A Coordinated Day-Ahead and Real-Time Networked Multi-Microgrid Scheduling Framework Considering System Robustness Enhancement: An ADMM-Assisted State Adversarial Deep Reinforcement Learning Approach

This study introduces a coordinated two-level scheduling framework for networked multi-microgrid systems, integrating day-ahead planning and real-time control under a “day-ahead guided real-time optimization” paradigm. The core innovation lies in harmonizing long-term planning with real-time adaptability to enhance both cost efficiency and system robustness. At the day-ahead level, this work proposes a hybrid algorithm that combines State-Adversarial Deep Reinforcement Learning (SA-DRL) with the Alternating Direction Method of Multipliers (ADMM). SA-DRL enables each microgrid to learn robust scheduling strategies in the presence of potential state disturbances, while ADMM ensures distributed coordination among microgrids for optimal energy exchange. At the real-time level, control policies derived from the trained SA-DRL models are refined using a dynamic penalty mechanism, aligning real-time actions with day-ahead plans to reduce forecast-induced uncertainty. Furthermore, an extended Long Short-Term Memory (xLSTM) network is employed for multi-feature rolling prediction of renewable generation and load demand, supporting both scheduling levels with more accurate temporal forecasting. By integrating prediction, learning, and optimization within a two-level architecture, the framework enhances resilience against data noise, cyber-physical threats, and temporal coupling. Experimental evaluations confirm that the proposed method reduces operating costs and improves both control stability and system robustness.