Explainable Deep Reinforcement Learning for Multi-Agent Electricity Market Simulations

As electricity systems evolve in the light of increased volatility and market variety, understanding market dynamics through simulations becomes crucial. Deep reinforcement learning (DRL) in combination with agent-based models (ABM) progressively garners attention as it allows the modeling of strategic bidding behavior of electricity market participants. However, as DRL is a black-box model, the learned behavior of market participants is hardly explainable nor interpretable for modelers. We bridge the gap in explainability of DRL in agent-based electricity market simulations by leveraging explainable DRL methods. The reviewed literature underscores the novelty of this approach, especially in multi-agent DRL settings. A case study comparing DRL and rule-based bidding strategies within the German electricity market showcases our method's potential. By analyzing DRL bidding strategies of 118 competitive DRL agents with clustering approaches and DeepSHAP, we investigate the underlying factors driving agent decisions, contributing to the development of transparent ABMs.