Optimization and Scheduling of Mobile Energy Storagefor Voltage Reactive Power Control in DistributionNetworks Based on Deep Reinforcement Learning

Introduction: This study addresses the challenge of enhancing grid resilience through the optimal coordination of Mobile Energy Storage Systems (MESS) under extreme events and high renewable energy penetration. It proposes a Parameterized Deep Q-Network (P-DQN) framework to unify long-term planning, real-time dispatch, and decentralized control, thereby bridging critical gaps in existing methodologies. Methods: A hybrid framework that combines genetic algorithms for long-term capacity planning with consensus-based decentralized control for voltage-constrained power flow management is developed. The P-DQN algorithm jointly optimizes MESS dispatch and grid operation under renewable energy uncertainty, incorporating hybrid action spaces and embedded system dynamics. Results: The proposed method is evaluated on an improved IEEE 33-node system with PV installations, MESS charging stations, and real-world data sources, employing a P-DQN framework featuring three-hidden-layer networks and ε-greedy exploration. Training converges by ~4,000 episodes with residual reward fluctuations caused by persistent exploration (ε ≥ 0.01) and environmental variability. Penalty terms for constraint violations remain nonzero, indicating that occasional feasibility trade-offs are made to optimize multi-objective performance. Discussion: The P-DQN framework advances existing methods by integrating adaptive learning, hybrid action spaces, and multi-objective optimization. Its robustness under partial observability and scalability in complex grid networks highlight its practical relevance for modern power systems. Conclusion: The proposed framework provides a versatile solution for grid resilience, directly addressing the limitations of centralized models and model-based approaches through decentralized, data-driven optimization.