Quicker to Train and Cheaper to Run: Edge-Ready DRL for Microgrid Resilient Energy Management

Resilient microgrid control only delivers value in remote settings if it can be trained quickly and run on resource-constrained hardware. This paper applies Simple Policy Optimisation (SPO) to energy management and benchmarks it against Proximal Policy Optimisation (PPO) and a Model Predictive Control (MPC) baseline under historical Cyclone Laila conditions. A short two-stage curriculum first secures resilience, then trades within the feasible set to extend battery life using either a balanced summation reward or a Lagrangian resilience constraint; goal-based early stopping breaks training once targets are met. Under the same actor–critic scheme, SPO produced steadier updates and stronger battery preservation than PPO, achieving up to 19.7 expected years of battery life (32% longer than MPC and 7% longer than PPO) while maintaining near-perfect resilience (RI = 0.998, compared with 0.993 for PPO Lagrangian). Compared with PPO, training compute is reduced by nearly twelve times with similar inference cost, whereas MPC is around a thousand times heavier at inference. SPO therefore offers an edge-deployable, resilience-first controller that protects batteries without sacrificing continuity of supply.