電力市場参加者モデルに関する研究：ファジイ推論システムを用いるエージェントモデルとリスク管理を考慮したエージェントモデルの提案

摘要

With the restructuring in electricity industries around the world, electricity markets have emerged in many countries. Under the market mechanism,generation companies (generators) and consumers are exposed to markets to meet their own needs, i.e. generators sell power to earn a pro t and consumers buy power for consumption. A market operator mediates by requiring participants to submit bids they desire to sell or buy for 24 hours of next day and then determining market price and dispatch schedules.Electricity markets are expected to promote competition and ultimately benefit consumers. However,the characteristics of electricity such as nonstorability, little demand side elasticity in short term, etc., make it difficult to apply traditional economic theory to analyze and predict market behavior.The strategic bidding behavior is often observed to severely influence market prices.Agent-based simulation is widely applied in social science and economics.Especially in electricity markets analysis, a participant is modeled as a rational and adaptive agent who learns to select bids that maximize its profit under various market conditions, thereby how local interaction among market participantsu27 bidding behavior influence global level market price can be investigated. Agent-based simulation also has the potential to treat heterogeneous costs, capacity limits, price caps and other practical issues directly.The thesis is motivated by the observation that participants in electricity markets are faced with various uncertainties such as fluctuating prices,varying demand, competitorsu27 strategic bidding behaviors etc. The information based on which participants make decisions is therefore imprecise and uncertain. Therefore the decision-making under uncertainty plays an important role in an agent model. However, so far, agent-based electricity market research has not dealt with such uncertainness explicitly. Moreover, verification of agent-based models, which is an important issue for agent-based modeling, is also not sufficiently explored.Fuzzy logic is proven to be a practical method to handle decision making problems under uncertainty. We propose to use a fuzzy inference system(FIS) to model a generatoru27s decision-making in bidding. To construct a fuzzy inference system, fuzzy sets and fuzzy rules need to be designed. Since the submitted bidu27s performance can be evaluated by the profit it brought about, such evaluative signals, rather than instructive training data, can be employed by reinforcement learning methods to construct an FIS. Symbiotic evolution is able to learn and tune FIS based on reinforcement signals and it is proved to be effective in a single agent situation. SE is adopted in our multi-agent environment to tune each agentu27s FIS.Experiments show that our model agrees with Cournot game and Bertrand game solutions in game theory methods if generatorsu27 strategies are restricted to quantity competition or price competition. The experiment results demonstrate the agreement of agent-based simulation with game theory analysis.Risk management is a critical issue for market participants. Therefore bidding strategies with risk management is strongly required by market participants. So far, bidding strategies with risk consideration are developed under the assumption of price-taker. However, a participant is not always a price-taker. Agent-based modeling can work well for situations without price-taker assumption. However, few effort is worked on agent models with explicit risk management, which are capable to help investigate how individualsu27 risk management influences market price.In the second part of our work, an agent model with explicit risk management for generators is proposed.Two risk metrics are considered: standard deviation, VaR (value at risk). An agent evaluates its action using weighted sum of expected return and risk for the action. Since the probability distribution for the reward received after the execution of an action is unknown, the estimates for VaR are based on historical data which should be memorized for evaluation. We apply the model to the analysis of market dynamics in electricity pool market. By experiments we show that generators with risk management bid differently compared to the case when they bid based on expected return. When genera-tors use standard deviation to measure risk, the variance in pro t is reduced at the cost of reduction of pro t. VaR metric can bring higher profit for generators compared to standard deviation measure.The model is applied to market simulation with discriminatory pricing and uniform pricing rules. Our experiment demonstrated that uniform pricing leads to higher variance in market price than discriminatory pricing when demand is high and uncertain.