報告書番号

R02014

タイトル（和文）

電力市場とCO2排出権市場の相互関連を考慮した取引シミュレーション -共進化型アルゴリズムを用いたマルチエージェントシミュレーションの学習方式の開発-

タイトル（英文）

Simulation of Electricity Market and CO2 Emission Market with Interdependency - Coevolution Algorithm for Learning Strategy in Multi-agent Simulation -

概要 （図表や脚注は「報告書全文」に掲載しております）

電力の生産には副産物としてCO2が排出されることから，2つの市場への参加者の電力市場での戦略と排出権市場での戦略は密接に関連する。さらに発電設備の投資戦略も2つの市場での戦略に大きく影響する。このため電力とCO2排出権の2つの市場をマルチエージェントシステム（MAS）で模擬する場合，エージェントが取り得る戦略の組合せの数は膨大となり，従来のMASで用いられる学習方式では，効率的に最適な戦略を学習することは困難となる。そこで，こうした場合にも効率的に最適な戦略が学習できる，共進化型アルゴリズムを用いた新たな学習方式を開発した。開発した学習方式を用いて電力とCO2排出権市場のシミュレーションを行ったところ，排出超過の罰則金を市場の取引価格を基準に決定する取引ルールの下では，罰則金が安くなるように市場価格を操作したうえで，排出権を大量に超過するCO2を排出する戦略を学習するエージェントが現れた。

概要 （英文）

This paper proposes a multi-agent simulation (MAS) system that can simulate an electricity market and an emission market. In the system, a participant in the markets, i.e. power company, is modeled as an autonomic agent that acts to minimize its cost. Because CO2 emissions are pumped out when electricity is generated, a bidding strategy of an agent in an electricity market depends on its bidding strategy in an emission market. Furthermore, the bidding strategies in the two markets depend on its capacity investment, because CO2 emissions and generating cost can be reduced by the introduction of new facilities. Therefore, the number of available strategies of an agent in the simulation becomes very huge, and it is hard to obtain optimal strategies of agents with the popular learning method of MAS.Thus we develop a new learning method of MAS that finds optimal strategies of agents using a coevolution algorithm. The coevolution algorithm is an algorithm that derives its behavior from a metaphor of the "host - parasite" relation. In the "host-parasite" relation, a host evolves its immunization against a parasite, and a parasite evolves to break down the immune system of the host at the same time. The coevolution algorithm aims to find a good solution (strong host/parasite) from the competitive race among the host and the parasite. In the proposed learning method, we use the coevolution algorithm to evolve the economical strategies of agents in competitive environment. To examine the performance of the proposed method, we applied the proposed method to games, in which optimal strategies can be obtained by traditional game theory. In the games, we attained strategies that are equivalent to the optimal strategies obtained by game theory, and the effectiveness of the proposed method was confirmed.Using the proposed MAS system, we simulated the trading in both an electricity market and a co2 emissions market. We first ran a simulation when penalty for the emission above the permits is 50% more than the average market price of emission.In this case, some agents learn a strategy that declines the penalty by manipulating the market price and emit the huge amount of co2 that exceeds their permits. Next, we ran a simulation when penalty for the emission above the permits is 20,000 yen, independent of market price. In this case, most of agents comply with their emission permits, and the market price of emission permits rises nearly fourfold. Furthermore, the electricity market price rises reflecting the trend of the emission price.

報告者

キーワード