Deep Reinforcement Learning-Based Mobile Battery Energy Storage System Control With Partial Observability and Data Imputation

Mobile battery energy storage systems (MBESS) play a crucial role in balancing loads, supporting energy integration, and providing backup power. However, accurate and reliable MBESS control is significantly challenged by the synergistic issues of partial state observability and data incompleteness. To address these issues, we propose a novel deep reinforcement learning MBESS control framework based on the partially observable Markov decision process (POMDP) and data imputation. The problem is formulated as the POMDP and the partially observable safe twin delayed deep deterministic policy gradient algorithm is proposed to solve the problem. In addition, to improve the control performance, a Transformer-based neural network data imputation method is proposed to reconstruct missing information. Furthermore, we introduce a priority experience replay mechanism that leverages Bayesian uncertainty estimation to improve learning efficiency. Experiment results in IEEE 30-bus and 57-bus systems present the robustness and adaptability of the proposed MBESS control framework under partially observable conditions.