A Load Transfer Decision Method for Distribution Networks Based on Mask-Constrained Deep Reinforcement Learning

In recent years, with the rapid expansion of urban areas and the increase in electricity consumption, the number of nodes in distribution networks has significantly grown, resulting in a more complex network structure. Moreover, after fault isolation in distribution networks, there is considerable uncertainty in the topological changes, and the selection of load transfer paths faces the challenge of combinatorial explosion due to numerous switch combinations. Traditional load transfer algorithms may not be suitable for the rapid load transfer demands following the occurrence of faults. Efforts have been made to introduce deep reinforcement learning algorithms to address these challenges. However, pure reinforcement learning methods face challenges of low policy safety and reliability as well as inefficient convergence. To improve the safety of actions, the paper makes use of human empirical knowledge to establish rules and constraints when designing reward functions. To improve the algorithm’s convergence efficiency, an action masking strategy based on an enhanced mask is introduced, allowing the direct masking of ineffective actions during the learning process. Taking the IEEE 123-bus system as a simulation example, the results demonstrate the superiority of maskDDQN in fault recovery improvement.