Optimal Control of Battery System by Reinforcement Learning Considering Profitability

Government-led pilot experiments for Virtual Power Plants (VPPs) are being conducted in Japan. However, these experiments have faced challenges such as deviations between predicted and actual values of photovoltaic (PV) generation, as well as insufficient battery capacity, leading to imbalances and associated penalties that reduce the revenue of VPP aggregators. In this study, we propose a method that combines probabilistic PV forecasting and deep reinforcement learning to determine the charging and discharging schedules of batteries within the VPP. Our approach considers the risks associated with deviations between forecasted and actual values, enabling the generation of operational plans while considering these uncertainties.