The ordinary neural network model is susceptible to the interference of system topology changes and data missing in state estimation, which affects the accuracy of integrated energy state estimation. In order to solve the above problems, a graph neural network ( GNN ) model that embeds a gated recurrent unit ( GRU ) into a graph neural network that can deeply extract neighborhood features is proposed. The proposed model improves the neighborhood matrix of the graph neural network to become an adaptive k-order neighborhood matrix to improve the ability of neighborhood feature information extraction. Then, by embedding GRU into the graph neural network to consider the fusion of time and topological correlation of feature learning, the model can capture more fine-grained network topology changes. So as to improve the anti-interference ability of the model. The simulation results show that the proposed model has better estimation accuracy and stronger robustness than other dynamic models.