Resilient microgrid operation under adverse weather conditions with deep reinforcement learning

The increasing occurrence of high-impact, low-frequency (HILF) events necessitates real-time dynamic response strategies in microgrid operations. This paper proposes a unified, real-time optimization framework to address the challenge of minimizing the impact of HILF events on microgrid operations. The four-stage approach integrates network reconfiguration, dispatchable distributed generator scheduling, and constrained deep reinforcement learning (DRL) to manage energy storage and flexible loads through an incentive-based demand response program. A constrained actor-critic DRL model ensures compliance with physical constraints while optimizing multiple objectives, including operational cost, network losses, and system resilience. The framework includes an HILF event simulation module that models adverse weather events and random line outages. Two case studies on a 33-bus microgrid with integrated renewable energy sources demonstrate the framework’s effectiveness under normal and HILF conditions. The uncertainties in renewable energy sources are addressed using fast interval optimization. Results show that the DRL-based demand response performs comparably to particle swarm optimization, with the added benefit of real-time adaptability. During simulated HILF events, the framework enhanced resilience and economic metrics by 9.27% and 8.13%, respectively, while reducing power losses. The DRL algorithm dynamically learns optimal control strategies for any set of line outages, ensuring robust microgrid operation.