The data-driven distribution network state estimation algorithm has shortcomings such as poor physical interpretability

low estimation accuracy under small samples

and poor convergence performance. Therefore

this paper proposes a distribution network state estimation method that integrates physical information and the GraphSAGE model. Firstly

in view of the characteristic that the line parameters of the distribution network play a decisive role in the voltage amplitude and phase

multi-dimensional edge feature information is integrated in the process of node feature aggregation to enhance the model’s ability to represent the characteristics of the power grid. Secondly

combined with the multi-head graph attention mechanism

the importance weights of different neighboring nodes are dynamically allocated

which can not only focus on more important nodes during node sampling to improve the training efficiency of the model

but also enable the model to rely on the limitations of the fixed adjacency matrix. Furthermore

physical constraints are embedded into the GraphSAGE model

and the model is forced to follow physical laws during the training process

which improves the interpretability and convergence speed of the model. Finally

simulation experiments were carried out in the IEEE 118-node distribution network system and compared with other neural network algorithms. The results show that this algorithm has better estimation effects in large-scale distribution networks and small samples.