Uncertainty-Aware Safe Soft Actor-Critic for Coordinated Battery and Electric Vehicle Energy Management in Active Distribution Networks

The increasing penetration of photovoltaic generation, battery energy storage systems, and electric vehicles is transforming the operation of active distribution networks. Although these resources provide considerable operational flexibility, their stochastic behavior introduces substantial challenges for real-time energy management. In particular, photovoltaic uncertainty, time-varying electricity demand, and uncertain electric-vehicle availability can lead to a significant mismatch between the conditions used for controller training and those encountered during operation. This paper proposes an uncertainty-aware safe Soft Actor-Critic framework for coordinated battery energy storage and electric-vehicle energy management in active distribution networks. The proposed framework combines continuous-action reinforcement learning with an explicit safety layer that filters infeasible actions before their execution. To improve generalization, domain randomization is employed during training by varying load demand, photovoltaic generation, electric-vehicle availability, and initial state-of-charge conditions. The resulting controller is evaluated in a physics-based IEEE ۳۳-bus distribution-network environment containing distributed photovoltaic generation, battery energy storage, and electric-vehicle charging resources. Four control strategies are considered: rule-based control, conventional Soft Actor-Critic, safe Soft Actor-Critic, and the proposed uncertainty-aware safe Soft Actor-Critic. Performance is evaluated using operating cost, peak grid demand, voltage violations, branch loading, electric-vehicle charging satisfaction, battery utilization, and degradation under unseen operating conditions. The experimental design explicitly separates the effects of safety enforcement and uncertainty-aware training. Consequently, the study does not merely evaluate the performance of a reinforcement-learning controller under nominal conditions, but investigates its ability to maintain safe and economically efficient operation when the operating distribution changes.

کلیدواژه ها:

Active distribution network ، battery energy storage ، electric vehicles ، safe reinforcement learning ، Soft Actor-Critic ، uncertainty-aware learning ، energy management ، photovoltaic generation ، smart grid