A multi-hierarchical interpretable method for DRL-based dispatching control in power systems

Abstract

Timely, effective, and robust artificial intelligence (AI) technology is urgently needed to improve decision-making efficiency in the presence of renewable energy with high penetration by elevating the level of power grid intelligence. However, at this stage, AI technology lacks reliability and transparency, making it unable to play a full role in application areas with high-security requirements such as power systems. This paper presents a multi-hierarchical interpretable method for power system dispatch and operation based on the graph deep Q-network (GDQN) model to achieve active power corrective control. The multi-hierarchical interpretable method combined with an improved sample-balanced deep shapley additive explanation (SE-DSHAP) method and a subgraph explainer can promote a more intuitive and comprehensive explanation of decision-making for power systems with complex topology. Operators will obtain more noteworthy power grid areas through the proposed interpretable method as the basis of auxiliary decision-making to realize efficient and accurate control. Then, two cases studied in a modified 36-bus system in the IEEE 118-bus system and the 300-bus network of the China power grid validate the effectiveness of the proposed method.