The construction of advanced metering infrastructure and the rapid development of deep learning technology make it possible to quickly find the optimal strategy for reactive power optimization by mining historical data and prior knowledge instead of relying on physical models. Therefore, a method for reactive power optimization based on graph convolutional network (GCN) is proposed. Through representing the topology information between nodes in distribution network with the adjacency matrix, the proposed algorithm can effectively mine the correlation between the node loads, mapping the complex nonlinear relationship between the power equipment status and the load data with the deep graph convolutional architecture. The simulation results show that the accuracy and robustness of the GCN are better than that of the existing data-driven methods such as the convolutional neural network, the multi-layer perceptron and the case-based reasoning. Its solution time is much shorter than the traditional heuristic algorithm, which can meet the real-time demand of the reactive power optimization in distribution networks.