Optimization for resource management using multi-agent systems

(1)

Dissertations in Forestry and Natural Sciences

DISSERTATIONS | THOMAS HO CHEE TAT | OPTIMIZATION FOR RESOURCE MANAGEMENT USING MULTI-AGENT SYSTEMS | No 4

THOMAS HO CHEE TAT

Optimization for resource management using multi-agent systems

PUBLICATIONS OF

THE UNIVERSITY OF EASTERN FINLAND

(2)

(3)

PUBLICATIONS OF THE UNIVERSITY OF EASTERN FINLAND DISSERTATIONS IN FORESTRY AND NATURAL SCIENCES

N:o 421

Thomas Ho Chee Tat

OPTIMIZATION FOR RESOURCE MANAGEMENT USING MULTI-AGENT

SYSTEMS

ACADEMIC DISSERTATION

To be presented by the permission of the Faculty of Science and Forestry for public examination in the University of Eastern Finland, Joensuu, on March 30, 2021, at 10 o’clock.

University of Eastern Finland School of Computing

Joensuu 2021

(4)

Grano Oy Jyväskylä, 2021

Editors: Pertti Pasanen, Matti Tedre, Jukka Tuomela, and Matti Vornanen

Distribution:

University of Eastern Finland Library / Sales of publications julkaisumyynti@uef.fi

http://www.uef.fi/kirjasto

ISBN: 978-952-61-3739-1 (print) ISSNL: 1798-5668

ISSN: 1798-5668 ISBN: 978-952-61-3740-7 (pdf)

ISSNL: 1798-5668 ISSN: 1798-5676

(5)

Author’s address: University of Eastern Finland School of Computing

P.O. Box 111 80101 JOENSUU FINLAND

email: thomasho@uef.fi Supervisors: Professor Pasi Fränti

University of Eastern Finland School of Computing

P.O. Box 111 80101 JOENSUU FINLAND

email: franti@uef.fi

Reviewers: Senior Lecturer John Page

University of New South Wales

School of Mechanical and Manufacturing Engineering 2052 NSW SYDNEY

AUSTRALIA

email: j.page@unsw.edu.au Dr. Matti Kutila

VTT Technical Research Centre of Finland 02150 Espoo

FINLAND

email: matti.kutila@vtt.fi

Opponent: Dr. Janne Sorsa

KONE Industrial Ltd 02150 Espoo

FINLAND

email: Janne.Sorsa@kone.com

(6)

(7)

Thomas Ho Chee Tat

Optimization for Resource Management Using Multi-Agent Systems Joensuu: University of Eastern Finland, 2021

Publications of the University of Eastern Finland Dissertations in Forestry and Natural Sciences

ABSTRACT

In our fast-paced modernizing world, there are many proposals for building infrastructure to improve human lives and urban livability. However, such endeavors require substantial amounts of effort and resources; if the proposal fails, then the investments are wasted. Furthermore, an infrastructure project’s operational resource consumption and environmental effects must be managed effectively. These issues must be addressed before undertaking extensive infrastructure projects. Estimat- ing an infrastructure project’s operational resource consumption is more complex than planning its construction, as the outcome depends on intermediary decisions over time. Hence, the methodology should be researched and formalized prior to undertaking an infrastructure project.

A methodological framework is important, as it offers insight to help us better manage building resources for a massive infrastructure project. Furthermore, it will offer insights into the waste or byproducts generated so that we can better manage them and avoid destroying the environment. Since individual human decisions are the main determinant in this research problem, we used the artificial intelligence technique of a multi-agent system in the methodology. A multi-agent system simulates the causes and effects of decisions made by intelligent agents modeled after humans in a controlled environment. This technique links one effect to one cause in a deterministic manner; everything happens for a reason and is explainable. Insights can be derived from the effects, which can be helpful in preparing for upcoming situations. In this study, we examine the impacts of closing critical infrastructure on the operations of a township. The results could help municipal governments manage similar situations effectively.

Some insights gained from the data are pertinent to Singapore’s energy optimization efforts. For instance, we found that Singapore’s energy supply is sufficient to fulfill a complete car electrification. We also found that closing certain bridges in Joensuu, Finland will impact the traffic situation. We verified the synthesized data against real data. Finally, we devised a framework to synthesize non-existing systems’ data for impact analysis and apply it to another kind of traffic system. All the studies were progressively done to reduce Singapore’s energy consumption, beginning with cars and then moving on to elevators. Elevators similar to cars, but they run on vertical tracks and have unpredictable energy consumption due to uncertain human usage in buildings serving many purposes.

Universal Decimal Classification:005.591.1, 004.85, 332.144, 519.863, 711.4, 711.7

Library of Congress Subject Headings:Mathematical optimization; Multiagent systems;

Intelligent agents (Computer software); Artificial intelligence; Machine learning; Computer simulation; Forecasting; Causation; Infrastructure (Economics); Energy conservation; En- ergy consumption; Electric power consumption; Urban transportation; City traffic; Electric

(8)

vehicles; Elevators; Singapore

Yleinen suomalainen ontologia: optimointi; älykkäät agentit; tekoäly; koneoppiminen;

simulointi; ennusteet; kausaliteetti; infrastruktuurit; resurssit; energiansäästö; sähkönkulu- tus; liikennejärjestelmät; kaupunkiliikenne; sähköajoneuvot; hissit; Singapore

(9)

ACKNOWLEDGEMENTS

First, I would like to thank Professor Pasi Fränti for giving me the wonderful opportunity to be his student at the University of Eastern Finland. He has provided valuable and generous insights and guidance during my studies. He spent considerable time and effort during my publication endeavors to ensure that the papers would be accepted and published.

Next, I would like to thank Dr. Susanto Rahardja for recommending me to Pro- fessor Pasi Fränti to be his PhD student. Without his guidance, I would not have had the opportunity to study in the University of Eastern Finland.

Finally, I would also like to thank my family, friends and colleagues for their en- couragement and generosity that they have unselfishly given so that I could pursue the PhD wholeheartedly. I am especially grateful for my ex-colleagues in the Insti- tute for Infocomm Research (I2R) whom offered valuable advice during my doctoral program. To all of you, I offer you my gratitude.

Joensuu, February 20, 2021 Thomas Ho Chee Tat

(10)

(11)

LIST OF PUBLICATIONS

This thesis consists of the present review of the author’s work in the field of optimization using multi-agent systems, and the following is a selection of the author’s publications:

I T.C.T Ho, R.S. Yu, J.R. Lim and P. Fränti, "Modelling implications & impacts of going green with EV in Singapore with multi-agent systems," Signal and Information Processing Association Annual Summit and Conference (APSIPA), Asia- Pacific, 1–6 (2014).

II W.X. Yang, T.C.T. Ho, L. Xiang Liu, C.C. Chai and R.S Yu, "An overview and evaluation on demand response program in Singapore electricity market,"

IEEE Conference on Energy Conversion (CENCON), 61–66 (2014).

III T.C.T Ho and P. Fränti, "Real-time Electric Vehicle Load Forecast to Meet Timely Energy Dispatch," IEEE International Conference on Service Operations and Logistics, and Informatics (SOLI), 148–153 (2018).

IV T.C.T Ho and P. Fränti, "Multi-Agent Approach Traffic Forecast for Planning Urban Road Infrastructure," TENCON IEEE Region 10 Conference, 1795–1800 (2018).

V J.H. Zheng, T.C.T Ho and H.B. Yuan, "Traffic Prediction for Efficient Elevator Dispatching,"TENCON IEEE Region 10 Conference, 2232–2236 (2018).

Throughout the overview, these papers will be referred to by Roman numerals.

(12)

AUTHOR’S CONTRIBUTION

I The idea was jointly created and developed by the author and co-authors. The author built the framework, synthesized the data and executed the experiments. The problem was provided by the Singapore Power Group Pte Ltd.

II The author built the multi-agent system for the experiments and participated in writing the paper. He wrote about the artificial intelligence used in building the multi-agent system used and the experiments performed using the system.

III The idea was created and developed by the author. It was refined by the co- author. The author designed and executed the experiments, as well as built the system required to conduct the experiments.

IV The idea was jointly created and developed by both authors. The author built the framework, synthesized the data, designed and executed the experiments.

The co-author suggested that we broaden the scope of the experiments.

V The authors jointly developed the idea. The author designed the multi-agent system for the experiments. He participated in the writing of the paper and described the methods used by the system.

(13)

1 INTRODUCTION

1.1 IMPORTANCE OF RESOURCE MANAGEMENT

The world’s resources are ﬁnite. However, mankind’s industrialization efforts over the past century were not focused on resource conservation. Callous acts of wastage have been the main cause of excessive resource usage, leading to many problems, such as depleting energy, water, food and other natural resources. This wastage also created secondary problems such as increased carbon dioxide (CO2) and reduced forested land. Impacts from secondary problems are felt in the environment, including global warming due to greenhouse effects caused by excessiveCO2 and reduced oxygen production. These effects have also generated derivative resources like carbon credits and land usage limits for urbanization.

1.2 ENERGY WASTAGE

Figure 1.1:Rising energy needs with industrialization since the 1800s.

Energy production and the consumption of fossil fuels began in the 18th century

(16)

due to the first industrial revolution [1] and escalated after World War II (WWII) and decolonization. Prior to WWII, energy consumption was mainly used for powering locomotives, heating homes and cooking [2]. Energy production mainly involved burning wood and coal to power steam engines, as well as stoves and fireplaces to cook and provide heat [2]. In the late 1940s, energy consumption surged, as shown in Fig. 1.1 due to massive adoption of electricity in manufacturing [2]. Moreover, to accommodate the growing demand, the world turned to fossil fuels, such as petroleum and natural gas, as they are more efficient in producing heat than wood and coal.

Figure 1.2:Rising percentage of carbon dioxide in the atmosphere since 1800s. [3]

Increased greenhouse gas emissions, especiallyCO2, ensued similarly during the post-war industrialization period. This trend is shown in Fig. 1.2, as the rate ofCO2

emissions rose sharply after 1950. The sharp increase in CO2 emissions has led Shaﬁee and Topal [4] to speculate that the world’s oil, coal, and natural gas reserves will only last another 35, 107, and 37 years, respectively. The sudden increase in the accumulation of CO₂in the world’s atmosphere had an adverse effect on global temperatures.

As seen in Fig. 1.3, the sudden increase in the global temperature was no mere coincidence but a result of mankind’s sudden industrialization. In this century, we are beginning to feel its effects. The melting of Muir Glacier in Alaska is one of the many visible consequences of the rising global temperatures, as shown in Fig. 1.4.

(17)

Figure 1.3: Rising global temperatures from 1880 to 2012 with NASA projection from 2012 to 2020. Image source: NASA

Moreover, the accelerated energy consumption could also be due to its unnecessary use. For example, leaving the room lights or air-conditioning on when no one is

Figure 1.4: Comparison of Muir Glacier between 1941 and 2004. Image source:

NASA

present. This is particular true for commercial buildings, where a switch controls a section of lights. If a single person works alone at night, the whole section is lit. Nu-

(18)

merous energy conservation campaigns have been undertaken over the years, and all of them have reminded people not to waste energy.

Energy mismanagement can be seen in automobiles today that run on petrol.

Over the past decade, there has been a sharp increase in Singapore’s population, which also increased the number of vehicles on the roads. Currently, most of the vehicles on the roads in Singapore areinternal combustion engine vehicles(ICEVs), referred as normal cars, but there are someplug-in hybrid electric vehicles(PHEVs) and electric vehicles(EVs). Thus, the total energy consumption by electric vehicles has not been significant. At the end of 2019, there were only 1,120 EVs and compared to the 615,452 ICEVs on Singapore’s roads [5].

EVs have many advantages over ICEVs. Most importantly, energy generation can now be consolidated, which allows for the use of cleaner fuel resources and provides a focal point for meaningful carbon recovery fromCO2. Hence, if all of Singapore’s ICEVs were converted to EVs, energy consumption and carbon recovery could be managed more effectively. Energy for EVs could be produced by power generation plants. However, given the difficulty of predicting traffic conditions and the weather, it is difficult to plan how much energy is required for EVs on an hourly basis. Moreover, energy must be consumed when it is produced. Over-generation causes wastage, whereas under-production means some EVs will not have adequate power. Hence, energy regulation is also important as an irregular supply like solar as it can destabilize the grid. Therefore, optimized energy management is becoming increasingly important.

As the world becomes more urbanized, many cities are growing upwards. This is especially true for cities like Singapore, where space is limited. Elevators are becoming the first-and-last mile transportation for people living and working in high-rise buildings with mixed functions. These buildings serve many functions, such as residential, commercial and transportation. Such buildings are regional hubs in Singapore’s northern, eastern and western areas [6]. Therefore, we should manage the increased elevator energy consumption as early as possible for effective energy resource management.

1.3 ADVANTAGES OF OPTIMIZATION

In general, optimization is the process of finding the best value that fits an objective function, which is defined by a set of constrained valuables [7]. For example, the main advantage of energy optimization is fulfilling consumption needs while minimizing the fossil fuel requirements for its production. However, peripheral or secondary effects may be generated. For example, in an energy grid, reducing the peak or ’filling the valley’ in daily energy load reduces stress on the power stations’

energy generators as explained in paperII.

Energy production optimization generates secondary benefits. For example, load-following power plants operate in direct response to changes in power demand. They either shut down completely or curtail production when demand is low. A power plant owns multiple power generators, which are categorized based on their level of efficiency. Hence, when future energy requirements are known

(19)

prior to production, engineers in power generation plants can determine the most efficient sequence for turning on generators to meet demand [8, 9]. This effectively minimizes the required number of generators needed, thus reducing the amount of time power generators are run unnecessarily, maintenance costs [10–13] and equipment downtime [14]. These are some of the many secondary benefits of energy optimization.

1.4 RESEARCH CHALLENGES

Traditionally, the success of a solution can only be determined after its implementa- tion or deployment. In real world solution deployments, regardless of size, financial investments are considered well spent if the goals are met. The measures of success are found in the data collected over time when the deployment is in operation.

Hence, this study seeks to meet the following objectives:

• Determine the model’s success factors

• Determine the method for data aggregation

• Generate the artificial model

• Generalize the created model

First, we develop a systematic approach to study the scenario of a non-existent de- ployed system consistently. Then, we identify the constraints and processes that best represent the scenario. Next, we identify tools that can synthesize data that represents the scenario. Finally, we generalize the processes to create a framework that can help manage resources. Each of these four objectives has its own challenges.

The first challenge presented here is how to determine success without real- world deployment. With reference to EVs and elevators, the first challenge frames how we determine the amount of energy required to power a nation’s electrified transportation system and elevators in all buildings with mixed functions before they are built.

To determine success, data is needed to justify the result. Since there is no real- world deployment, the second challenge is how to obtain the data. There exist many ways to synthesize data for the above purpose. However, the data need to have properties similar to the real-world data. The synthetic data need to include the effects of causal relationships, similar to those found in the real world. Some of them may be complicated, which means an initial action may lead to a chain of cause and effects that are link to one another consecutively. Hence, we need to closely model the required simulation to the real-world situation. In this study, we need to map EVs on a road network system and elevator systems in a building sep- arately but accounting for possible daily variations in both cases.

In creating data that meet the above requirements, the third challenge is to search for a model capable of performing this function. Empirical methods employ measurements and observations made under different conditions to build models. How- ever, due to the lack of real-world deployment, empirical evidence is unavailable.

Statistical methods translate problems into constraints and equations to create models. However, it is difficult to define all equations to include all real world possibili- ties that will characterize the data. We need to find a data creation model capable of

(20)

taking in constraints without losing flexibility to account for uncertainties in human behavior.

The fourth challenge is to generalize the model so that it can be applied to similar problem statements. There is also a need to find relevant real-world information to support the generalized model. Hence, we need to consider the commonalities in the steps used in each case study and set them in a serial chain of processes.

The remaining thesis is structured as follows. In Chapter 2, we illustrate the conceptual framework adopted to optimize the efficient use of resources. It does so by explaining the sequential steps of modeling a research problem with real-world details, assumptions and constraints and executing a simulation. The final step in achieving the optimized outcome is through forecasting with data generated from the simulation.

In Chapter 3, we examine three methods of forecasting and discuss the approach used to select the most appropriate method to achieve the required outcome. In Chapter 4, we illustrate applicable areas where benefits can be obtained when applying the methodology formulated herein. In Chapter 5, we summarize the author’s contributions to this study. The thesis concludes with Chapter 6, in which we suggest future avenues of research.

(21)

2 OPTIMIZATION

Figure 2.1:Optimization framework used in this study.

The most effective way to optimize the required resources for a system is to know the precise requirements in advance. This is trivial for a system that has processes consuming constant resources in a repeating pattern. However, it becomes difficult when the processes’ consumption start to vary, as the necessary data are not readily available. In light of resource depletion, there is a strong desire to optimize systems for reducing resource wastage.

Fig. 2.1 shows the framework used for optimization in this study. To optimize systems, two important elements must be available: historical data that capture trends intrinsic to the system (’Data’ in Fig. 2.1) and forecast methods capable of accurately predicting future data (’Insights’ in Fig. 2.1). For the former, if the system does not exist in real world, we will need to synthesize the relevant data (’Problem’

and ’Modeling’ in Fig. 2.1). For the latter, simple methods (e.g., linear regression) and more complex ones (e.g., neural networks, support vector machines) exist for forecasting.

2.1 SYNTHETIC DATA

Synthetic data are acceptable in the absence of actual data, as it has similar attributes, relationships, and other factors. There are many ways to synthesize data. However, to solve real-world problems, synthesized data need to contain the same flexibility

(22)

and unpredictability as actual data.

Many ways exist to synthesize data. Classical approaches are either empirical [15, 16] or stochastic [17, 18], as shown in other fields, because massive energy consumption measurements for EVs do not exist yet. Empirical approaches involve structural and well-defined processes that produce verifiable observations and evidence. They are not based on pure theory or logic. They are measured using physical means, such as measuring temperature with a thermometer, length with a ruler, or, as in paperIV, counting human traffic with sensors. The evidence produced by these approaches must be valid and reliable. Validity means the investigation is applicable to the problem statement or query. Reliability means similar observations are reproducible under the same conditions [15, 16]. However, an existing physical model must first be in place so that measurements can be taken to build the empirical model.

As previously stated, the physical model needs to exist for taking measurements to build a data-synthesizing empirical model. Currently, most cars are not built to log travel data for each journey. Some cars may have additional sensors to log such data, but they are not representative of the car population. Although event data recorders exists, they do not capture vehicle data under normal operations on a long-term basis [19]. Even if data loggers could be easily installed, considerable effort would still be required to enlist cars to participate in this event-capturing research. Hence, an empirical model is not realistically feasible at this time.

Stochastic approaches involve random processes that produce observations based on probability. They describe systems that change in undefinable ways, such as where the ball will land on a roulette wheel or the hourly total residential energy load of a given community. In these examples, the result space is bound, but how each observation is made cannot be determined. Each observation is independent [17, 18]. Hence, in artificial intelligence,simulated annealing[20, 21],neural networks[22,23] andgenetic algorithms[24,25] which involves the use of probabilities to solve problems are considered stochastic methods.

Since logging journey information can be difficult, many researchers use stochastic methods to model and perform data traffic flow analysis [26–28] because no real- world deployment is required to acquire similar related data. As actual massive deployment of data logging sensors can be exorbitant and time-consuming, stochastic methods are well suited for modeling road traffic.

However, empirical approaches cannot be used to generate energy consumption data for traveling EVs because road network systems capable of generating the data have not been built. Hence, stochastic approaches, which involve modeling the system and estimating its outputs, are currently the only available option. The method chosen to synthesize the data depends on the modeled situation.

In papers II and V, Monte Carlo [29, 30] and Poisson [31, 32] distributions are used, respectively, in conjunction with multi-agent systems to synthesis the data required for model analysis. A Monte Carlo distribution process is a computerized mathematical method that augments analysis and decision-making by accounting for uncertainty. It considers a set of possible outcomes and simulates a likelihood

(23)

table for each of them by re-calculating their likelihoods over many trials. Using a Monte Carlo distribution is advantageous because it shows not only the possible outcomes but also the likelihood of each outcome. It is also able to show the relationship between inputs and outcomes [30]. It was used in paperIIto model the power-brokers’ participation in Singapore’s energy market using a newly introduced curtailment program as an energy brokering instrument. Monte Carlo distributions were used to synthesize users’ likelihood of successfully curtailing their energy ex- penses when required by the curtailment program.

A Poisson distribution describes the likelihoods for a set of independent occur- rences at fixed intervals in a discrete probability distribution [31]. In paperV, a Poisson distribution was suitable for synthesizing the number of passengers arriving at the lift lobby on the various floors of a building. The number of passengers per arrival is independent, which fits the description of how a Poisson should be used [32]. In paperVthe distribution accounted for the types of businesses on each floor, as it affects the various peak lift-usage times. For example, buildings with product service centers operating during the lunch hour may have more lift activity than a building without such businesses

More complex methods have been used in modeling for data synthesis. Artifical neural networkshave been used to model steel structures and create data for reliability analysis [33]. Genetic algorithms have been used to create wind conditions data for the optimization of wind turbine design [34]. However, in this study, we are modeling human behavior and need a method that does not make random and arbitrary decisions. Furthermore, we must be able to account for uncertainty and make deterministic adaptive changes during the simulation. Multi-agent systems meet the requirements of this study, and we will use them to explore and perform data synthesis.

2.2 SYNTHESIZING DATA WITH MULTI-AGENT SYSTEMS

A multi-agent system (MAS) is a self-organized computerized system that has many intelligent agents interacting with one another solving problems that are difficult to solve by a single tight and rigid system [35–37]. The agents’ programmed intelligence can be procedural, functional, methodical or, in more complicated instances, use various methods, such as finite-state machine learning or algorithmic searches.

This means the agents’ intelligence is scalable from using a simple reactive rule- based decision-making mechanism to choosing the best option given by machine- learning algorithm. The MAS used in papersIIandVare simple and involve agents that only need to make decisions that have binary choices. In paperII, the agents simulating energy brokers only need to decide if they want to participate in the brokering. Likewise, in paperV, the passengers participating in the lift simulation only need to decide on which floor they want to disembark the passenger lift.

In complex situations where there are many considerations and agents can have more than binary choices, data can be synthesized by using a MAS that has agents with built-in artificial intelligence (AI) [38]. Their AI is explainable and deterministic. In papersIandIII, a MAS is used to simulate traffic movements on Singapore’s road network. Each vehicle’s velocity was measured while traveling on the road,

(24)

and the data was used" to improve clarity and flow here was used to estimate the corresponding energy consumption of all vehicles in Singapore. In paperIV, a MAS is used to simulate the before and after traffic movements of Joensuu, Finland with and without a new bridge. These results could be used in a study to determine the town’s road congestion if, for example, a bridge were closed for repairs.

A MAS can also have emergent behaviors, which are not programmed initially as part of the agents’ original intelligence; instead, they are an outcome of interactions, which result in a cause-and-effect type of relationship among the outcomes [39]. For instance, in papersIandIII, traffic congestion was observed as a result of all drivers on the road converging on common paths in the road network that they have chosen. In paperIV, new route choices that did not pass through town were observed due the closing of bridges. These behaviors need not be purposely considered in the simulation. Unlike cases where traditional stochastic modeling is used, the modeler will need to factor in such behaviors. This is an added benefit of using a MAS for modeling [39].

MAS are closed modeling simulations in which every action and reaction in it is deterministic. The agents’ and environment’s behaviors are all pre-programmed into the simulation [38]. This means the interactive agent-to-agent and agent-to- environment behaviors are explainable. However, the agent-to-agent and agent-to- environment interactions may have emergent behaviors that were incorporated into the modeling [39]. For example in paperIVwhen a bridge was closed, some roads became congested due to drivers selecting new routes. In papersI,IIIandIV, MAS was used to synthesize the data required for optimization. Furthermore, due to the model’s complexity, data pre-processing and post-processing were required so that it could be executed using a MAS and raw data produced by the MAS simulation could be meaningfully used.

In paperI, we begin to answer the question of energy load estimation for cars, as mentioned in Chapter 1. We begin to Explore different ways to achieve energy load estimation for cars. We surveyed extensively and found that A MAS could be useful in estimating the energy load of cars driven by humans, modeled after human driving behavior. To ensure that our synthesized data would be similar to actual data, we searched Singapore’s Land Transport Authority database, main- tained by the government agency in charge of managing road traffic in Singapore, for the country’s vehicle statistics. In paperI, actual Singapore’s transportation de- mographics were used (Table 1). In total, there were 957,006 vehicles on Singapore’s roads in 2018. However, not all them were modeled. Essentially, only vehicles that contribute to traffic congestion were chosen. For instance, motorcycles and scooters were not modeled, as the do not contribute significantly to traffic congestion, but other vehicle types were modeled. Some vehicle characteristics are accounted for in the model and can be implemented in the agents’ scheduling in the form of road network entry and exit timings, as explain in paper hjnmI.

The data used in this MAS modeling were combined with additional data, such as Singapore residents’ home locations and manpower distribution with work locations. This combination estimated the number of residents who drove vehicles and where they were likely to drive to during the day. Asorigin-destination(OD) matrices are efficient for keeping track of traffic conditions [40], we produced one with the

(25)

Table 1. Distribution of vehicles in Singapore. Image source: Land Transport Authority of Singapore

combined information described above. In addition, assumptions must be added here so that the MAS will account for them during execution. The assumptions below were used in building the MAS simulation model in paperI. Subsequently, the same MAS model was used in paperIII.

Assumption 1. All drivers stay in the designated residential areas.

Assumption 2. Jobs are distributed according to Singapore’s manpower statistics.

Assumption 3. All drivers work on the island.

Assumption 4. Every driver in Singapore will drive out at least once per day.

A caveat exists for Assumption 3, as the number of Singaporeans working in nearby Malaysia is small and does not affect traffic conditions significantly. There- fore, we did not consider its effects in the simulation. To improve the simulation’s

(26)

accuracy, we added the drivers’ lifestyle information. In papersIandIII, working- class drivers may travel between home and work regularly every day and may visit other places for dinner and other leisure activities after work. Hence, we model three different group of drivers for this simulation. The first group drives to work and home only, making two trips per day. The second group may drive to work and one other destination. The last group drives to multiple locations in a day. This level of detail makes the simulation more similar to real-life daily situations and may be inconsequential at this point in the thesis, but it will impact the processing time consideration later in Section 4.3. The pseudo codes for generating the cars’

departure schedules are provided in Section A.1 of the Appendix. The schedules generated serve as the cars’ entrance time into the road network system and do not define their exit times because we do not want cases of cars disappearing off the road network system due to road congestion and being mysteriously transported to their end destinations.

The most important reason we use a MAS for our study is because each agent can be programmed with intelligence, and each agent is capable of acting indepen- dently according to its surroundings [35]. We want to study human behavior during driving; hence, we only program driving-related behaviors into our MAS. These behaviors are used for route planning in the pre-journey stage and some reactions required during the journey, such as taking detours due to traffic delays or route unavailability in the original planned route. Intelligence for post-journey behavior is not required, as it does not affect the journey. What drivers do and how they react to situations at their destinations are inconsequential to the journey data in our study. These are high-level driving behaviors. Low-level behaviors, such as vehicle acceleration, turning and stopping should be programmed into the agent if applicable.

Many MAS, such as NetLogo [41], Swarm [42], Repast [43] and Mason [44] are available for our study. These are basic simulators that require users to build everything from scratch. Hence, using them would be excessively time- and effort- consuming for our purposes. Multi-agent simulators specific for traffic studies, such as SuMO [45], FreeSim [46] and MATSim [47] are also available. These simulators are more suitable for our study. They have built-in packages and libraries, including different types of vehicles and roads, as well as junctions and traffic light timing systems specific to traffic simulations. They also have built-in rudimentary driver intelligence,such as acceleration, turning and stopping. These built-in functions save tremendous time and effort in creating the traffic model’s environment and the agents’ programmed intelligence. There are many multi-agent simulators built for different purposes, such asMASGriP[48] for smart-grid modeling andMA- SCEM[49] for modeling competitive electricity markets. Hence, choosing the right MAS for the right purpose saves time and effort.

MATSim is used as the default traffic MAS; this simulator can simulate many cars concurrently in a real-world road network system [47]. It was compared to other traffic simulators in paperIand was found to be the best in terms of degrees- of-freedom for programming the cars’ action on road networks. More importantly, MATSim allows users to program the drivers’ intelligence and choose the route creation mechanism for route planning. MATSim plans all routes with the shortest path algorithm [50] but gives users the option to travel by major or minor roads or

(27)

avoid tolls. It also allows users to specify whether drivers will take an alternative route when they encounter trafﬁc delays. These behaviors are very human-like. The processes and data described above for the synthesis of trafﬁc data are summarized and shown in Fig. 2.2. Recently, green open vehicle routing problem (GOVRP)"

models were used to optimize route scheduling [51, 52]. Our model is similar, but GOVRP models consider multiple variables to determine the shortest route. Hence, their solutions have considerations from a single dimension.

Figure 2.2:Synthesizing trafﬁc data with multi-agent system.

In paper III, we extend paperI’s work to create data for more days. However, we do not want to create data for arbitrary days. To build a relationship between each day of data, we used the hourly temperatures of consecutive days for the air- conditioning energy load calculation. The air conditioning system was found to be a major energy consumer compared to other electrical components in an EV;

temperature control consumes 35%, followed by the power-steering system (5%), the braking system (5%), and other systems (5%) [53]. Table 2 shows estimated energy consumption values for an EV [54].

(28)

Table 2. EV Electrical Components

Component Power Consump-

tion (Watts(W)) Power Steering 3000

Braking System 3000 Signaling System 42 Windscreen wipers 50

Audio System 20

The use of these values for data post-processing will be discussed in Further detail in Section 2.3. The key point of using this method of data synthesis in paperIII is that we can generate continuous daily data for day-ahead energy load forecasting.

This forecast is then used to optimize energy production and achieve the benefits explained in Chapter 1.

In paperIV, we build a new model using the same techniques learned in paper I. However, in paperIV we used the MAS to investigate change in drivers’ behavior when the roads they typically use during their work commute are unavailable.

Roads are closed due to, for example, accidents, repairs, and traffic diversions. In paperIV, the problem statement involves investigating changes in Joensuu’s road traffic conditions when one of the bridges crossing Joensuu’s Pielisjoki river is closed for repairs. This is especially important because there are only four bridges crossing Pielisjoki river in Joensuu. "The study in paperIVwas also conducted to compare the similarity of MAS synthesized data to its real-world counterpart, which is explained in further detail in Section 2.4.

After the MAS has finished with simulating the scenario, its raw output data may not be immediately usable. Sometimes, post-processing is required. In papers I-IV, post-processing is applied regardless of whether the MAS was used. The next section elaborates on the post-processing required in each paper.

2.3 SYNTHESIZED DATA POST-PROCESSING

Depending on the data needed for optimization, post-processing of the synthesized data may be required. In papersI,II,III andIV, data post-processing is done. In paper I, the MAS, MATSim, used to simulate cars moving on the roads of Singa- pore produced data related to the cars’ states every second. These data included traveling velocities, cars’ entry and exit times onto a road and roads traveled during a journey. We first need to transform the car-related data into energy load data for each journey. In the study conducted for paperI, the literature on calculating energy consumption is reviewed (e.g., [55, 56]). The current study is not focused on calculating the exact forces contributing to energy consumption of a car. However, we wanted a physical model rather than one based on estimation, so we adopted a hybrid model in which only the most pervasive kinetic and potential component is considered in the calculation. As the study topic involves EVs, we can assume that they will begin charging their batteries immediately after each trip, which transfers the load to the power grid, as EVs must draw energy from the power grid to re- charge their batteries. Charging is not trivial, as the type of EV battery, charging cable and current grid utilization affects the charging rate. Hence, the EV’s load is amortized to a charging load on the power grid. The pseudo codes used for data

(29)

post-processing can be found in Section A.2 of the Appendix.

The study described in paperIIused a MAS to determine whether a new energy curtailment program would benefit residential consumers if they chose to participate. The residents’ aggregated shiftable energy loads from, for example, washing machines, electric rice-cookers, water heaters and air-conditioners were considered in this study. The objective was to determine whether residents who participated in the curtailment program benefited from it. The MAS, NetLogo, used in this study produced data for electrical component time-of-use, duration and the load that needs to be curtailed. We first transform the time-of-use, duration and power ratings into energy load data. Then, with the curtailed load requirements, we transform the load into cost-benefit data. We aggregated the cost benefits over one month to determine whether the curtailment program is beneficial to residents.

The study described in paperIII, as explained in Section 2.2, is a expansion of the effort in paperI. Instead of producing an instance of the data, we are building time-series data for the energy load on the grid. We use the MAS to create many data instances and connect them with continuous days of weather data. We begin by transforming raw car data into energy consumption data, which are then trans- formed. Hence, the data post-processing steps in paper III are same as in paper paper I, except that in calculating energy consumption in the first step, we must ensure that consecutive days of weather are considered in the calculation.

In paperIV, we use the traffic modeling techniques learned in paperIto simulate traffic in Joensuu, Finland. Unlike papersIandIII, we are not calculating the energy load on the power grid. We use a MAS to investigate the effects of bridges on Joensuu’s traffic situation. We remove the use of them one at a time and investigate how it affects drivers’ behavior and behavior, as well as the resulting traffic situation on Joensuu’s roads. MATSim also monitors the number of cars on the roads when simulating traffic. There are a total of four bridges crossing the Pielisjoki River in the vicinity of Joensuu. We begin with a control simulation in which all bridges are usable. Then, we remove the bridges one at a time and simulate the traffic again.

After all simulations are finished, we post-process the number of cars that cross the bridge in hourly periods and use it for traffic flow data. Then, we perform comparison studies on the bridges. Hence, only one step of post-processing occurs in paper IV. Although not studied in paperIV, the energy load for EVs can also be calculated by applying the post-processing methods described in papersIandIII, but with two changes. We apply Joensuu’s hourly temperature and change the air-conditioning energy load formula to calculate heating or cooling energy consumption depending on the seasons.

2.4 SYNTHESIZED DATA VERIFICATION

Modeled and simulated data differ from real-world data. For the data produced by a MAS to have value, it must be close or similar to real life. Hence, the modeled and simulated data need to be compared to real-world data. In paperIV, the difference between simulated and real-world data is compared.

In paper IV, the MAS models Joensuu inhabitants’ daily routines to produce

(30)

traffic data. The main focus is on tracking the number of vehicles crossing the river using the four bridges. Real-world bridge utilization data was collected by automatic counters placed on the bridges. The two datasets were compared and initial assessments were performed. The data showed that morning patterns were not comparable to evening patterns. In both datasets, we observed that Sirkkalan- silta bridge was used more frequently than Suvantosilta bridge during the evening hours. Hence, when data used for modeling closely resemble real-world data, data produced by the simulation more closely approximates real-world data. To complete the optimization process for real world resource management, detailed forecast techniques are used, which will be discussed in Chapter 3.

(31)

3 FORECASTING

Minimizing delays while dispatching perishable resources requires knowing the level of demand in advance. To do this, we forecast the demand. Since we are implementing a system that will be used in real-time, the chosen technique must produce forecast results within a given time-frame.

3.1 STATISTICAL METHODS

Statistical forecasting methods consist oflinear[57],polynomial[58] and other types of non-linear regression [59, 60]. These regressive techniques are robust but only suitable for single-dimension data. Moreover, the data generated by the modeling are all time-series that exhibit different trends at various points in time. Simply applying these regressive methods to the time-series are not sufficient to capture changes in trends.

The data can be re-shaped through principal component analysis (PCA) [61] to make the salient information more visible. However, in the long-term, as the data are aggregated, salient information maybe lost or superseded. Moreover, examining various ways to re-shape data for real-time systems is tedious, and existing methods can perform this task. For completeness, these techniques are examined in this study. They can forecast quickly enough for real-world systems but produce an unacceptable margin of error as the data aggregates. These methods are discussion in the next section.

3.2 MACHINE LEARNING METHODS

As mentioned in the previous section, some methods can be used to learn salient information from datasets. These are machine learning methods. During the course of this study, three methods were very popular among the research community:multi- layer perceptron[62],recurrent neural networks[63] andlong short-term memory[64].

The premise for machine learning methods is that they sufficiently distill trends from the historical data and crystallize them into forecast values for the future.

In paperIII, historical EV charging data are fed to the machine learning methods and used to forecast future consumption values. The most important and decisive value is the amount of time needed to compute the forecast figures, as it deter- mines whether the machine learning method can be used for real-time system im- plementation. Secondarily, the accuracy and consistency of the forecast will also be considered for the system; a quick but inaccurate forecast is useless.

3.2.1 Multi-Layer Perceptron

Multi-layer perceptron(MLP) utilizes the feed-forwarding technique to infer relationships between inputs and outputs [62]. Layers can be added and used as intermedi-

(32)

aries for information mapping. Learning is supervised through back-propagation.

Traditionally, MLP is useful for solving classification problems. However, the model can be modified to perform forecasting. First, we separate the time-series into priori and its outcome data. We then run the MLP to classify the all the data. Subsequently, for forecasting, we postulate that the new priori data will fall into the classification.

Corresponding outcomes to those classified priori are likely to be the forecast value.

In paper III, we determined that the MLP works quickly with large datasets.

Computational speed depends on the number of intermediate layers used to per- colate the data for insights and does not degrade too much as data size increases drastically. Because of this property, MLPs have been used for solving time-sensitive problems, such as energy-related [65, 66] and environment-related [67] forecasting.

They are periodic and require half-hourly results re-calculated with the latest input data. The input data size can be large when a more accurate result is needed.

3.2.2 Recurrent Neural Network

Recurrent neural network(RNNs) are commonly used for forecasting. A RNN differs from a traditional artificial neural network (ANN). Its information flow is cyclic and directed, whereas it is linear in an ANN. Its accuracy is determined by a fitness or reward function rather than verification against labels provided with the input data [63].

However, in paper III, we found that a RNN does not work well with large datasets. It’s computation time increases exponentially as the data size increases linearly. Hence, RNNs are typically used to solve problems that are not bound to time constraints but require high accuracy, such as river flow [68] and long term energy load [69] forecasting. Based on these two publications, the calculated results need not be immediate.

3.2.3 Long-Short Term Memory

Long short-term memory (LSTM) is a neural network (NN) that can adaptively determine whether information from a distant or closer past has a greater effect on the future outcome [64]. This makes LSTM powerful because it can keep track of information entropy and effectively determine when to use it. LSTM’s computation time increases as the data size increases. However, its computation time does not increase exponentially like it does with a RNN.

LSTM works well for short-term forecasting problems that exhibit seasonal patterns, such as residential load and traffic conditions [70,71] because it can determine which trends affect the forecast. This is seen in the numerous publications involving the forecasting of seasonal energy load using LSTM [72, 73] on the 2014 Global Energy Forecast Competition Data [74].

3.3 DATA SYNTHESIS

In papersIandIII, we wanted to use the aggregated daily energy consumption of all electric vehicles in Singapore to compare the accuracy and efficiency of the forecast methods. However, this information is unavailable since, by the end of 2019, only

(33)

0.1% of the total vehicles in Singapore were EVs, which is an inadequate sample size for extrapolating to 100% EV use in the future. Therefore, we had to synthesize the data and used the method explained in Chapter 2. The data synthesis process is shown in Fig. 3.1.

Figure 3.1:Data generation processes for daily EV energy consumption.

Fig. 3.1 describes the ’Modeling’ and ’Data’ sections from Fig. 2.1. The modeling involved using Singapore residents’ information to create the required situation as input for the MATSim simulator. Collecting the data involved combining the cars’

distances, times and velocities with daily temperatures and drivers’ air-conditioning preferences within their cars to produce energy consumption data for each journey.

This information serves as the energy load added to the power grid once the journey ends. Furthermore, it can also be the localized load information for the local power distribution system. This is important as various sub networks of the power grid have different degrees of wear-and-time and capacitance. Hence, this information is useful to determine whether the local power distribution system requires upgrades before it can handle new energy loads.

Fig. 3.1 shows the processes that synthesize data for a day. To produce different daily energy load profiles for cars, we first add variance to the activities’ starting times. This allows each driver to start some time either before or after his or her usual starting time. However, we should be careful, as this may produce schedules that are impossible, such as starting the next activity before the previous one ends.

Just like in the real-world, we cannot begin a new car journey without ending the current one. Usually, for routine activities, we have a time at which we need to leave for the next activity. However, there is always a tendency for people to either leave a few minutes earlier or later. This type of variation can be added to the ’Modeling’

section. Changing the input data indirectly affects the simulation output for the

’Data’ section ultimately altering the final energy consumption data.

(34)

Changes can also be added in the ’Data’ section to create additional energy profiles. Instead of using hourly weather temperatures, we use a daily time-series covering one month. Likewise, we add variance to the drivers’ air-conditioning preferences because, on some days, they may prefer a lower or higher temperature than their usual one. Such small differences may not affect drivers individually, but it may affect the power grid as an aggregated load. Fig. 3.2 shows the final processes used to generate energy load profile data for different days. The final process in Fig.

Figure 3.2: Final electric vehicle energy consumption data generation processes in this study.

3.2 generates the cars’ energy load profile for one day. To produce a time series of load profiles over a one-month period, we take each month’s daily hourly temperatures and feed them to the energy load’s consumption. With the synthesized data, we are then able to move on to the next step of using the forecast methods studied in chapter 2.

3.4 FORECAST METHODS

In paperIII, the methods used in the comparison were RNN, MLP and LSTM. To compare their efficiency and accuracy, we used each method without optimizing their tuning parameters. This is because we did not want to spend additional computational time or risk perceived bias in evaluating its accuracy due to optimization.

Three methods are used to forecast the future energy consumption in the comparison study. Singapore’s energy market functions in periods of thirtyminutes, which limits the maximum time available to forecast the next period’s energy consumption. However, Singapore’s energy market has dictated that all related action- to-transact should be submitted five minutes before the end of a period. This further reduces the amount of time available to perform forecasting.

(35)

The synthesized data consisted of traveled distances and time spent on the roads by each vehicle, along with the average flow speed of each road segment on the network. We should note that previously, in Section 2.3, we studied several methods for calculating a car’s energy consumption. However, as this study’s main focus is not on optimizing a car’s energy consumption, we decided to use a simple physical hybrid model comprised of a kinetic and a potential energy component to provide some estimated data that reflect traveling and environmental effects for the purpose of forecasting. The data can be a crude estimate of the actual amount since what we need is the trend in daily energy consumption. Average EV energy consumption statistics are available online [75], and the total energy consumption can be calculated easily by multiplying the total mileage of all cars in the simulation against the EV energy consumption of choice. However, since MATSim can provide per second velocities for each car, we decided to make use of that detailed data to capture nu- ances in energy consumption, especially in start-stop situations when cars are stuck in congested traffic.

The engine and electrical devices are the two sources of energy consumption in vehicles. Kinetic force components for traveling consists of forces used to overcome, for example, tire rolling resistance, air resistance and inertia. However, as stated previously, we wanted to make use of MATSim generated granular velocities, which we calculated as the required force to overcome the air resistance faced by the car multiplied by the distance traveled in this study. This is a crude estimation, and it cannot provide a full accounting of cars’ energy consumption. Although this calculation could be improved in future work, we needed a method that accounts for all of the components affecting a car’s energy consumption and accepts detailed data as input. The drag force component [76], which is also the calculation for an object moving through a fluid body, is given as follows:

FD= ¹

2ρv²CDA (3.1)

whereF_Dis the drag force component generated,ρis the mass density of air,vis the velocity of the vehicle,C_D is the drag coefficient and Ais the reference area of the vehicle going against air. Energy consumed is thus the product of distance traveled dtand drag forceFD. This method is crude, as the equation given is based on wind axis. This implies that the velocity is not the car speed but the speed of the air that the car needs to overcome to move forward. This force is one of the contributing forces acting against the car. Perhaps a better force component for determining the force on the car’s axis is described in [55]. However, when the experiment was conducted, there was no information available about Singapore’s road grade in a public database. The force component can be easily modified in future work once the road grade information is known.

In addition to the energy consumed to power a vehicle during travel, its electrical components, such as air-conditioning, ignition, radio, and signaling lights, consume power. Most of a car’s electronic equipment has a fixed power consumption rating, except for the air-conditioning unit. This is because the energy consumed is proportionate to the work required to reduce the external air’s temperature to the driver’s preference before pumping into the car’s interior. And because outdoor temperatures can change hourly and are seasonal, we cannot use a fixed power consumption

(36)

rating for the air-conditioning system. This energy calculation is detailed in paperI.

The amount of data used for training is variable because we want to investigate increases in accuracy as more data are used for training the three methods.

Data increments are in 24-hour periods, and increases in data points are always in multiples of 48, as there are 48 half-hour periods in a day. The data consisted of 110 days of non-continuous daily time-series data in total. The data are separated into training, testing and verification datasets at a ratio of 6:3:1, respectively. As we are dealing with critical systems and the motivation in this study is to eliminate as much resource waste as possible, we want to ensure that the system is as accurate as possible.

The forecasting algorithms are built in Python. Libraries for data-science processing, such as keras [77], tensorflow [78], pandas, numpy and statsmodels were used. The forecasting algorithms were executed on a 3.8 Ghz quad-core processor with a 6 MB cache, 8 GB of memory and a NVIDIA GeForce GTX 1050Ti with 2GB of GDDR5 memory. Tensorflow invoked the CUDA library to expedite the machine- learning process of the algorithms.

The algorithms were used to forecast the next period’s data based on historical data up to the current period, given the length of time taken,dfor a single loop in a closed system and size of the historical data,n. The three methods must forecast the results before timedis up. This can be expressed as a function f where:

0≤n≤max∞f(σ,n)≤d, where

σ={LSTM,RNN,MLP} ^(3.2)

Hence, we maximize the amount of data used for improved accuracy in each method such that the computing time is d at most. The execution time for each method was recorded against the amount of data used for each experiment. The accuracy of each method was also recorded against the amount of data used. Further details are provided in paperIII.

(37)

4 REAL WORLD RESOURCE MANAGEMENTS

As explained in Chapter 1, the world’s resources are finite. In this chapter, the processes and methods explained in Chapters 2 and 3 are applied to real world situations to optimize resource consumption. We will also explore the application of scenario modeling, simulation, and forecasting to energy and traffic management.

4.1 ENERGY

Energy is life. Without it, everything will come to a stop. The world has become more reliant on it than ever, but leaders have recognized that conventional energy sources are finite. The framework describe in previous chapters can be helpful in optimizing energy consumption. Based on various power distribution analyses in the transmission and control literature, we conceptualized the energy production and consumption cycle shown in Fig. 4.1 [79–83].

Figure 4.1:Energy generation and consumption life cycle

Fig. 4.1 shows the factors and considerations affecting the production and consumption of energy. Delivery reliability is important because power loss will result in voltage dips, brownouts and blackouts in the grid. It is simple to always produce an energy surplus to ensure reliability, but this solution wastes energy, as the excess is grounded. Given that fuels are limited and using them excessively contributes to climate change in a harmful way, wasteful approaches to the operational production and consumption of energy are not acceptable.

(38)

A balance exists between energy production and consumption for optimal energy management. However, in a power transmission system, there are losses due to heat, resistance and other factors [84]. To prevent blackouts or brownouts in an energy distribution system, generators will produce a surplus. This is described in the below system:

P(Estimation) +P(Surplus) =C(Load) +C(Loss) (4.1) wherePrepresents production andCthe consumption energy functions described in Fig. 4.1. According to Singapore’s power systems operators,P(Surplus)is always an estimated percentage ofP(Estimation), andC(Loss)is always minimized. Hence, for the current system, it can be reduced to:

P(Estimation)uC(Load) (4.2)

The energy load is traditionally comprised of residential, commercial and industrial loads. Commercial and industrial loads are controllable. Residential loads, although not as controllable as the former two, can be predicted with some confidence. Its consumption follows a pattern with slight deviations. However, the introduction of EVs increases uncertainty in energy consumption patterns.

However, an EV’s load is not easy to estimate. It is a multi-variate estimation involving various factors, such as traffic flow, routes taken and various factors, such as. In paperI, we reported that the determinants are aggregations of sub-factors, as shown below:

Tra f f icFlow= f(Drivers⁰Schedule,Drivers⁰Activities) HumanDrivingBehavior= f⁰(Tra f f icFlowDecision,DrivingPattern)

VehicleEnergyLoad= f⁰⁰(Velocity,VehicleRe f erenceArea,Distance)

(4.3)

However, velocity is proportionate to traffic flow, while distance proportionate to the routes chosen and traveled. Routes can be selected before the journey begins. How- ever, because traffic congestion may be encountered during a simulation, drivers have the programmed intelligence of selecting an alternative route to avoid congestion. Hence, the final traveled distance may not be equivalent to the initial selected one. The vehicle’s energy load can be substituted as follows:

VehicleEnergyLoad= f⁰⁰(Tra f f icFlow,VehicleRe f erenceArea,HumanDrivingBehavior) (4.4) To ensure that the simulated model meets these requirements, the MAS described in paperIwas built. Each determinant was incorporated into different parts of the MAS simulation. For example, traffic flow planning occurred in the simulation’s pre-execution. Driver behavior is determined during execution, while energy load is estimated after it. This type of estimation is essential in energy resource management because accurate energy forecast reduce the need to activate the contingency power generators.

After obtaining vehicle traveling data from numerous MAS simulations and taking in account of temperature information, the energy loads of all EVs for various days are obtained. The data can then be used in lieu of real-world data to train various neural networks to forecast future data, as describe in paperIII.

(39)

4.2 TRAFFIC

4.2.1 Directing Road Usage

In paper I, the MAS simulation was primarily used to generate the traffic flow data.

It was used to calculate and estimate the energy load of electric vehicles traveling on roads. To further explore MAS simulations of traffic flow, the impact of road infrastructure allocations and de-allocations was examined in paperIV.

The initial motivation was to determine whether the data generated by MAS simulations of traffic flow closely approximate real-world conditions. In paperIV, our simulations focused on the number of vehicles crossing a river using four bridges.

The number of vehicles using the bridges was automatically counted by electronic sensors.

The collected data were compared with the MAS simulation’s to determine whether the synthesized data closely approximate real-world data. We observe some similarities to the number of cars crossing the river over a 24-hour period in both datasets. The time periods where bridge traffic increases and decreases in the simulations coincides with the same instances in real life.

The findings were sufficiently promising, so we decided to examine closing the bridges to traffic one at a time. This approach was used to simulate what would happen to the traffic flow if one of the bridges could not be used. Subsequently, we can examine the congestion that might occur due to bridge closing. s

4.2.2 Directing Human Behavior for Lift Traffic Management

MAS were originally used in video games; Space Invaders and Pac-man were pio- neers in its use of MAS. They became more prolific in real-time strategy video games until it became possible to model real-life human behavior. In paperV, a MAS modeling human traffic for lifts was created. A lift manufacturer wanted to develop an AI algorithm to better serve its customers. In addition to developing new functions, such as intelligent responses to emergency situations and self-predictive maintenance to manage lift engineer shortages, they wanted a new lift dispatch algorithm that would reduce lift passengers’ wait times and overall energy consumption.

We obtained lift passenger arrival time-series data from our lift manufacturer- collaborator. We tried to model the data. However, given the seemingly stochastic nature of the time-series data, it was difficult to determine which model had the best fit. Three probabilistic distribution models were examined: poisson, average and k-means clustering. A poisson distribution is not sufficiently granular to represent real-time human traffic. An Average distribution will not be able to simulate outlier data that exist in the real world. Traditional k-means clustering is unable to simulate the different daily stochastic arrivals of passengers to the lift lobby due to the Euclidean measure used to group the time-series data. In the end, measure used to group the time series data was changed to dynamic time warping for better grouping.

(40)

Moreover, throughout the day, each floor in the building has varying levels of human traffic. For example, in a typical office building, the ground floor experi- ences more passenger arrivals in the morning, while floors above the ground level will have more passenger traffic in the evening. This is because people are arriving for work in the morning and leaving from work in the evening. Other factors, such as the occupancy rates or business types on each floor is not considered as this research is developing a foundational lift dispatching scheduling algorithm that can be customized for each building.

We simulated a building with 3 lifts and 10 floors. Passenger arrivals are simulated with the given premises described previously. There are assumptions on the reaction of each lift to the arrival of passengers for group optimization:

Assumption 5. Passenger information, such as the destination floor, is available.

Assumption 6. The requested floor is served by only 1 lift at any point in time.

Assumption 7. At any point in time, each lift path is either going up or coming down.

Assumption 8. Each lift returns to its terminal floor after fulfilling all requests.

Assumption 9. The number of passengers will not exceed maximum allowable load.

Assumption 10. Each lift will only change its traveling direction once per trip.

A MAS that simulates the interactions between passenger arrivals and lift dispatches was built to collect data that will used to derive the lift dispatch optimization model. The MAS generated deterministic outcomes based on variation in passenger arrivals. Two datasets were created. The first set includes input data regarding passenger arrivals and the predetermined terminal floors. The second set contains output data of total distances traveled by the lifts. From the output, the shortest distance represents the most optimal route for lift dispatches. Corresponding predetermined terminal floors with the shortest distances resulted in the optimal settings for the lifts.

Hence, to minimize the energy load against the usage and distance traveled reduction, a lift scheduling algorithm was derived. However, it may have shortcom- ings. For instance, there may be a upper-bound to the amount of people per floor it can serve. Therefore, it may be possible to use this algorithm to alter the lift usage behaviors of people in the building.

4.3 MANAGING COMPUTATION FOR REAL-TIME NEEDS

For this study to benefit real-world resource management, we have to be mindful of the total process time required to produce the forecasts. Many types of simulation software support real-world applications. Multi-Agent Simulator of Competitive Electricity Markets (MASCEM) [49] and General Electric’s Multi-Area Production Simulation (MAPS) system [85] were created to help power system operators and energy traders perform real-time electricity grid and market operations. The Na- tional Energy Modeling Systems (NEMS) [86] helps power plants plan energy production strategies based on fuel prices. Long-range Energy Alternatives Planning (LEAP) [87] mitigates climate change effects through continuous real-time energy

Optimization for resource management using multi-agent systems

Dissertations in Forestry and Natural Sciences

THOMAS HO CHEE TAT

Optimization for resource management using multi-agent systems

Thomas Ho Chee Tat

OPTIMIZATION FOR RESOURCE MANAGEMENT USING MULTI-AGENT

SYSTEMS

TABLE OF CONTENTS

1 INTRODUCTION

2 OPTIMIZATION

3 FORECASTING

4 REAL WORLD RESOURCE MANAGEMENTS