Fast Pooling Based Multi-Scale Graph Neural Network and Its Application in Electric Power System

With the development of power system intelligence, data in the electric power system has shown high complexity and uncertainty. For this reason, the knowledge graph (KG) has been applied in various applications in the power grid. By combining with machine learning, deep learning, and other graph computing technologies, the knowledge graph can effectively find out the useful information in the data, and use the existing data information to optimize the business process and help in decision-making. However, the knowledge graph in the electric power system has multiple nodes, heterogeneous edges, and complex structures, and the applications on the knowledge graph are shallow. The traditional graph neural network models do not consider much on the topological structure of input graphs, and thus the problems of low efficiency and high complexity come out. In order to solve the problems mentioned above, we proposed a multi-scale graph neural network with fast pooling in this paper. This model could improve the computing efficiency through the fast calculation of the pooling layer, and prevent the decrease of the test accuracy by introducing a multi-scale network structure. Also, the model could adapt to different sizes of input graphs, where the pooling layer gave a fixed size of output regardless of the input size. According to the experimental results on the popular benchmark dataset PROTEINS and the self-built dataset CEPRI-Transmission, the proposed model outperforms baseline models when considering both accuracy and efficiency.