSY and ENCO will test and integrate the developed algorithms. WP No. WP3 Related task Task 3.2 - Development of demand response, prediction, and optimisation methods for 4th 5th DHC networks Lead Beneficiary 3 - KTH Author(s) Mustapha Habib (KTH), Qian Wang (KTH) Contributor(s) Jorge Leao Pirocchi (IDP) Type R Dissemination PU Language English – GB Due 30/09/2024 Submission date 30/09/2024 Version Date Authors Description V.0.1 13/09/2024 Mustapha Habib (KTH) First version for internal reviewing V.0.2 27/09/2024 David Verez (ARC) Review V.1.0 28/09/2024 Jorge Leao Pirocchi (IDP) Review V.2.0 28/09/2024 Mustapha Habib (KTH) Final version after review D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 4 Table of Contents 1. Executive Summary .................................................................................................... 7 2. Introduction ................................................................................................................ 9 2.1. Scope ......................................................................................................................... 9 2.2. Audience ................................................................................................................... 9 2.3. Abbreviations ............................................................................................................ 9 2.4. Contributions of partners ....................................................................................... 10 2.5. Relation to other activities ..................................................................................... 11 2.6. Structure ................................................................................................................. 11 3. Methodology ............................................................................................................ 12 3.1. Machine and deep learning methods applied in this research. ............................. 12 3.1.1. Data clustering ........................................................................................................ 12 3.1.1.1. Euclidean distance measure ................................................................................... 14 3.1.1.2. DTW distance measure ........................................................................................... 14 3.1.2. DHN heat power prediction using neural networks ............................................... 15 3.1.2.1. Multi-layer perceptron ........................................................................................... 16 3.1.2.2. Training ................................................................................................................... 18 3.2. Metaheuristic optimisation applied on DHN-HP .................................................... 19 3.2.1. System modelling and validation ............................................................................ 20 3.2.2. Optimisation problem formulation ........................................................................ 21 3.2.3. Particle Swarm Optimisation .................................................................................. 22 3.3. Optimal control of HP-RES-TES within DHN. .......................................................... 23 3.3.1. System modelling .................................................................................................... 24 3.3.1.1. Heat pump and energy storage system. ................................................................. 24 3.3.1.2. state of charge modelling of the battery energy storage system .......................... 24 3.3.1.3. Nonlinear optimisation ........................................................................................... 25 4. Method validation within use-cases ........................................................................ 26 4.1. Nordic climate – data clustering to handling low-quality data from DHN meters .. 26 4.1.1. Time-series clustering .............................................................................................. 27 4.1.1.1. Nursing homes data clustering ................................................................................ 27 4.1.1.2. Schools data clustering ............................................................................................. 29 4.2. Load prediction ........................................................................................................ 31 4.2.1. Long-term assessment ............................................................................................. 31 D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 5 4.2.2. Seasonal assessment ................................................................................................ 34 4.3. Continental/Alpine Climate – optimal demand response management in DHN at a community level ..................................................................................................................... 37 4.3.1. Use case description. ................................................................................................ 37 4.3.2. Simulation results ..................................................................................................... 39 4.4. Mediterranean Climate – optimal control of heat pump with a photovoltaic/storage system in centralized DHN substation ................................................................................... 41 4.4.1. Use-case description ................................................................................................ 41 4.4.2. Simulation results ..................................................................................................... 43 4.4.2.1. System parameters and method validation environment ....................................... 43 Conclusions ............................................................................................................................. 48 Summary of achievements .................................................................................................... 48 Relation to continued developments. ................................................................................... 49 References .............................................................................................................................. 50 List of Figures Figure 1 Data collection and analysis flowchart. ................................................................................. 13 Figure 2 Different aligning rules for (a) Euclidean distance and (b) DTW distance matching. ........... 15 Figure 3 Example of an MLP model structure. .................................................................................... 17 Figure 4 DH/HP and TES connection topology in the proposed substation scheme. ......................... 19 Figure 5 Simple overview of the connection topology of HESS in the centralized heating substation. ............................................................................................................................................................. 24 Figure 6 Clustered datasets for nursing homes for the years 2017, 2018, and 2019 using Euclidean kmeans. ................................................................................................................................................. 28 Figure 7 Clustered datasets for nursing homes for the years 2017, 2018, and 2019 using DTW k-means. ............................................................................................................................................................. 29 Figure 8 Clustered datasets for schools for the years 2017, 2018, and 2019 using Euclidean k-means. ............................................................................................................................................................. 30 Figure 9 Clustered datasets for schools for the years 2017, 2018, and 2019 using DTW k-means. ... 31 Figure 10 Annual prediction assessment using a school dataset extracted from Cluster 1 (a) using Euclidean k-means (b) using DTW k-means. ....................................................................................... 32 Figure 11 Annual prediction assessment using a school dataset extracted from Cluster 2 (a) using Euclidean k-means (b) using DTW k-means. ....................................................................................... 32 Figure 12 Annual prediction assessment using a nursing home dataset extracted from Cluster 1 (a) using Euclidean k-means (b) using DTW k-means. .............................................................................. 33 D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 6 Figure 13 Annual prediction assessment using a nursing home dataset extracted from Cluster 2 (a) using Euclidean k-means (b) using DTW k-means. .............................................................................. 34 Figure 14 Prediction performance for a school dataset (a) in winter (b) in summer. ........................ 35 Figure 15 Prediction performance for a nursing home dataset (a) in winter (b) in summer.............. 35 Figure 16 Prediction performance in mild-season (a) for a nursing home (b) for a school ................ 36 Figure 17 Satellite view of Großschönau community with buildings considered for optimal energy coordination. ....................................................................................................................................... 38 Figure 18 Energy optimisation flowchart. ........................................................................................... 38 Figure 19 Heat demand with PV power production of the concerned buildings. .............................. 40 Figure 20 HP optimal condenser temperature setpoint with associated TES SOC variation in a 2-day simulation (user-centric scenario) ....................................................................................................... 40 Figure 21 HP optimal condenser temperature setpoint with associated TES SOC variation in a 2-day simulation (community-centric scenario) ........................................................................................... 41 Figure 22 Electric power distribution in the ENVIPARK research facility. ........................................... 42 Figure 23 Heating distribution network in ENVIPARK. ........................................................................ 42 Figure 24 Simulation conditions (a) heat load (b) PV power (c) electricity price. ............................... 44 Figure 25 HP condenser temperature setpoint: (a) with flexibility Scenario 1 (b) with flexibility Scenario 2. ........................................................................................................................................... 45 Figure 26 BESS power setpoint: (a) with flexibility Scenario 1 (b) with flexibility Scenario 2. ............ 45 Figure 27 HP supplied heat: (a) with flexibility Scenario 1 (b) with flexibility Scenario 2. .................. 46 Figure 28 BESS SOC variation: (a) with flexibility scenario 1 (b) with flexibility scenario 2. ............... 47 Figure 29 TES maximum water temperature: (a) with flexibility Scenario 1 (b) with flexibility Scenario 2. .......................................................................................................................................................... 47 List of Tables Table 1 MLP training parameters ........................................................................................................ 19 Table 2 Statistic measures of the generated normalized datasets in cluster-wise. ............................ 31 Table 3 MSE and MAE evaluation using nursing homes normalized data for the year 2019. ............ 37 Table 4 Simulation parameters ........................................................................................................... 39 Table 5 PSO parameters ...................................................................................................................... 39 Table 6 System parameters ................................................................................................................. 43 Table 7 Flexibility impact assessment of different optimisation methods. ........................................ 48 D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 7 1. Executive Summary The goal of the HYPERGRYD project is the development of a set of replicable and scalable costeffective technical solutions to allow the integration of Renewable Energy Sources (RES) with different dispatchability and intrinsic variability inside Thermal Grids as well as their link with the Electrical Grids, including the development of innovative key components, in parallel with innovative and integrated ICT services formed by a scalable suite of tools for the proper handling of the increased complexity of the systems from building to Local Energy Community (LEC) levels and beyond, and accelerate the sustainable transformation, planning, and modernization of District Heating and Cooling (DHC) towards 4th and 5th generation. HYPERGRYD also aims to develop real-time management of electrical and thermal energy flows in the coupled energy network complex, including the synergies between them. Therefore, HYPERGRYD aims at three over-arching General Objectives: • To prove Smart Energy Networks as the future of Efficient Energy Management in DHC in synergy with the Electrical Grids in LEC/Smart Cities of the future; • To define the roadmap to design and plan future DHC as well as the modernization of the existing ones in different climates and RES penetration levels toward 4th-5th generation, • To demonstrate HYPERGRYD RES-based Enabling Technologies, Smart Energy Grid Solutions empowered by new ICT tools and services as the key for this evolution. During the project, HYPERGRYD’s solutions will be implemented across four Live-In-Labs cases in three representative climates, with special consideration to their cost effectiveness and potential replicability to finally achieve these three main objectives. One of the main solutions proposed by HYPERGRYD is the exploitation of DHN data, typically generated by smart sensors and energy meters, to come up with applicable tools that enable the implementation of optimal energy management of modern DH networks that incorporate renewables and low-temperature supply. This will open the path towards a smooth transition of actual DH networks to 4th and 5th generation DH. The purpose of this deliverable in short is to: • Develop different machine learning techniques designed for data clustering to deal with lowquality data issued from sensors and energy meters installed in DHN substations, and to identify abnormal patterns and profiles. • Develop a deep learning technique for long-term prediction of DH load in different building categories. • Design and simulate an optimal coordinated demand side management based on disaggregated data for a DH-powered community. • Design and simulate a centralized demand-side management based on aggregated data applicable to the main DHN substation in the presence of different energy storage technologies, renewables, and heat pumps. D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 8 • Validate the abovementioned techniques on real datasets generated in three different European climates. On behalf of Authors Mustapha Habib, KTH D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 9 2. Introduction 2.1. Scope This deliverable outlines the development and validation of data-driven techniques for optimizing energy usage in DHN-powered buildings and communities, leveraging data from smart meters, sensors, and building management systems. It explores how such data can reveal end-user behaviour to inform advanced control and management policies. The scope includes clustering and prediction methods, combined with metaheuristic and deterministic optimisation algorithms, to achieve optimal control and demand management. The solutions presented capitalize on energy flexibility within DHN communities, offering both distributed control via advanced IoT communication and centralized control based on aggregated data at the main DHN substation. In the scope of these techniques, this deliverable proposes a set of tools that can serve in upgrading the current DHN toward the 4th and 5th generations. 2.2. Audience This deliverable provides outcomes that can be applicable inputs for Task 3.3 (Development of novel machine learning techniques driven by IoT for LEC-integrated DHC networks), in addition to the activities in WP5 concerning demonstration and validation strategies in demo sites. 2.3. Abbreviations RES: Renewable Energy Sources LEC: Local Energy Communities DHN: district heating network DC: Direct current AC: Alternative current ML: Machine learning DL: Deep learning ESS: energy storage system TES: Thermal energy storage BESS: Battery energy storage system HP: Heat pump IoT: Internet of things D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 16 One model is to be created for each cluster (developed in 3.1.1). In addition to outdoor weather conditions, heat energy demand in buildings can be affected by many other factors. In this work, the chosen data inputs, for each hour, are listed below: • The forecasted outdoor temperature (in °C) is typically accessible via an online service. • Hour of the day: from 1 representing "1:00-2:00" to 24 representing "23:00-00:00", • Day of the week: an integer, from 1 for Monday to 7 for Sunday, • Month of the year, an integer, from 1 for January to 12 for December, • Holiday indicator, a Boolean, 1 for a holiday and 0 for a working day. Those parameters are set basically to give an estimation of the heat load trend, which changes essentially according to different timeslots. They can also give an approximate assessment of the occupancy rate and human activities related to the building type (M.M. Alam et al., 2017). As the analysed buildings are not residential, it can be possible to find some indicators having correlations with the occupancy rate with an accepted approximation, which is the reason behind the choice of the last four ANN inputs listed above. In this context, "hour of the day" can give an approximate guess of the daily scheduled work time for schools and nursing homes. In the same way, "day of the week" has been introduced to differentiate between working days and weekends, which is more significant in the case of schools as nursing homes still have partial work activities on weekends. With the same impact, the inputs "month of the year" and "holidays indicator" have been introduced to highlight vacation periods. 3.1.2.1. Multi-layer perceptron A Multi-Layer Perceptron (MLP) is a type of feedforward artificial neural network that consists of multiple layers of interconnected nodes, called neurons or units. It is a foundational and widely used architecture in the field of deep learning [4-5]. The term "perceptron" refers to a basic building block that models a single neuron, and "multi-layer" indicates that multiple layers of these perceptrons are stacked together. In an MLP, the neurons are organized into layers, typically consisting of an input layer, one or more hidden layers, and an output layer. The input layer receives the input data, and the output layer produces the final output of the model. The hidden layers are intermediate layers between the input and output layers, responsible for processing and transforming the input information through a series of non-linear operations. Figure 3 shows an illustrative architecture of an MLP with an input layer of five neurons; three hidden layers with 12 neurons each and one output layer with one single neuron. D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 17 Figure 3 Example of an MLP model structure. Each neuron in the MLP is associated with a set of learnable parameters, including weights and biases. The weights determine the strength of the connections between neurons, while the biases introduce an offset or bias term to the neuron's output. These parameters are adjusted during the training process to optimize the performance of the MLP on a given task. The functioning of an MLP involves two main steps: forward propagation and backward propagation. In forward propagation, the input data is fed through the network, and the activations of each neuron are computed layer by layer, ultimately producing the output. Backpropagation involves computing the gradients of the network's parameters concerning a chosen loss function, allowing for the adjustment of the weights and biases in a direction that minimizes the loss. This iterative process of forward and backward propagation is repeated until the model converges to an optimal state. Eq (4) formulates the forward process through MLP architecture. 1 ( ) ( ) ( ) ( 1) 01 ( ) ( ( )) L U L L L L ii i f X a w w f X −− = =+  (4) Where: L is the layer number. ()L f is the L function layer. () 0L w and ()L i w are respectively bias and weight vectors corresponding to layer L . X is the input vector. a is the activation function (it can represent different activation functions between different layers). D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 18 1L U− is the number of layers before L. The chosen activation function is the Sigmoid function, it is formulated as: 1 () 1x ax e− =+ (5) 3.1.2.2. Training During the training phase, an optimisation algorithm is used to update the weights and biases of the MLP to minimize a loss function, which is formulated in (6). As an optimisation method, this study addresses the use of Limited-memory Broyden-Fletcher-Goldfarb-Shanno (LBFGS). It is an optimisation algorithm used for solving unconstrained optimisation problems. The algorithm combines the BFGS method with limited memory storage, allowing it to approximate the inverse Hessian matrix without explicitly computing or storing it. The limited-memory aspect of LBFGS makes it memory-efficient and suitable for large-scale optimisation problems. However, the detailed mathematical description of LBFGS is outside the scope of this article. (6) Where and are the actual and predicted DH load respectively, P N is the number of data samples. The method "neural_network.MLPRegressor", which is a part of the Python package "sklearn", was chosen as an implementation tool. The training parameters are set in Table 1, and those parameters are reached after some trial-and-error tests. As training data, two datasets for hourly DH heat consumption for 2017 and 2018 of nursing homes and educational buildings located in Trondheim, Norway are used (more information about the used datasets will be given in 4.1.1). The pseudocode for a trained MLP is summarized below: Initializing (1) 0 w and Acquiring training data X While (training stopping criteria is not met){ Calculating the loss function Updating weights and bias for all layers MLP feedforward} (1) i w D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 19 Table 1 MLP training parameters MLP parameter value Number of neurons in the input layer 5 Number of hidden layers 5 Number of neurons in one hidden layer 50 number of neurons in the output layer 1 Activation function Sigmoid Optimizer LBFGS Learning rate Constant (0.1) Loss function MSE 3.2. Metaheuristic optimisation applied on DHN-HP In this study, we assume a virtual scenario, in which, we have a residential community connected to DHN, in which, HPs with TES are installed at each building. The purpose behind this assumption is to assess the impact of the maximum energy flexibility offered by HP-TES in the optimal exploitation of the DHN energy. However, in reality, this assumption may not always be valid. the proposed method consists of adjusting at each time step (one hour), in coordination, the state of charge (SOC) of TES units in all buildings in the DHN community by varying the corresponding HP condenser temperatures. For method validation, this section addresses the development of mathematical models necessary for simulation, as well as the optimisation problem formulation. As stated earlier, a virtual scenario is tested, in which, all buildings are connected to the same DHN network via HP/TES units. The connection topology assumes that HPs are used as DH temperature boosters as described in Figure 4. Therefore, the developed methods can be applied as an assessment framework toward the integration of 4th and 5th generation DH (4GDH and 5GDH) which is characterized by low-temperature supply with high penetrations of local RES. Figure 4 DH/HP and TES connection topology in the proposed substation scheme. For a successful method implementation, a few simplifying assumptions are proposed as below: D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 20 The DHN temperature supply is reduced by 20 % (compared to the current temperature level) to simulate 5GDH. 1. All HP units are equipped with variable-speed compressors. 2. The possibility for the buildings connected to the same low-voltage (LV) grid to share their extra PV powers (inter-community). 3. The only considered electricity consumption (for optimisation) is the one related to the HP compressor's operation. 3.2.1. System modelling and validation For HPs, the coefficient of performance (COP) is defined based on the source and sink temperatures as described in (7), which, source T and sink T are the source (DHN supply) and sink (condenser) temperatures in HP respectively [Kelvin]. 𝜂𝐻𝑃 is an empirical coefficient that calibrates the ideal COP to the real one, in our case, it is fixed at 0.15. sin sin s k HP k ource T COP TT  =− (7) The consumed electricity _el HP P [kW], related to a supplied HP heat _HP out Q [kW] is formulated in (8). It is mainly related to compressor operation. _ _ HP out el HP Q PCOP = (8) Ideally, the heat consumed by HPs in the evaporation process is formulated in (9) (neglecting heat wastes in the HP units): _ _ _HP in HP out el HP Q Q P=− (9) As it will be explained later, _HP in Q is nothing but the heat extracted from DH network [kW]. Regarding TES, a stratified multi-node model for water tanks is adopted. The differential equation for the evaluation of water temperature in one node i in the water tank is formulated in (10): , , , , , i i p IN i OUT i COND i CONV i LOSS i dT mC Q Q Q Q Q dt = − + + − (10) Where ,IN i Q and ,OUT i Q are respectively the supplied and extracted heat at the node i [kW]. In this study, this applies only to the first node ( 1i= ); i m is the water mass of layer i [kg]; Cp is the water specific heat (4182 J/kg.°C). ,CONV i Q is the heat power gained or lost via convection mode at the node i . it is formulated as below [kW]: D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 21 ,.( ). .( ( 1) ( )) .( ). .( ( ) ( 1)) CONV i UP IN OUT DOWN OUT IN Q m m Cp T i T i m m Cp T i T i  = − − − − − − + (11) Where IN m and OUT m are respectively the inlet and outlet water flows [kg/s]; UP  and DOWN  are respectively the flow direction indicators: 1 0 IN OUT UP if m m else    =  , 1 0 OUT IN DOWN if m m else    =  (12) ,COND i Q is the heat transferred to or from layer i via conduction mode [kW], it is formulated as below: , .. ( ( 1) ( )) ( ( ) ( 1)) COND i k A k A Q T i T i T i T i dX dX = − − − − + (13) Where k is the thermal conductivity of water (0.598 W/m.K), A is the cross-section area of the water [m2]; dX is the model node thickness [m]. ,LOSS i Q is the heat loss through the tank metal wall in layer i [kW], it is formulated as below: ,. .( ( ) ) LOSS i AMB Q kwU T i T=− (14) Where kw is thermal conductivity of the wall (2.5 J/m2.°C); U is the water node lateral surface [m2]; and AMB T is the ambient temperature, chosen 23 °C. To update the water temperature at each node, the differential equation is solved numerically using the equation below: , , , 2 ( ) ( ) . . . . COND i CONV i LOSS i t T i Q Q Q Cp R dX    = + − (15) Where ()Ti is the temperature incrementation at layer i [° C];  is the water density [kg/m3]; R is the diameter of the water cross-section area in the tank [m]. 3.2.2. Optimisation problem formulation As a close approximation to the TES SOC [%], in this study, the average water temperature across the tank height is used as described in (16), in which MAX T is the maximum temperature that can the storage tank be operated with (fixed at 65 °C); N is the number of nodes in the tank model. 1 100. . N i i MAX T SOC NT = = (16) The heat supplied by HPs to TES units is formulated as below: D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 22 . .( (0)) (0) 0 p HP COND HP COND HP OUT mC T T if T T Qelse −− − −  =  (17) Where HP COND T− is the HP condenser setpoint, to be defined by the optimizer at each time step [° C]; (0)T is water temperature at the TES upper layer ( 0i= ) [° C]. When the optimizer wants to turn off the HP, it sends condenser temperatures setpoints lower than the source temperature (55 °C) to prevent heat flow reverse in simulation. Optimisation has two main weighted terms in the objective function formulated in (18): one related to the energy cost reduction at each building, and one related to the whole community as one entity. 2 _ , , , , _ , , , , , , 1 1 1 1 1 1 ( ). ( ) PV BN N H N H M el HP i t PV i t t el HP i t PV i t loss j t t i t i i j J P P EP P P P  = = = = = = = − + − −      (18) Where: H is time optimisation time horizon in hours [hour]; B N is the number of concerned buildings; ,,PV i t P is the supplied energy from the PV system installed on building i . i at the time t [kWh] (for buildings having no PV systems, this variable is set to zero); t EP is the electricity price at the time t [€/kWh], it has two possible values: one related to buying price and one to selling one; PV N is the number of installed PV systems; ,,loss j t P is the electric loss of the LV connection cables j at the time t [kW];  and  are weighting coefficients. A set of constraints are to be respected during the optimisation, which are the minimum and maximum SOC limits of each TES, and the HP condenser temperature setpoint as formulated below: MIN MAX SOC SOC SOC (19a) HP COND MIN HP COND HP COND MAX T T T − − − − −  (19b) According to the values of the weighting coefficients, the optimisation criteria can be defined, either by going toward reducing the electricity cost for each consumer individually (by increasing  ) or toward maximizing the energetic independence of the entire community by sharing available PV energies between different buildings (by increasing  ). For the last objective function term, the electric power losses between different buildings are to be taken into consideration, therefore, information about the internal low-voltage cabling inside the community is needed. 3.2.3. Particle Swarm Optimisation Particle Swarm Optimisation (PSO) is selected as the optimisation technique. It was developed by Eberhart and Kennedy (Kennedy and Eberhart 1995), inspired by the social behaviour of bird flocking or fish schooling. PSO shares many similarities with evolutionary computation techniques such as Genetic Algorithms. It is suitable for optimisation problems with a high number of decision D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 23 variables and flexible objective functions, which is the use-case of this study. The system is initialized with a population of random values of solutions and searches for optima by updating generations. However, unlike GA, PSO has no evolution operators such as crossover and mutation. In PSO, the potential solutions, called particles, fly through the problem space by following the current optimum particles. Detailed information will be given in the following sections. PSO is initialized with a group of random particles (solutions) and then searches for optima by updating generations. In every iteration, each particle is updated by the following two "best" values: the first one is the best solution (fitness) it has achieved so far, and the value is called pbest . Another "best" value that is tracked by the particle swarm optimizer is the best value obtained so far by any particle in the population. This best value is a global best and called gbest . After finding the two best values, the particle updates its velocity and positions with the following equation (20a) and (20b): ( ) ( ) 1 1 2 2 ( 1) ( ) . ( ) ( ) . ( ) ( )v i v i c r pbest i p i c r gbest i p i+ = + − + − (20.a) ( 1) ( ) ( 1)p i p i v i+ = + + (20.b) Where i the number of iterations; ()vi is the particle velocity; ()pi is the current particle (solution); ()pbest i and ()gbest i are the best positions for the particle and the group respectively. 1 r and 2 r is a random number between 0 and 1. 1 c and 2 c are learning factors. 3.3. Optimal control of HP-RES-TES within DHN. This subsection presents and explains the optimisation-based control of the hybrid energy storage system (HESS) for local energy community (LEC) management. First, the proposed hybridization architecture is presented and explained. Then, the mathematical models of the involved systems in HESS are defined, which will be the basis for the simulation study later. Using the developed models, the nonlinear optimisation algorithm will be formulated. Finally, the use-case specifications used for validation will be described, highlighting the virtual scenario that is going to be implemented in the simulation. Proposed HESS Hybridization Architecture Based on the use-case specifications, which will be described at the end of this section, this study proposes the hybridization architecture shown in Figure 5. It consists of TES fed by an air-source HP. The HP compressor is powered by a combination of a battery energy storage system (BESS) and photovoltaic (PV) system in addition to the low-voltage (LV) grid. Two main controllable assets are adopted to implement the optimisation-based control of HESS, which are HP and BESS. For the first one, a variable speed compressor is needed to adjust continuously the refrigerant flow, and consequently, the HP sink temperature level. For the second one, a DC-to-AC inverter/charger is needed to interface BESS with the local AC bus and control the charging/discharging rates. However, D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 24 this study does not include any optimal sizing of these sub-systems, and the related investment cost is excluded from the final evaluation. Figure 5 Simple overview of the connection topology of HESS in the centralized heating substation. 3.3.1. System modelling 3.3.1.1. Heat pump and energy storage system. For this task, the same mathematical approaches explained in subsections 3.1.3.1 for HP and TES, are adopted. 3.3.1.2. state of charge modelling of the battery energy storage system BESS SOC level for the next time interval is estimated via the following equation: . ( ) ( ) 100. B IT SOC t T SOC t C  + = − (21) ()SOC t is the measured value [%] while ()SOC t T+ is the predicted one at the time horizon T ; C is the BESS rated capacity [Ah] and B I is the running current [A] which is calculated thanks to the next simple equation: D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 25 B B B P IV = (22) B V is the BESS voltage [72 V], for simplicity it is considered constant (three 24V units in series); B P is the exchanged BESS power [W], it takes positive value in case of charging and negative values in case of discharging. For simplicity, BESS self-discharge process is not modelled. 3.3.1.3. Nonlinear optimisation In this study, Sequential Quadratic Programming (SQP) is chosen as an optimizer for the optimisation problem, it is an approach that starts with an initial guess for the decision variables x . At each time step, it forms a quadratic approximation of the Lagrangian, which is the objective function ()fx combined with a weighted sum of the constraint functions ()hx and ()gx , and find the optimal solution in an iterative way (Boggs, P. and Tolle, J, 1995). 1 min ( ) 2 . . ( ) 0 ( ) 0 TT f x x Qx F x st g x hx =+  = (23) ( , ) ( ) ( ) ( ) TT L x f x h x v g x  = − − (24) Q and F are respectively the quadratic cost matrix and the linear cost vector;  is a vector of Lagrangian multipliers for the equality constraints; v is a vector of Lagrangian multipliers for the inequality constraints. The problem is then shifted from minimizing ()fx to minimizing ( , , )L x v  which holds the objective function along with the constraints. The optimisation algorithm iteratively updates the decision variables x , Lagrange multipliers  and v until convergence. The goal is to find an x that minimizes the Lagrangian subject to the constraints. The objective function adopted in this study is formulated as follow: ,,, 22 ,2 ,, .( ) ( ) ( ) () HP i i B i PV i i TH i MAX i MAX MIN i MIN HP i B i PV i i TH i Q EP P P if EP EP COP J T T T T QP P if EP EP COP  − −    = + − + −  − −    (25) Subject to the hard constraint formulated below: ,. 100. Bi MIN i MAX IT SOC SOC SOC C   −  (26) D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 32 The datasets extracted from Cluster 2 for schools, generated by both clustering methods, show relatively less modelling capability, which is translated to the average prediction performance summarized in Figure 11. Figure 10 Annual prediction assessment using a school dataset extracted from Cluster 1 (a) using Euclidean k-means (b) using DTW k-means. Figure 11 Annual prediction assessment using a school dataset extracted from Cluster 2 (a) using Euclidean k-means (b) using DTW k-means. Mainly due to the different occupants' activity schedules, the heat load annual profile for nursing homes is less fluctuating, as Figure 12 shows, compared to schools. This is because DH load profiles D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 33 for schools are much sharper, with distinct night setbacks and weekday/weekend shifts than those for nursing homes. Moreover, heat loads in summer for schools are moderately lower than those for nursing homes, which is highlighted by the black dashed rectangles in Figure 10 and Figure 11. The reason is that no official scheduled activities are carried out during the summer in schools, which is not the case for nursing homes. Cluster 2, issued by the two studied clustering techniques for nursing home data, shows different patterns compared to Cluster 1 (see Figure 13). In this case, the annual heat load for the corresponding buildings is considerably lower. Low SH demands for these buildings can represent one potential justification. Moreover, Cluster 2 also includes some data anomalies, such as the outlier value in Figure 13 (a) and the measurement gap in Figure 13 (b), which are both highlighted by black dashed rectangles. Figure 12 Annual prediction assessment using a nursing home dataset extracted from Cluster 1 (a) using Euclidean kmeans (b) using DTW k-means. D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 34 Figure 13 Annual prediction assessment using a nursing home dataset extracted from Cluster 2 (a) using Euclidean kmeans (b) using DTW k-means. 4.2.2. Seasonal assessment To gain a deep insight into the prediction process in different seasonal periods, three-week scenarios have been selected: cold, warm, and mild. Below, more explanations, along with some graphs, are presented for this context. Datasets, chosen arbitrarily from Cluster 1, generated by DTW k-means for each building type, have been selected. In Figure 14, the ANN model shows a high prediction capability for schools in the winter compared to the summer scenario. The reason is that in winter, heat load peaks appear clearly during working days daily, while in summer, this load pattern is negligible. The heat demand is substantially lower during the weekend in winter, while in summer this difference is not significant. This matter is highlighted by the black dashed rectangles in Figure 14 (a) and Figure 14 (b). Peak energy demands in winter are linked mainly with the need for SH and the noticeable occupancy rate. In the summer, SH demands are insignificant, and occupancy is almost null due to the annual vacation. Concerning nursing homes, the ANN model shows clearly low prediction capabilities, especially in summer (see Figure 15 (b)), as in this case, the outdoor temperature factor is less dominant for DH load (SH demand is significantly low). In this particular case, the DHW load represents the main energy demand portion, which is related to factors other than weather data. This effect is less dominant in the case of schools due to the impact of other important aspects, mainly related to the occupant's activities. D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 35 Figure 14 Prediction performance for a school dataset (a) in winter (b) in summer. Figure 15 Prediction performance for a nursing home dataset (a) in winter (b) in summer. For load prediction assessment in mild seasons, the fact that outdoor temperature is no longer a dominant parameter for determining the heat load leads to inaccurate predictions for nursing homes (see Figure 16 (a)), however, ANN is still showing relatively better results when it comes to school heat load prediction. D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 36 Figure 16 Prediction performance in mild-season (a) for a nursing home (b) for a school Prediction performance evaluation Table 3 lists mean squared error (MSE) and mean absolute error (MAE) values for the prediction performance of the proposed ANN regression method for all clusters, compared to a reference method developed by (T. O Timoudas et al., 2022). The authors here developed an ANN approach that includes, along with the historical outdoor temperatures, and historical heat loads with different time lengths. It has been found that better results were reached using historical values of outdoor temperature and load up to 24 hours. This approach has been validated on the same datasets used in this article, either for training or testing. Better results have generally been obtained using Cluster 2 datasets using both clustering techniques (Euclidean and DTW k-means). One justification for that is that Cluster 2 generally gathers less fluctuating load profiles, as shown in Figure 13. This makes it easier for MLP to track smoother load curves appearing in the winter as they reflect SH demands. Such load curves are highly dependent on weather data, mainly, outdoor temperatures, which explains the high prediction performance in this case. Figure 12 shows more fluctuating curves, which can be affected by other additional factors, such as occupancy change rates. For all clusters, the table shows clearly that the comparison is in favour of the MLP model, presented in this study, compared to the ANN baseline method. D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 37 Table 3 MSE and MAE evaluation using nursing homes normalized data for the year 2019. Evaluation criteria Proposed MLP ANN proposed in [10] MSE Cluster 1 (Euc k-means) 0.0153 0.0219 Cluster 1 (DTW k-means) 0.0154 Cluster 2 (Euc k-means) 0.0073 Cluster 2 (DTW k-means) 0.0072 MAE Cluster 1 (Euc k-means) 0.0975 0.1106 Cluster 1 (DTW k-means) 0.0977 Cluster 2 (Euc k-means) 0.0639 Cluster 2 (DTW k-means) 0.0638 4.3. Continental/Alpine Climate – optimal demand response management in DHN at a community level 4.3.1. Use case description. The optimisation-based demand side management (DSM) policy developed in this study is validated on data generated by smart meters which are installed in the Sonnenplatz live-in lab in Großschönau, Austria. It consists of a local energy community powered by LV and local DH networks (see Figure 17). The users connected to the network consist of three public buildings, with an overall heated surface of 2000 m2 and an estimated average power consumption of 150 kW: two commercial buildings with an estimated power consumption of 170 kW, and four residential buildings with an estimated power consumption of 24 kW. The generation sources include one of the existing biomass boilers with a capacity of 130 kW, connected to a 10 m3 water tank, and the proposed solar system, with an overall capacity of 250 kW. This study assumes that the DH network is connected to each building through the substation topology shown in Figure 4. HP plays the role of a temperature booster that upgrades the DH supply temperature from ~55°C to ~70°C. It is also considered a connection point of the electric grid (via the compressor operation) with the thermal networks/assets. In this study, the optimisation is carried out based on the flowchart presented in Figure 18. Because computing time increases exponentially with the number of involved buildings, in this study, the simulation of the DSM strategy is limited to four buildings, labelled conventionally as follows: Building 4.5, Building 6.0, Building 7.0 with 10.5 kWp PV, and Building 8.0 with 11 kWp PV. D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 38 Figure 17 Satellite view of Großschönau community with buildings considered for optimal energy coordination. Figure 18 Energy optimisation flowchart. D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 39 Table 4 Simulation parameters Parameter Value Electricity buying price 0.45 €/kWh Electricity selling price 0.15 €/kWh TES volume 500 liters TES inlet water flow 0.018 kg/s TES outlet water flow 0.013 kg/s TES max SOC 90 % TES min SOC 10 % DH supply temp 55 °C HP condenser min temp 50 °C HP condenser max temp 70 °C PV power cabling loss 3 % Average CO2 emissions 127 g/kWh Table 5 PSO parameters Parameter Value Number of particles 80 Decision variables 4 Maximum iterations 50 Inertia weight w 0.5 Cognitive coefficient 1 c 0.6 Social coefficient 2 c 1.6 4.3.2. Simulation results In this section, a simulation study is carried out for the assessment of the developed optimisationbased DSM. The simulation spans a two-day scenario using IoT data generated on May 30-31, 2023. Data for the heat load of the four buildings, concerned for simulation, with the PV power production of two of them, are displayed in Figure 19. The main parameters used in the simulation are summarized in Table 4 For better control performance assessment, the study assumes the presence of under-sized TES units at each relevant building's substation, specifically utilizing 200-liter TES. The optimisationbased DSM changes, at each time step (one hour), all HP condenser temperature setpoints. The main PSO optimisation parameters are listed in Table 5. The simulation asses the optimisation outcome for two cases with two objective functions weighing modes: D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 40 Figure 19 Heat demand with PV power production of the concerned buildings. Case 1 (user-oriented optimisation): in this case, only the first term of the objective function is considered. Therefore, energy optimisation is in favor of peer-to-peer or peer-to-utility trading. The goal is to reduce the energy cost for each building individually by driving different HP units optimally. The HP temperature setpoint, as well as the SOC evolution of each TES, are displayed in Figure 20. The simulation showed an energy cost of 0.00 €, 0.00 €, -3.95 €, and -3.80 € for Building 4.5, Building 6.0, Building 7.0 and Building 8.0 respectively. Positive values refer to energy buying prices while negative ones refer to selling prices. The total exchanged energy with the grid utility was –51.76 kWh, similarly, the negative sign refers to an energy exported to the grid utility. Figure 20 HP optimal condenser temperature setpoint with associated TES SOC variation in a 2-day simulation (user-centric scenario) D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 41 Case 2 (community-oriented optimisation): in this case, only the second term of the objective function is considered. The goal is to maximize the energetic independence of the concerned buildings. The method is based on sharing optimally the PV energy between buildings to give higher flexibility to the HP units to work in optimal coordination and independently from the grid utility. The HP temperature setpoint, as well as the SOC evolution of each TES, are displayed in Figure 21. The simulation showed an energy cost of 6.24 €, 3.57 €, -2.89 € and –3.22 € for Building 4.5, Building 6.0, Building 7.0 and Building 8.0 respectively. The total exchanged energy between the community and the grid utility was –19.63 kWh. Figure 21 HP optimal condenser temperature setpoint with associated TES SOC variation in a 2-day simulation (community-centric scenario) The simulation results reveal that in Case 1, the cost-effective HP operation is evident for each end-user. In fact, in Building 4.5 and Building 6.0, HP units did not experience any significant work. Though, a large unused surplus of PV energy was exported to the grid utility. In contrast, Case 2 addresses this surplus through the optimal coordination of all involved HP units. However, Carbon emission evaluation shows that, in Case 1, a total of 0.39 kg CO2 emissions related to the HPs operation is calculated, while in Case 2, 3.58 kg was calculated. Striking a balance between user-centric and community-oriented optimisations is essential and tailored to each use-case specification, which is driven by economic and ecological targets. 4.4. Mediterranean Climate – optimal control of heat pump with a photovoltaic/storage system in centralized DHN substation 4.4.1. Use-case description The Live-in Lab ENVIPARK consists of ten buildings, four of which are mainly used as laboratories, five as offices, and one as a canteen and restaurant. The DHN is supplied by two heat exchangers with 1.2 MW each with a supply temperature around 75°C during the winter season, and for domestic heat D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 48 in Figure 24 (c) and the electric power imported from the grid for each scenario. The table shows how significant is the impact of extending the operation flexibility boundaries given to each ESS technology (TES and BESS) in the performance of the heating cost optimisation. Adjusting the variation interval of the TES maximum water temperature from 10°C to 15°C and BESS SOC variation interval from 40% to 80% has led to a cost saving of up to 12.4% using SQP chosen in this article, and 13.9% using pattern search algorithm. However, extended flexibility limits have led to more computing time, which is explained by the prolonged search space for each algorithm. If the energy cost optimisation process is a part of real-time control, this issue is crucial and should be considered. Table 7 Flexibility impact assessment of different optimisation methods. Optimisation algorithm Heating cost Execution time Cost saving (Baseline Scenario 1) Scenario 1 Scenario 2 Scenario 1 Scenario 2 SQP 1466 € 1284 € 8 min 12 min 12.4% Interior-point (F. Capitanescu et al., 2007) 1456 € 1262 € 22 min 28 min 13.3% Active-set (Zhong et al., 2023) 1468 € 1290 € 6 min 7 min 12.1% Pattern search (Tao et al., 2021) 1434 € 1234 € 1.7 hour 1.6 hour 13.9% Conclusions Due to the wide availability of data from smart meters, sensors, and building management systems in DHN-powered buildings, data-driven approaches, for optimizing energy usage at building and community levels, represent powerful tools. Such data can help in uncovering end-user behaviour enabling by that the development of advanced control and management policies. In this context, this deliverable shows the development and validation steps of several data-driven techniques offering clustering and prediction functions. Combining prediction models with metaheuristic and deterministic optimisation algorithms, this deliverable shows also optimal control and demand management that take advantage of available energy flexibility in DHN energy communities, either as a distributed control solution thanks to advanced IoT communication, or as a centralized one that relies on aggregated data in the main DHN substation. Summary of achievements Based on the conclusions described above, the main achievements of the works presented in this deliverable are listed below: • According to each use case specifications, different data collection solutions have been adopted for better data analytics. This includes working with classical files-based data (.xls or D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 49 .csv), relational and time series databases (through SQL and API querying), or through IoT communication. • Advanced ML data mining techniques may represent powerful tools to deal with low-quality and noisy data, obtained usually from temporal malfunctioning of sensors or internet disconnection. This can also identify uncommon DH consumption patterns and abnormal user behaviour, which improves significantly the performance prediction models that are built in cluster-wise approach. • Leveraging DL for long-term prediction of DH load showed adequate to good performances in different scenarios. • By taking advantage of prediction models with metaheuristic and deterministic optimisations, it is possible to build advanced model-based control and management policies that optimize energy usage in DHN communities. • When disaggregated data in DHN communities are not available, aggregated one can be utilized to enable centralized management based on the optimal exploitation of the energy flexibility in the main substation. Relation to continued developments. Following the last point listed above, there is still an ongoing activity that consists of disaggregating power consumption data at a building level using ML-based non-instructive load monitoring (MLNILM). This technique can discover the operation state of different HVAC installations by analyzing their electric power consumption indicators (active and reactive energies, power factor, phase current). This technique has the potential to assess power and heat flows in the building and up to the community level, which enables the implementation of energy management strategies applicable for sector coupling. D3.3 Methodologies and frameworks to optimize demand response and peak load shaving for 4th5th DHC in three typical European climates. 50 References A. Javed, B.S. Lee and D.M. Rizzo, “A benchmark study on time series clustering”, Machine Learning with Applications, 1, 2020, 100001. Boggs, P. and Tolle, J, 1995, Sequential Quadratic Programming. Acta Numerica, 4, 1-51. J. Kennedy and R. Eberhart (1995). Particle Swarm Optimisation. Proceedings of the 1995 IEEE International Conference on Neural Networks (Perth, Australia): IEEE Service Centre, Piscataway, NJ, IV: pp 1942-1948. M. Jesper, F. Pag, K. Vajen and U. Jordan, “Annual Industrial and Commercial Heat Load Profiles: Modelling Based on k-Means Clustering and Regression Analysis”, Energy Conversion and Management: X, Volume 10, 2021, 100085. M.M. Alam, H. Bao, P.X.W. Zou and J. Sanjaya, “Behaviour Change of Building Users and Energy Consumption”, Encyclopedia of Sustainable Technologies 2017, Pages 189-196. O.I. Abiodun, A. Jantan, A.E. Omolara, K.V. Dada, N.A.E. Mohamed and H. Arshad. State-of-the-art in artificial neural network applications: A survey. Heliyon, Volume 4, 2018, 00938. S. Aghabozorgi, AS Shirkhorshidi and TY Wah, Time-series clustering–a decade review. Inf Syst 2015; 53:16–38. Thomas OhlsonTimoudas, Yiyu Ding, Qian Wang. A novel machine learning approach to predict shortterm energy load for future low-temperature district heating. REHVA 14th HVAC World Congress 22 – 25 May, Rotterdam, The Netherlands Zheng Zhang, Ping Tang, Lianzhi Huo, and Zengguang Zhou. MODIS NDVI time series clustering under dynamic time warping. International Journal of Wavelets, Multiresolution and Information Processing HYPERLINK "https://www.worldscientific.com/toc/ijwmip/12/05". (2014) . Florin Capitanescu, Mevludin Glavic, Damien Ernst, Louis Wehenkel, 2007, Interior-point based algorithms for the solution of optimal power flow problems. Electric Power Systems Research 77, 508–517 H. Zhong, S. Meng, X. Zhang, Z. Wei, C. Zhang and L. Du, 2023, "Multiphysics-Constrained Fast Charging of Lithium-Ion Battery With Active Set Predictive Control," in IEEE Transactions on Intelligent Transportation Systems Hai Tao, Faraedoon Waly Ahmed, Halkawt Abdalqadir kh ahmed, Mohsen Latifi, Hiroki Nakamura, Yafeng Li, 2021, Hybrid whale optimization and pattern search algorithm for day-ahead operation of a microgrid in the presence of electric vehicles and renewable energies, Journal of Cleaner Production 308, 127215