Scheduling Optimization of Hybrid Microgrid Generators Based on Deep Reinforcement Learning

Abstract

Unit scheduling in hybrid microgrids (PV/wind–generator–battery) is nonlinear, multi-constraint, and affected by uncertainties in load and renewable energy forecasts. Conventional rule-based or deterministic optimization methods often require accurate models and are less robust to forecast errors, while large-dimensional exact solutions are not always feasible for real-time operations. This study proposes a Deep Reinforcement Learning (DRL)-based generator scheduling optimization framework that formulates the problem as a Markov Decision Process. The state vector includes multi-horizontal load/renewable energy forecasts, battery state of charge, fuel price, and unit operating limits; actions are the genset power setpoint and battery charge/access rate. A reward function internalizes fuel costs, battery degradation, emissions, curtailment, and unsupplied energy penalties, while also encouraging reserve provision. To ensure operational safety, we add a safety layer that projects policy actions onto the feasible set (SOC limits, ramp rate, minimum on/off, and converter capacity). Training is performed offline with domain randomization over weather and load profiles, and then evaluated in a rolling horizon scheme with minute resolution. Simulation results demonstrate operating cost savings and curtailment reduction compared to the MILP/MPC baseline, with high constraint compliance and sub-second inference times, making it suitable for implementation in edge controllers. This approach demonstrates scalability across a wide range of microgrid configurations and remains robust to uncertainties, offering a practical path to low-cost and low-emission operation.