Abstract: The randomness and volatility of power grid operation are increasing, which brings great challenges to the state control and analysis of system. Therefore, by deeply exploring the spatial-temporal characteristics in power big data, a graph neural network distribution grid state estimation method based on spatial-temporal correlation characteristics is proposed. The method transformed the distribution network topology into a graph structure in the graph domain and used graph convolutional neural network (GCN) to extract high-dimensional spatial features. The temporal characteristics of the electrical information timing series were extracted by gated recurrent unit (GRU). The IEEE39 model was improved for case analysis, which verified that the method had good estimation accuracy, tolerance and generalization ability.