Multi-agent Reinforcement Learning for Low-Carbon P2P Energy Trading among Self-Interested Microgrids

A team of researchers, including Junhao Ren, Honglin Gao, and others, has published a paper proposing a novel multi-agent reinforcement learning (MARL) framework designed to optimize peer-to-peer (P2P) electricity trading between autonomous, profit-driven microgrids. The core challenge addressed is the uncertainty in renewable generation (like solar and wind) and local demand, which complicates day-ahead energy scheduling. In this system, each microgrid acts as an independent AI agent that learns to bid both a price and a quantity of energy into a local market, aiming to maximize its own financial return through strategic use of its energy storage for arbitrage against time-varying main-grid prices.

The proposed framework includes a central market-clearing mechanism that processes all bids to finalize trades, designed to be incentive-compatible—meaning it encourages microgrids to report their true costs and capabilities. Simulation results, as reported in the paper accepted by IEEE ICC 2026, demonstrate that the AI-learned bidding strategies lead to a "win-win" outcome. The policy successfully increases the utilization of locally generated renewable energy within the community of microgrids, thereby reducing dependence on and purchases from the often higher-carbon main electricity grid. Concurrently, this efficient local trading boosts the collective economic welfare of the participating microgrids, proving that decentralized, AI-coordinated markets can align individual profit motives with broader environmental goals like carbon reduction.