Explainable reinforcement learning for distribution network reconfiguration

Abstract

The lack of transparency in reinforcement learning methods' decision-making process has resulted in a significant lack of trust towards these models, subsequently limiting their utilization in critical decision-making applications. The use of reinforcement learning in distribution network reconfiguration is an inherently sensitive application due to the need to change the states of the switches, which can significantly impact the lifespan of the switches. Consequently, executing this process requires meticulous and deliberate consideration. This study presents a new methodology to analyze and elucidate reinforcement learning-based decisions in distribution network reconfiguration. The proposed approach involves the training of an explainer neural network based on the decisions of the reinforcement learning agent. The explainer network receives as input the active and reactive power of the buses at each hour and outputs the line states determined by the agent. To delve deeper into the inner workings of the explainer network, attribution methods are employed. These techniques facilitate the examination of the intricate relationship between the inputs and outputs of the network, offering valuable insights into the agent's decision-making process. The efficacy of this novel approach is demonstrated through its application to both the 33- and 136-bus test systems, and the obtained results are presented.