Introduction: Dynamic network reconfiguration is a critical tool for enhancing the operational efficiency and security of power distribution networks. However, traditional model-based optimization methods are often computationally intensive and rely on precise system models, whereas conventional reinforcement learning (RL) approaches frequently overlook critical safety constraints. Methods: To address these challenges, this paper proposes a novel safe deep reinforcement learning (SDRL) framework, Prioritized Experience Replay Safe Soft Actor-Critic for Network Reconfiguration (PER-SSAC-NR). The proposed framework balances operational cost reduction and safety constraint adherence by combining a Safe Soft Actor-Critic (SSAC) architecture with parameter-sharing twin critics and prioritized experience replay. Results: Comprehensive experiments on the standard IEEE 33-bus and 69-bus test systems show that PER-SSAC-NR outperforms baseline RL methods in both learning speed and final policy quality. The agent achieves near-optimal power losses of 0.1410 MW and 0.0992 MW on the 33-bus and 69-bus systems, respectively, closely approaching the 0.1395 MW and 0.0985 MW solutions from the model-based optimization benchmark while satisfying all operational safety constraints. Discussion: These results demonstrate that the proposed PER-SSAC-NR framework can provide safe, efficient, and scalable data-driven decision-making for power distribution network reconfiguration.

The volatility of renewable energy generation and nodal electricity prices provides an arbitrage opportunity for Mobile Battery Energy Storage Systems (MBESS) leveraging both temporal and spatial advantages. However, the inherent high complexity and strong randomness of both power and transportation systems lead to complex risks for MBESS control. Existing works mainly consider the market price risk and ignore the transportation system risk caused by traffic congestion. Specifically, they are constrained by two critical limitations: (1) decisions can only be made upon arrival at a destination, making the agent unresponsive on the road, and (2) traffic congestion risk is neither quantified nor controlled, leading to suboptimal routing strategies. To address these limitations, the MBESS needs more flexible “on the road” decision making and multiple risk management capabilities. Guided by this objective, a flexible deep reinforcement learning-based MBESS control framework is proposed, considering both market and traffic congestion risk. First, dynamic routing ability is integrated with the MBESS agent to provide more flexibility in making decisions, regardless of whether the agent has reached the designated location or not. Second, two risk metrics are proposed to quantitatively assess the traffic congestion risk based on moving time, and then the agent can make decisions considering both market and traffic congestion risk. Finally, considering the inefficiency of learning caused by introducing multiple risks, a risk curriculum learning method is proposed to improve the training efficiency and reduce learning costs. These components are unified in the Multiple Risk Estimation SDDPG (MRE-SDDPG) algorithm, which jointly maximizes profitability while controlling electricity price and traffic congestion risk. Simulations in the IEEE 30 bus environment show that the proposed framework can increase profit by 8.6% while reducing the traffic time by 15.8% on average, demonstrating the superiority of our design in considering traffic congestion risk.