| Summary: | This paper proposes a novel framework for conducting sealed-bid double auctions in power trading for multi-microgrid networks, addressing the critical challenge of jointly optimizing bidding decisions and battery scheduling under uncertainty in renewable energy generation and load demand. In contrast to existing approaches that treat these components independently, our method explicitly models their interdependency for maximizing trading efficiency. We assume a normal distribution of prediction errors and introduce an uncertainty range and a bid buffer capacity to account for expected variations in forecasted generation and load, enabling more robust coordination between bidding and storage operations. While Q-learning determines the exact bid, the feasible power availability or demand is derived from the uncertainty range, ensuring consistency between learned bidding decisions and forecast-aware constraints. The Q-learning relies solely on its historical bidding outcomes without attempting to predict the bids of other participants. In parallel, battery operations are optimized using a hybrid method that combines Genetic Algorithm (GA) and Simulated Annealing (SA), explicitly incorporating the bid buffer capacity to align scheduling with market commitments. We also propose a fully decentralized and tamper-resistant execution architecture based on a consortium blockchain, where multiple aggregator agents within each microgrid, representing renewable sources, loads, storage systems, and the bid agent, function as independent blockchain nodes. Simulation results on a benchmark microgrid system with Monte-Carlo modeled prediction errors demonstrate that the proposed approach significantly enhances both economic benefits and trading robustness compared to conventional frameworks.
|
|---|
