Reinforcement-learning-based adaptive energy management strategy for microgrid energy storage systems

Abstract

High renewable penetration makes microgrid energy management sensitive to uncertain photovoltaic output, wind fluctuation, load variation, electricity price, and battery degradation. Conventional rule-based and model predictive strategies require manually tuned thresholds or accurate forecasts, which limits their adaptability when the operating condition changes rapidly. This paper proposes a reinforcement-learning-based adaptive energy management strategy for a grid-connected microgrid with a battery energy storage system (BESS). The method formulates the dispatch process as a constrained Markov decision process, constructs a state representation combining renewable generation, demand, price, state of charge, and forecast residuals, and trains an adaptive soft actor-critic controller with safety projection and dynamic reward weighting. The controller learns charging, discharging, and grid-exchange decisions while respecting power balance, state of charge (SOC) limits, and battery cycling constraints. A 15-min simulation study of a campus/charging-facility microgrid shows that the proposed strategy reduces operating cost by 22.7% compared with a rule-based benchmark and by 4.3% compared with a PPO controller, while improving renewable self-consumption and keeping SOC violations below 0.5%. The results demonstrate that reinforcement learning can provide an adaptive and computation-light dispatch layer for resilient microgrid storage operation.