The randomness and fluctuation of renewable energy output exacerbate voltage fluctuations in the power grid

significantly impacting the safety and economic operation of the system. To address this issue

a real-time reactive power optimization method based on a graph attention network (GAT) is proposed. First

the labels for the training samples are designed based on the physical characteristics of the power grid. Second

the GAT model is used to map the complex nonlinear relationships between the states of reactive power regulation devices and renewable energy and load data

thereby obtaining the optimal solution. Finally

the decision-making results of the GAT model are analyzed using the Shapley additive explanations (SHAP) interpretability mechanism. An application to an actual power grid demonstrates that the proposed method exhibits superior reactive power optimization performance and faster solution speed

while the interpretability approach aids operators in better understanding and trusting the model's decisions

thereby enhancing the safety and economic level of the new power system's operation.