With the advancement of the carbon peak and neutrality targets

the integration of large amounts of highly uncertain renewable energy sources into the grid has introduced the uncertainty of the power system and severely altered the power flow and voltage distribution in the distribution network

which poses new challenges for the operation and management of the distribution network. Artificial intelligence technology has experienced very rapid development in recent years. Graph reinforcement learning algorithms (GRL) effectively combine the feature extraction capabilities of graph neural network (GNN) for non-Euclidean structured data with the sequential decision-making capabilities of deep reinforcement learning (DRL). This enables GRL to possess a strong nonlinear fitting ability

excellent adaptability

outstanding scalability

and significant advantages in dynamic optimization decision- making. As a result

GRL is highly suitable for addressing complex uncertainty optimization and decision- making problems in distribution network with graph-based models. The principles and characteristics of GRL are introduced first by combining the principles of reinforcement learning with commonly used types of GNNs. Then

the applications of GRL in distribution network are analyzed in four aspects: optimal dispatch

dynamic reconfiguration

fault recovery

and collaborative optimization of coupled power- transportation networks. Then a comparison is made between the application status and performance of GRL and DRL

analyzing the advantages of GRL in distribution network optimization as well as the existing issues. Finally

the issues that need improvement in the application of GRL for power distribution network optimization are summarized

and the future prospect of its application is discussed.