AI-driven energy storage optimization in microgrids: Integrating deep reinforcement learning with BiLSTM and LightGBM models

The integration of intermittent renewable energy sources and dynamic pricing mechanisms into modern microgrids necessitates intelligent energy management systems (EMS) to ensure economic viability and grid stability. This paper proposes a novel framework that synergizes advanced forecasting models with Deep Reinforcement Learning (DRL) to optimize the operation of an Energy Storage System (ESS). We develop high-fidelity prediction models using LightGBM for solar and wind power forecasting and a Bidirectional Long Short-Term Memory (BiLSTM) network for load demand and electricity price forecasting. These predictions inform two distinct DRL environments: a 7-action discrete model and a more complex continuous-action model that holistically manages the entire microgrid. Four DRL agents-Deep Q Network(DQN), Double DQN (DDQN), Proximal Policy Optimization (PPO), and Soft Actor-Critic (SAC)-are trained and evaluated. The results demonstrate the framework's efficacy, with all agents learning profitable control policies. The SAC agent consistently yielded superior performance, achieving cost savings as high as 53.89% in the discrete environment with Austrian data and 54.91% with Finnish data. In the more complex continuous-action environment, the SAC agent also demonstrated the most robust and profitable strategy, validating the proposed system's potential for significantly enhancing microgrid operational efficiency and profitability.