Double deck elevator group control systems using evolutionary algorithms: interfloor and lunchpeak traffic analysis

Depósito de Investigación de la Universidad de Sevilla https://idus.us.es/ This is an Accepted Manuscript of an article published by Elsevier in Computers & Industrial Engineering, Vol. 155, on May 2021, available at: https://doi.org/10.1016/j.cie.2021.107190 © 2021 Elsevier. En idUS Licencia Creative Commons CC BY-NC-ND Double deck elevator group control systems using evolutionary algorithms: interfloor and lunchpeak traffic analysis Pablo Cortés 1 , Jesús Muñuzuri, Alejandro Vázquez-Ledesma, Luis Onieva Ingeniería de Organización, Escuela Técnica Superior de Ingeniería, Universidad de Sevilla, Camino de los Descubrimientos s/n, 41092 Sevilla, Spain Abstract.- The continuous development of high-rise buildings around the world requires the installation of efficient elevator systems able to vertically transport the different passengers along the buildings in their daily journeys. Double deck elevators can increase the efficiency of these vertical transportation systems. Double deck elevators consist of two adjacent cabins that are joined and travel together along the same shaft, so the handling capacity of the system can be improved by allowing the dispatch of passengers with destination to two consecutive floors at the same instant. This type of architecture emerges as especially appropriate for uppeak traffic conditions. However, its suitability has not been sufficiently analysed for non-dominant (up or down) traffic patterns, such as interfloor and lunchpeak traffic. Our paper deals with conventionally controlled double deck elevators, where the Elevator Group Control System (EGCS) requires specific car-landing call allocation algorithms able to manage such special car architectures. Along this line, we propose a genetic algorithm that demonstrated a good performance when compared to a tabu search algorithm, taking into account different fitness evaluation functions (overall dispatching time and nearest call). The analysis was undertaken for interfloor and lunchpeak traffics and the average waiting, transit and journey times, and the energy consumption are reported as performance indexes of the vertical transportation system. The algorithms produced efficient results outperforming the considered benchmark and emerged as very competitive algorithms considering all the performance indexes as a whole. Results were tested using ELEVATE, the standard simulation software for vertical transportation. Keywords: double deck; elevator group control system; vertical transportation; traffic pattern; genetic algorithm. 1 Introduction Vertical transportation problems arise as a challenging real-world field of research that includes many complex industrial engineering problems associated with the efficient and optimal use of resources in a dynamic way. Some of these problems relate to the scheduling and allocation of elevators to the calls that are issued in the different floors of the buildings. Tackling such problems requires knowledge engineering, intelligent optimization or artificial intelligence approaches that benefit from the use of information learned from data, such as the type of traffic associated to each optimization period dynamically. In fact, the continuous development of high-rise buildings around the world requires the installation of efficient elevator systems able to vertically transport the different passengers along the buildings in their daily journeys. The scientific literature is paying attention to the subsequent problems arising, resulting in a wide research field over the last years (Fernández and Cortés, 2015; Cortés et al. 2018). The Elevator Group Control System (EGCS) constitutes the core of such vertical transportation systems. An EGCS consists of hall call buttons located at every floor, car call buttons inside the cars of the elevator group, and a group controller that must manage the multiple cars to efficiently transport passengers. This group controller includes the dispatching algorithms that are responsible for the efficient allocation of landing calls (issued by pressing the hall call buttons) to the cars of the EGCS. Currently, EGCSs can be managed by 1 Corresponding author pc[email protected] conventional control or destination control technologies. In an EGCS with destination control the passengers have to select their destination floor before entering the car by pressing the corresponding button in the lobby, since there are no destination buttons inside the car. By contrast, in our paper we only consider the control of double-deck elevators with conventional up and down call buttons in the lobbies, and with the passenger issuing his/her destination when inside the car (see Figure 1 to appreciate alternative lobby button sets). Along this line, the dispatching algorithm associated to the EGCS has to solve the following problem: every time a passenger makes a landing call by pressing the corresponding hall call button, one car must be assigned to such call. This assignment is made according to a specified criterion that must be optimized. Different criteria can be considered (Cortés et al. 2006), with the most usual ones based on the minimization of service times (Fernández et al. 2014). These criteria are the Average Waiting Time (AWT), the Average Transit Time (ATT), or the passenger time to destination, also called Average Journey Time (AJT), which is calculated as AWT+ATT (Barney, 2005). However, criteria based on other aspects have started to gain attention, like energy consumption (Fernández et al., 2013) or the multi-criteria combination of several indicators (Tyni and Ylinen, 2006). It is important to note that the energy consumption is a relevant performance index from the building managers’ and owners’ perspective. Figure 1. Different existing button sets. Intelligent optimization algorithms (IOAs) have proved to be efficient approaches to deal with the elevator dispatching and scheduling problem. Today, there is a large set of relevant alternatives available (see BoussaïD et al., 2013 for a wide scope on optimization metaheuristics). Among them, genetic algorithms have been one of the most popular approaches. One of the first real-world application of a genetic algorithms to elevator dispatching was proposed by Tyni and Ylinen (2001). Other genetic approaches can be found in Cortés et al., (2003), Cortés et al., (2004), or Bolat et al., (2010). Other metaheuristics have also been successfully implemented, such as Particle Swarm Optimization (Li et al., 2007a; Bolat et al., 2013), immune systems algorithms (Li et al. 2007b), tabu search (Bolat and Cortés, 2011), Viral Systems (Cortés et al. 2013), or reinforcement learning algorithms (Wei et al., 2020) among Example of a lobby button set with destination selection (a) Example of a totem-type lobby button set with destination selection(c) Example of a lobby button set with destination selection (b) Example of a button set located inside the elevator cabin (f) Example of a lobby button set without destination selection, but with movement direction sellection(d) Example of a common lobby button set, without movememt direction (e) others. Future developments are conducting the research to the integration of the Internet of Things as a key enabling technology able to provide real-time data to the EGCS (Van et al., 2019; Van et al., 2020; and Zhou et al. 2018). The daily activity in buildings is ruled by traffic patterns associated to the passengers’ movements and determining the appropriateness of the different dispatching algorithms and even the criteria to be used in their objective function. Usually, four traffic patterns are considered depending on the dominant journey flow: this traffic can be from the main floor to the rest of floors in the building (uppeak traffic), from the different floors toward the main floor (downpeak traffic), a combination of uppeak and downpeak that takes place at the lunch hour (lunchpeak or midday traffic), or an assorted traffic without dominant directions (interfloor), (see Cortés et al., 2012 for a further description). The knowledge associated to the type of traffic constitutes a relevant source of data with dynamic variations. This characteristic associated to variations in the traffic provides an excellent framework for using evolutionary approaches to deal with vertical transportation problems since they have the ability to adapt dynamically their evolution-to-the-optimum by taking advantage of the information learned from data in realtime. Figure 2 provides a representation of the typical movements in office buildings (the vertical axis represents the population movement and the horizontal axis corresponds to the hourly period). On the one hand, lunchpeak, interfloor and downpeak traffics allow diverse strategies for dispatching and allocating cars to the landing calls. On the other hand, uppeak traffic is mainly managed by sending back the cars to the ground and basements to pick up travellers in order to be distributed in their destination floors. For this type of traffic, the Round-Trip Time (RTT) or the handling capacity are the most relevant vertical transportation system performance indexes. Figure 2. Traffic patterns in an office building. Despite the already discussed significant amount of research in the field of vertical transportation, there are not many contributions dealing specifically with double deck elevators, where the number of scientific contributions is much more limited. This paper is focused on such elevators and proposes different approaches to manage the dispatching problem, given that the singular architecture of these elevators requires specifically oriented control algorithms. Double decks are especially appropriate for heavy uppeak traffic conditions because they represent a significant increase in the handling capacity of the vertical transportation system (Hirasawa et Time Downward traffic Upward traffic Uppeak Lunchpeak Lunchpeak Downpeak Interfloor Interfloor al. 2008; Yu et al. 2007; Zhou et al. 2008 and Zhou et al. 2007). In addition, there are several studies considering downpeak traffic for double deck elevators (Hirasawa et al. 2008; Zhou et al. 2008; Sorsa et al., 2003). However, their performance for other types of traffic has not been generally studied in the scientific literature and must be carefully analysed and optimized too. This is especially true for the case of interfloor traffic, which has received very little attention so far from the scientific community. In this paper we conduct a detailed analysis of interfloor and lunchpeak scenarios and analyse different car-landing call allocation algorithms. In the remainder of the paper, section 2 follows with the description of double deck elevator architectures, including a discussion of the relevant literature on their specific control systems and optimization algorithms. The main contributions of our work, including the consideration of this particular type of elevators, of alternative efficiency criteria, and of interfloor and lunchpeak traffic patterns, are also discussed in this section. Then, section 3 presents the double deck elevator group control system model. Section 4 is dedicated to the car-landing call allocation genetic algorithm, as well as to the fitness function used to evaluate the appropriateness of the candidate solutions. Section 5 is devoted to the experimentation comparing the genetic approaches and the CC-Elevate algorithm that is used as a benchmark for the proposed traffic patterns (interfloor and lunchpeak). Finally, section 6 remarks the main conclusions of our research. 2 Double deck elevators Double deck elevators consist of two adjacent cabins (also called decks) that are joined and travel together along the same shaft. In fact, a double deck is an elevator with two cabins attached together, one on top of the other (see Figure 3 for an illustration of double deck elevators with two adjacent cabins travelling together). They constitute a sound technical solution for buildings higher than 60 floors, and for shuttle applications in very tall buildings (AlKodmany, 2015). If the number of elevators is more than 8 for conventional group control and the required car capacity is more than 26 persons or 2000 kg, then the application of doubledeck elevators is recommended (Al-Sharif et al, 2017b). Taller and larger buildings call in turn for more complex elevator architectures, like the one described recently by Liew et al (2018), who have integrated a double-deck elevator system into a multi-car elevator system (Tanaka et al, 2016) to develop a new hybridized elevator system, or the proposal by Kim et al. (2018) that suggests the use of flexible double-cage hoist for high operational efficiency in double deck architectures. Figure 3. Two double-deck elevators at Midland Square, Nagoya, Japan (Source: Wikimedia Commons). Given the special characteristics of double deck systems and to allow the access to all the floors in the building, the access to destinations in even and odd floors are differentiated at the main entrance access hall of the building (see Figure 4). Therefore, even floors are accessed only from the upper deck and odd floors from the lower deck (note that the highest floor of the building can only be accessed through the top deck of the elevator). See section 3 for a more detailed discussion on how the double deck elevator group control system manages this issue. Figure 4. Main entrance hall access to double deck elevators. This system increases the handling capacity of the system by allowing the dispatch of passengers with destination to two consecutive floors at the same time. It has been said that due to this fact, double deck systems can reduce the elevator footprint between 25% and 40% (CIBSE Guide, 2005). Since their apparition in the sixties of the past Century, the controllers of double deck elevators have progressed very much. The first control systems for double deck elevators consisted of a trailing-deck control system that divided the elevator into the leading deck (first deck in the direction of movement) and the trailing deck. The dispatching method was a simplistic one based on the principle of collective dispatching to the trailing deck. Therefore, the leading deck only attended to the landing-calls adjacent to those landing-calls already assigned to the trailing deck. To odd floors To even floors In recent years, more advanced control systems of double deck elevator groups have arisen in the industry. However, there are not many references to double deck systems in the scientific literature yet. Most of the contributions have come from KONE Corporation researchers and are associated to industrial patents. The first contributions were due to Kavounas (1989), Fortune (1996), and later, Siikonen (1998), who published scientific-technical papers in Elevator World showing the advantages of such double deck architectures. Since then, the research has been mainly undertaken by two different research groups that have followed different approaches to deal with the problem. The first research group has focused mostly on the elevator dispatching problem, using operations research methodologies. Most of the group members come from the KONE Corporation. This is the case of Sorsa et al. (2003), who presented one of the first approaches developing a graph-based formulation that is solved using a genetic algorithm empowered with the GeneBank technology (Tyni and Ylinen, 1999) to allow real-time operation. This GeneBank approach pre-computes the possible chromosomes and has been used in most of the papers of this research group thereafter. Sorsa et al. (2003) proposed a routing formulation based on ideas taken from the multiple travelling salesman problem and the pick-up and delivery problem. Simulation results, including waiting and journey times, were provided for lunchpeak and downpeak traffic in a building with 29 populated floors. The same authors proposed a generic procedure to control multi-deck systems (Tyni and Ylinen, 2001). However, the conclusions in Sorsa et al. (2003) remark that multi-deck elevators with more than two adjacent cars will not be suitable in the near future. Moreover, Sorsa and Ruokokoski (2013) argued that although double deck elevators produce better performance under certain specific traffic patterns (typically under uppeak traffic, improving the handling capacity and the round trip time), they can produce worse performance for other traffic patterns or even some conflicts between different criteria, such as the waiting and journey times. Recently, Sorsa (2019) has presented a genetic algorithm that is compared against branch and bound solutions of the model providing satisfactory solutions. The algorithms are simulated in a KONE proprietary tool, and results are provided for lunchpeak traffic. There is a second team of researchers working with this type of architectures. This research team comes from the Waseda University and the Fujitec Corporation, and follows an approach focused on the elevator group control system, which is more similar to ours. In fact, they have also proposed different genetic approaches to solve the problem. The main references in this case are: (i) Zhou et al. (2008) proposed a genetic programming network to be used as a trafficflow-adaptive controller able to manage other traffic patterns different from uppeak; (ii) Hirasawa et al. (2008) suggested an approach based on the same type of genetic network programming approach. First, their algorithm optimizes the genetic network programming for solving the double-deck system problem. Next, its performance is compared with the performance obtained by using the conventional methods for validation. They undertake an extensive experimentation for uppeak and downpeak periods, and some additional experiments for other traffic patterns; (iii) Yu et al. (2007a) described the main rules for the same genetic network programming; (iv) Yu et al. (2007b) proposed an allocation algorithm based on genetic network programming and optimized by an ant-colony system that provided fast convergence, tested with a real case study; and (v) Zhou et al. (2007) followed a very similar approach to Yu et al. (2007b) but replacing the ant colony algorithm by reinforcement learning techniques. These five papers corresponding to the group of authors Zhou, Yu, Mabu, Hirasawa and Markon that were published between 2007 and 2008 follow a similar line and apply the same genetic network programming procedure and the same structure of experiments, paying major attention to uppeak and downpeak traffics. The authors acknowledge that the two conference papers (Yu et al. 2007b and Zhou et al. 2007) are based on their past studies and make a limited contribution focused on improving the genetic programming network by using reinforcement learning and ant colony optimization. Table 1 summarizes the main contributions of the related publications to the state of the art of double deck elevator group control systems and the methodological approach that was followed in each case, as well as the performance criteria considered in the dispatching process. Table 1. Double deck elevator group control systems literature review. Contribution Methodology Performance criteria* Hirasawa et al. (2008) Genetic network programming AWT/Max WT Sorsa (2019); Sorsa and Rukokoski (2013) Genetic algorithm empowered with GeneBank AWT/ATT Sorsa et al. (2003) Genetic algorithm empowered with GeneBank AWT/AJT Yu et al. (2007a) Genetic network programming AWT/MaxWT Yu et al. (2007b) Genetic network programming with Ant Colony Optimization AWT/MaxWT Zhou et al. (2007) Genetic network programming with reinforcement learning AWT/MaxWT Zhou et al. (2008) Genetic network programming AWT/MaxWT Al-Sharif et al. (2017a) Monte Carlo simulation RTT/No. stops *AWT: Average Waiting Time ATT: Average Transit Time AJT: Average Journey Time MaxWT: Maximum Waiting Time RTT: Round Trip Time No. stops: number of stops of the cars In view of the existing scientific literature on the subject, we can state that the main difference between our paper and the KONE research team approach lies in the methodology. The KONE research team follows the traditional mathematical programming approach formulating the mathematical model and providing optimal solutions for small instances in order to validate their optimization approach (typically genetic algorithms). Then, they provide results for larger problems that are tested in a proprietary tool. On the other hand, our contribution is focused on the double deck elevator group control system (see next sections 3 and 4), which is tested in a publicly available commercial simulation tool instead of a proprietary tool. In any case, we have not been able to find references that manage conventional double deck elevators (the case we are dealing with in our paper) and capacity constraints using a mathematical model. Instead, we propose a control system based on a genetic algorithm more in line with the second group of researchers mentioned above. In fact, the scientific literature of elevators could be divided between those approaches based on the mathematical models and those others that tackle the problem directly, considering the innate constraints of the vertical transportation that are managed by the optimization approach and controlled by the simulation software. Our proposal follows this last approach. Moreover, our paper provides contributions representing progress in some other issues beyond the state of the art: – First, most of the double-deck scientific literature has been based on uppeak and downpeak traffic patterns (that is, traffic with dominant up or down traveling directions). Hence, the major future contributions should be based on testing and improving the system’s performance for interfloor and lunchpeak traffic patterns. And given the lack of attention to the interfloor traffic pattern (there has not been any paper dealing with such traffic until now), we have focused our research on this type of traffic. Indeed, interfloor traffic is the pattern that covers most of the day. An appropriate use of the energy during the interfloor traffic pattern can therefore represent significant energy savings (Fernández et al., 2013). – Second, we consider alternative and different efficiency criteria to analyse the problem and its real implementation viability: we provide results for waiting and journey times, as well as energy consumption. This last criterion has emerged as a relevant performance index in high-rise office buildings, particularly important from the managers’ and owners’ point of view. It is important to note that the energy criterion has not been considered before in double deck elevators. – Third, we compare the performance of our algorithm for several buildings of different heights, considering 12, 16, 20 and 24 populated floors, thus creating a realistic basis for comparison. 3 The double deck group control system The EGCS consists of a set of hall call buttons that are located on every floor, together with car call buttons inside the cabins and a group controller. The EGCS must manage the different allocations of landing calls from halls to cars efficiently. The allocation criterion is associated to different quality of service metrics. Elevator control follows a set of basic rules largely used and implemented in elevator control systems. Such rules are as follows: – An elevator may not stop at a floor where no passenger enters or leaves it. – An elevator may not pass a floor at which a passenger wishes to leave it. – A passenger may not enter an elevator carrying passengers and travelling in the reverse direction to the required direction of travel. – An elevator may not reverse its direction of travel while carrying passengers. In addition, the elevator control system follows the well-known collective control principle. Such principle, as Barney (2003) states, determines that the system serves floors with transportation requests in an ascending or descending order depending on its current travelling direction. Following these rules, we implemented the double-deck control system showed in Figure 5. For our problem, the candidate solutions in the tabu search algorithm are based on the description in Figure 6. The candidate solution identifies those cars being active and serving calls in a binary array stating a value equal to ‘1’ in case of using the car and a ‘0’ if the car is not being used. Once the candidate solution is configured, the car-landing call allocation is done following a random procedure to generate the solution encoding as described in Figure 6. The neighbourhood is constructed by considering possible permutations in the active cars. This reduces the neighbourhood size allowing a fastest computation of the tabu search. The neighbourhood of a candidate solution is stated by changing a ‘1’ to a ‘0’ with a 50% probability and vice versa. Consequently, the neighbourhood size is 2N-1. Every visited solution is added to the tabu list and solutions in the tabu list cannot visited again unless an aspiration criterion is satisfied. The aspiration criterion is satisfied when the new candidate solution has a quality of service value lower than 1.2 times the value of the best found solution. When the tabu list is full, it is emptied following a FIFO criterion. The tabu search algorithm is iteratively run until a maximum number of iterations (Max-Iter) is reached. The algorithm pseudocode is described below (Algorithm 2). Algorithm 2. Tabu search algorithm pseudocode. 1: Begin 2: Generate initial solution 3: Initiate Tabu List 4: Save best solution 5: for i =1 to Max_Iter do 6: Select new solution in neighbourhood 7: Evaluate new solution 8: if new solution is in Tabu List then 9: if new solution verifies Aspiration Criterion then 10: Move to new solution 11: Add new solution to Tabu List 12: end if 13: else 14: Move to new solution 15: Add new solution to Tabu List 16: end if 17: if new solution is better than best solution then 18: Update best solution 19: end if 20: end for 21: Landing call allocation according to best solution 22: End 4.5 Lab design analysis This subsection collects the results obtained for a short lab design library. The objective was to identify the most appropriate parameter values for the algorithms, as well as to compare both of them to select the most promising approach to undertake a large-scale simulation. The tests were carried out using ELEVATE, the elevator traffic analysis and simulation software by Peters Research Ltd. (Elevate, 2007), which has become the worldwide elevator industry standard (Caporale, 2000). We used a building case study with 12 floors and 5 double deck elevators. Lunchpeak was selected as traffic pattern to carry out the tests, since it depicts heavy and complex traffic conditions. The Average Waiting Time (AWT) was chosen as the performance index, since it is the most widespread performance index in vertical transportation (Barney, 2003; CIBSE Guide, 2005). In order to define the set of parameters of the algorithms, we implemented the NC fitness function. This was due to the fact that the fitness function should not condition the algorithm parameters’ design (that is, crossover/mutation probability and population size in case of the genetic algorithm, and tabu list size for the tabu search, as well as the maximum number of iterations for both cases). We compared values between 0.85 and 0.95 for crossover probability, which are commonly accepted as appropriate values for genetic algorithms, and tested populations between 50 to 70 individuals. The results obtained allowed us to identify a population of 64 individuals, a crossover probability equal to 0.9 (remaining 0.1 for mutation). Regarding the tabu search, we tested values of the tabu list size between 10 to 40, identifying a size equal to 30 as the best alternative (see Table 2). In both cases 115 was selected as the maximum number of iterations since a greater number did not produce better results, and in any case it consumed too much computation time over such number of iterations. Table 2. Parameter selection. GA TS Crossover Pop.size AWT Tabu List AWT 0.95 70 19.2 40 19.1 0.95 64 19.2 30 18.5 0.95 58 19.7 20 21.8 0.95 50 22.1 10 22.1 0.90 70 19.0 0.90 64 19.0 0.90 58 19.8 0.90 50 21.0 0.85 70 21.1 0.85 64 21.2 0.85 58 21.4 0.85 50 23.1 Once the parameters of the algorithms were selected, a short lab design library was defined to analyse the behaviour of both algorithms in order to identify the best-performing algorithm for the problem. Tests were carried out for 12 and 20 floor buildings with double deck elevator groups of 5 and 7, and 4 and 8 respectively. The analysis was undertaken using the NC and ODT fitness functions and lunchpeak and interfloor traffic. Table 3 shows the results and demonstrates the better performance of the genetic algorithm that outperformed the tabu search for all the analysed cases in terms of average waiting and transit times (AWT and ATT). Table 3. Lab design library analysis. 5 Experimentation In this section, we define the comprehensive experimentation carried out to analyse the behaviour of the selected allocation genetic algorithm according a large variety of performance indexes. The experimentation was undertaken with the ELEVATE software, and we carried out simulations for lunchpeak and interfloor traffic conditions. Note that uppeak traffic algorithms are focused on collecting passengers at the entrance floors to be delivered through the different floors up to the highest destination floor and then the elevator is sent back to the entrance floors again until the end of the uppeak period maximizing the round trip time. Downpeak traffic follows an equivalent but inverse strategy. By contrast, lunchpeak and interfloor are more variable traffic patterns and they appear as more relevant to analyse the potential improvements offered by alternative algorithms. We used the traditional AWT, ATT and AJT performance indexes, and added the energy consumption to complete the experimentation target of our paper. This analysis provides a detailed perspective of the most appropriate approaches depending on the type of traffic and intended objective for the vertical transportation system. We used the double deck conventional control algorithm included in the algorithm suite of ELEVATE to compare the proposed algorithms (Elevate, 2007). The algorithm is based on the collective principle and intends to minimise the AWT, by using the lowest Estimated Time of Arrival (ETA) as allocation criterion. This algorithm is used as an appropriate benchmark and is referred to as CC-Elevate in our tests. 5.1. Experimentation configuration Experiments were simulated with passenger demands generated by ELEVATE, where passenger arrival times are randomly generated according to a Poisson process. Because of that and to obtain an appropriate statistical validation, we simulated 120 minutes to guarantee statistically meaningful sample sizes. Tests were carried out for tall buildings from 12 to 24 floors at 3.3 m of distance, and EGCS with a number of cars varying from 4 to 8 depending on the height of the building, which corresponds No. Floors No. Cars AWT ATT AWT ATT AWT ATT AWT ATT 519 40.3 20 41.1 19.9 39.8 22.8 40.8 7 10.9 33.7 11.5 33.1 8.3 30.4 10.7 30.5 4 51.1 67.1 57.6 67.9 55.3 71.7 66.9 70.9 8 28.8 59.3 28.9 59.8 27.4 52.4 27.8 53.5 No. Floors No. Cars AWT ATT AWT ATT AWT ATT AWT ATT 5 17.1 42.1 18.5 43.6 18.8 42.4 18.9 42.9 7 7 28.5 8.5 28.7 6.1 28.2 11.7 29.4 4 40.1 70.3 48.1 71.1 44.4 70.8 53 71.2 8 14.8 47.5 17.2 47.7 14.6 46.5 27.9 46.6 GA TS GA TS Interfloor NC ODT 12 20 GA Lunchpeak GA TS 12 20 NC ODT TS to a realistic double deck conventional control system configuration. Under these conditions, the elevator speed also depends on the building height. Table 4 shows parameters for realistic elevator groups that were used as the basis of the experimentation. Note that the parameters have been chosen according to the suggestion for single deck elevators provided by the CIBSE Guide (2005) in its section 3.2.5 to provide an appropriate quality of service. In addition, it must be taken into account that the double‐deck elevator group needs to have two entrance floors below the populated floors, so the real height will vary between 14 and 26 floors. Table 4. Building configurations. Number of populated floors Number of elevators Elevator speed (m/s) Total population 12 5 2.5 900 16 6 3.0 1200 20 7 4.0 1500 24 9 5.0 1800 As previously discussed, we simulated interfloor and lunchpeak traffic to test the algorithms. The specific characteristics of the simulated lunchpeak traffic are 40% incoming, 40% outgoing, and 20% interfloor. To generate the interfloor traffic we used the following percentages: 25% incoming, 25% outgoing, and 50% interfloor. The actual demand was automatically generated by ELEVATE (note that, by default, ELEVATE does not generate interfloor traffic from odd to even numbered floors and even to odd numbered floors). Both traffic patterns were simulated with passenger demands of 10%, 12%, 14%, and 16% of population per five minutes for 120 minutes (%POP parameter in ELEVATE) respectively for the four building configurations to cover a wide range of traffic conditions in a standardized manner. In general, we have followed the recommendations for vertical transportation experimentation configurations that can be found in Hakonen and Siikonen (2009) and Siikonen (1993). 5.2. Experimentation results ELEVATE provides experimentation result reports like the example shown in Figure 10. This example depicts the results (AWT, ATT, AJT and energy consumption) for the GA using NC as the fitness evaluation function for interfloor traffic and considering the first configuration (12 populated floors and 4 elevators). The running times for the two genetic approaches (GA with ODT and NC fitness functions) considering all the possible configurations (combining lunchpeak and interfloor with the configurations of Table 4) were below 1 second. Consequently, these results show that the algorithms can be applied in a real-life EGCS. Figure 10. Example of the results provided by ELEVATE simulation software. Table 5 shows the results obtained after simulating with ELEVATE the genetic algorithm considering the two different fitness functions (NC and ODT), as well as the CC-Elevate that we used as benchmark. The table includes the values for all the performance indexes: AWT, ATT, AJT, and energy consumption. Time values are provided in seconds and energy values are provided in kWh. The table includes the results for interfloor and lunchpeak traffic patterns. The best result for each criterion (AWT, ATT, AJT and energy consumption) and for every combination of building height and number of cars is highlighted in bold and italic. Table 5. Experimentation results. The evaluation of the algorithms incorporates a statistical component since the genetic algorithms include certain random processes. Although the ELEVATE software takes into account such factors to provide average values, the interpretation of the results must consider this aspect carefully. Therefore, in order to do so, we have constructed Table 6 to assess the suitability of the different approaches by showing the degree of performance for each approach compared to the best result for the four criteria and each configuration. The percentage indicates the increase in service time or energy consumption with respect to the best solutions obtained in each case (represented as ‘BEST’). A colour code has been used for clarification: green colour for cases with a percentage of deviation lower than 5% with respect to the best Average of all runs Distribution of Passenger Waiting Times All Floors over complete duration 010 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170 180 time (s) Average Waiting Time (s) 19.0 (+0.0/-0.0) Longest Waiting Time (s) 139.4 (+0.0/-0.0) 0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000 no of passengers waiting less than time 0 10 20 30 40 50 60 70 80 90 100 % passengers waiting less than time Average of all runs Distribution of Passenger Transit Times All Floors over complete duration 010 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170 180 time (s) Average Transit Time (s) 40.3 (+0.0/-0.0) Longest Transit Time (s) 133.8 (+0.0/-0.0) 0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000 no of passengers in transit for less than time 0 10 20 30 40 50 60 70 80 90 100 % passengers in transit for less than time Average of all runs Distribution of Time to Destination All Floors over complete duration 020 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 360 time (s) Average Time To Destination (s) 59.3 (+0.0/-0.0) Longest Time to Destination (s) 240.2 (+0.0/-0.0) 0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000 no of passengers reached destination in less than time 0 10 20 30 40 50 60 70 80 90 100 % passengers who reached destination in less than than time Average of all runs Energy Consumption Cumulative 11:00 11:15 11:30 11:45 12:00 12:15 12:30 12:45 13:00 13:15 time (hrs:min) Total Energy Consumption 78.59711 kWh (+0.00000/-0.00000) Total Cost $4.72 (+$0.00/-$0.00) 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 kWh No. Pop. Floors No. Cars AWT ATT AJT Energy AWT ATT AJT Energy AWT ATT AJT Energy 12 519 40,3 59,3 78,59 19,9 39,8 59,7 78,87 18,3 41,1 59,4 81,09 16 6 25,3 58,2 83,5 79,79 28,6 57,2 85,8 80,62 24,6 58,2 82,8 82,86 20 7 38,7 76,7 115,4 77,97 36 75,9 111,9 81,59 39,4 77,2 116,6 86,22 24 939,8 89,7 129,5 80,52 41,7 90 131,7 83,92 43 95 138 93,73 No. Pop. Floors No. Cars AWT ATT AJT Energy AWT ATT AJT Energy AWT ATT AJT Energy 12 5 17,1 42,1 59,2 82,85 18,8 42,4 61,2 83,25 16,2 41,4 57,6 134,98 16 6 23,6 59,6 83,2 82,78 23,8 59,8 83,6 81,31 21,2 60,8 82 128,42 20 7 32,6 79,3 111,9 85,02 36,2 79,6 115,8 86,10 31,2 80,4 111,6 124,32 24 9 50,7 94,2 144,9 81,71 47,7 93,9 141,6 87,05 46,1 93,7 139,8 123,82 Interfloor GA - NC CC-Elevate GA - ODT Lunchpeak GA - NC CC-Elevate GA - ODT value; yellow colour for cases between 5% and 15%; red colour for cases between 15% and 30%; and black colour for cases with a deviation greater than 30%. Anyhow, note that absolute differences outside of a range interval of 5 seconds rarely took place, which demonstrated the adequacy of all the approaches regarding waiting, transit and journey times. However, the energy consumption provided more relevant differences. According to this criterion the best results were clearly obtained by the GA implementations, outperforming the CC-Elevate algorithm by up to 45 kWh as average for lunchpeak traffic for the 120 simulated minutes, which is relevant. Table 6. Relative analysis of the obtained results. The detailed analysis of the interfloor traffic shows a better performance by both genetic algorithms for the AWT, AJT and ATT criteria, especially for those cases with higher complexity (higher buildings with larger elevator groups). This better behaviour is also appreciated for the energy consumption performance index, whose difference between the CC-Elevate and the genetic approaches increases with such complexity. Figure 11 depicts this analysis graphically. The x-axis shows the four considered configurations described in Table 1, and the y-axis the corresponding performance index. Figure 12 shows the radial chart representation. In such graphics a lower area corresponds to a general better performance. The area for both genetic approaches corresponds to the best results, especially in the more complex configurations. Figure 11. Performance indexes graphical results for the interfloor traffic pattern. AWT ATT AJT Energy AWT ATT AJT Energy AWT ATT AJT Energy 3,8% 1,3% BEST BEST 8,7% BEST 0,7% 0,4% BEST 3,3% 0,2% 3,2% 2,8% 1,7% 0,8% BEST 16,3% BEST 3,6% 1,0% BEST 1,7% BEST 3,8% 7,5% 1,1% 3,1% BEST BEST BEST BEST 4,6% 9,4% 1,7% 4,2% 10,6% BEST BEST BEST BEST 4,8% 0,3% 1,7% 4,2% 8,0% 5,9% 6,6% 16,4% AWT ATT AJT Energy AWT ATT AJT Energy AWT ATT AJT Energy 5,6% 1,7% 2,8% BEST 16,0% 2,4% 6,3% 0,5% BEST BEST BEST 62,9% 11,3% BEST 1,5% 1,8% 12,3% 0,3% 2,0% BEST BEST 2,0% BEST 57,9% 4,5% BEST 0,3% BEST 16,0% 0,4% 3,8% 1,3% BEST 1,4% BEST 46,2% 10,0% 0,5% 3,6% BEST 3,5% 0,2% 1,3% 6,5% BEST BEST BEST 51,5% Interfloor Lunchpeak GA - NC GA - ODT CC-Elevate GA - NC GA - ODT CC-Elevate Figure 12. Radial chart results for the interfloor traffic pattern. From the analysis of Tables 5 and 6 and Figures 11 and 12 we can conclude that the genetic algorithm with NC fitness evaluation performs slightly better than the genetic algorithm with the ODT as fitness function, although the absolute values are in the same order of magnitude for the four performance indexes. The analysis for lunchpeak traffic conditions shows a slightly better behaviour for the average waiting time (AWT) for the CC-Elevate algorithm. However, the absolute difference is just between 1 and 3 seconds, which is really short. In addition, this better performance is not appreciated for the transit and journey times (ATT and AJT). And the CC-Elevate is clearly outperformed by the genetic approaches in terms of energy consumption. In a similar line, we could conclude that the genetic algorithm considering the NC as fitness function provides the better overall performance. This analysis is also confirmed by the graphic analysis. See Figure 13 for the performance indexes results and Figure 14 for the radial chart that depicts clearly the significant energy savings. Figure 13. Performance indexes graphical results for the lunchpeak traffic pattern. Figure 14. Radial chart results for the lunchpeak traffic pattern. The analysis of the figures shows some conflicts between different criteria, e.g. the waiting and transit times, as well as the energy consumption. However, the results are consistent with the design of the different approaches. The CC-Elevate algorithm is mainly focused on minimising the AWT and therefore once a landing call is issued it intends to minimise such AWT by sending a dedicated and even empty elevator. This produces a higher consumption of energy. On the other hand, the GA-NC tries to promote the allocation to those nearby elevators travelling in the direction of the landing call issued at the halls, and finally the GA-ODT tries to balance the AWT and ATT through the computation of the ODT. Therefore, these genetic approaches produce a decrease in the energy consumption by saving energy in their movements along the building. Another relevant issue is related to the similar behaviour regarding the transit times (ATT) of all the approaches. However, a detailed analysis points the GA-ODT out as the best choice, since it is the only one considering the overall dispatching time including the expected time of travel of those passengers already inside the car that have issued their destination. However, note that this performance index (ATT) is specially considered in destination control systems and not in conventional control systems, which is the case we have considered in this paper. In destination control algorithms, the system knows in advance the destination and is able to conduct the allocation of cars to minimise transit times. This phenomenon cannot be reproduced in conventional control algorithms that only know the direction of travel. Finally, we can conclude that in vertical transportation neither algorithm will always perform best. This depends on the actual traffic conditions, and also depends on what is meant by "best", which will in turn depend on the service policy chosen by the owner or the manager of the building. However, if a unique recommendation should be given, the comprehensive analysis of all the tests will lead us to recommend the selection of the GA using NC fitness function as the best alternative. 6 Conclusions Double-deck elevators constitute a sound technical solution to service high numbers of users in large, high-rise buildings. These configurations are similar to two elevators moving together through the same shaft and servicing different floors at the same time. This incorporates a high level of additional difficulty to their management and control, particularly in the case of complex systems where several double-deck elevators coexist. These management and control processes are even more complex when multi-criteria considerations are incorporated, considering the minimization of energy consumption together with waiting times. And a further level of complexity is attained in the case of interfloor and lunchpeak traffic conditions, where the demand presents a more complex and unpredictable pattern than uppeak or downpeak scenarios, where a vast majority of the demand moves in the same direction. We have implemented a genetic algorithm considering two different fitness evaluation functions that provide different approaches to evaluate the conventional control double deck elevator architecture under interfloor and lunchpeak traffic conditions. The genetic algorithm was previously compared to a trajectory-based algorithm such as the tabu search algorithm depicting a better performance and validating its selection as a good approach to deal with the problem. The genetic algorithm was then compared to the CC-Elevate results that is already installed in the ELEVATE algorithm suite, i.e., the elevator traffic analysis and simulation software that we used for analysing the behaviour of the proposed algorithms. The selection of such traffic patterns was oriented to providing a detailed analysis of the behaviour of double deck architectures for alternative traffics different from uppeak and downpeak. The former has been widely studied and appears as especially adequate for double deck architectures, while the latter experiences an inverse behaviour. These traffic patterns offer the largest variability of movements and allow to better highlight the contribution of our proposals. In fact, several studies have proved that double deck elevators improve the handling capacity and the round trip time during uppeak traffic (the most relevant performance indexes for such traffic pattern) but not many works presented results attending to other traffic patterns. Particularly the analysis of interfloor traffic appears very rarely in the scientific literature, when it is the traffic pattern that takes place most of the time. Our experimentation was carried out attending to the most relevant criteria: the average waiting time, the average transit time, the average journey time and the energy consumption. The consideration of energy consumption in the analysis of double deck architectures has received very low attention in the scientific literature. The running times allow us to confirm the viability of real-life implementation of the two proposed approaches. Nevertheless, it is difficult to state that an algorithm is better than another in vertical transportation, since it depends on the actual traffic conditions and the building configuration. In any case, we can conclude that attending to the analysed configurations, the GA-NC provided the best performance. However, the results provided by GA-ODT were very well balanced between the four criteria producing the lowest area values in the radial graphics (Figures 9 and 11) for complex situations of interfloor and lunchpeak traffic. The main contributions of our paper can be listed as follows: a) The comprehensive and detailed analysis of conventional control double deck algorithms for non-dominant traffic patterns such as interfloor and lunchpeak traffic. These patterns have received very low attention by the scientific community, and interfloor traffic has not been previously considered in the scientific literature to the best of our knowledge for double deck elevators. b) The implementation of a novel performance evaluation function considering the overall dispatching time for conventional control systems through a constructive procedure, such as ODT. c) The adaptation of the NC evaluation function from single deck to double deck elevators. d) The consideration of the energy consumption for conventional control double deck elevators, which has not previously considered in the scientific literature to the best of our knowledge. Future research lines should consider the introduction of uncertainty in passengers’ movement, since we managed deterministic percentages, but its integration in ELEVATE software should require an adaptation of the simulation platform. To do so, improvements in the genetic algorithm to deal with such scearios could be based on fuzzy or neural approaches. Additionally, the consideration of Internet of Things as a new paradigm allowing real-time data capture should lead to a new generation of real-time algorithms. Researches based on integrating Internet of Things into double deck elevator group control systems have not been presented yet, although some papers are appearing for common architectures (see the introduction section). Finally, in terms of handling capacity and space savings, other constructive architectures based on multi-car elevators could be widely evaluated in a near future. References Al-Kodmany, K. (2015) ‘Tall buildings and elevators: A review of recent technological advances’, Buildings, Vol. 5, No. 3, pp. 1070-1104.