Abstract

This research paper explores the application of offline reinforcement learning (RL) in controlling battery operation in a grid-connected microgrid. The study investigates the impact of different parameters on the performance of the RL algorithm, such as the number of discretization levels, gamma, and alpha values. The results show that the convergence time and optimality of the RL algorithm are affected by the choice of these parameters. The research concludes that carefully selecting the discretization levels of state-action spaces and RL hyperparameters is crucial for optimal RL algorithm performance. The benchmark offline sensitivity analysis can be compared in the future with other RL approaches, such as function approximation or DRL methods.