Battery Scheduling Control of a Microgrid Trading with Utility Grid Using Deep Reinforcement Learning

Abstract

Managing microgrids (MGs) with variable renewable energy (VRE) is challenging because of uncertainties of electricity production, loads, and energy price, so we need flexible control strategies for battery energy storage system (BESS) to handle those challenges. Model-based approaches require precise models of the MG to give accurate results but having an accurate model can be difficult in continually changing environments. We introduced a new day-ahead optimization method to control BESS scheduling and power exchange between the utility grid and an MG with a load, photovoltaics, and BESS with the aim of energy cost minimization. Deep reinforcement learning (DRL) was used for the optimization of sequential actions of BESS over a time horizon. A theoretical optimum scheduling was derived using a linear programming optimization to be compared with the DRL agent. Both no-battery and greedy control algorithms were used as baselines. It was shown that the results of the proposed technique were better than the baselines through numerical simulations using whole year's data.