Explainable deep reinforcement learning for resilient and battery-aware microgrid control

Microgrids support renewable integration and resilience, but operation is challenged by intermittency, forecast uncertainty, and battery-life constraints. This paper presents an eXplainable Deep Reinforcement Learning (XDRL) framework for hourly microgrid energy management under uncertainty, trained with Proximal Policy Optimisation (PPO). The reward jointly optimises a priority-weighted Resilience Index (RI) and a life-cycle-aware battery term, and robustness is strengthened via curriculum learning over progressively noisier scenarios. Post-hoc explainability using Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) is used to interpret how charging, discharging, and load-allocation decisions depend on net-energy balance, recent state of charge, and different load demands. Simulation on a cyclone-prone coastal microgrid in Kothapatnam, India (PV–wind–battery with prioritised loads) shows that, under uncertainty, the proposed policy achieves RI = 0.9956 (0.33% below an MPC benchmark of 0.9989) while increasing expected battery life by about 5% (15.9 vs. 15.1 years) and producing smoother SOC trajectories. From a computational perspective, online inference is about 5800 × cheaper than solving MPC at each step, and the 15-year lifetime compute cost (including one-off training) is approximately three times lower. A 4000-run Monte Carlo study confirms robustness (median RI 0.992, 5–95%: 0.984–0.999; median battery life ∼ 16 years).