MSGCN: Multi-task Spectral Graph Convolutional model for identification of branch parameters considered grid topology

Abstract

This paper addresses issues associated with existing methods for power system branch parameter identification, such as the inability to consider transmission line topology, handling large-scale power grid data, and high sensitivity to data contamination. A method called Multi-Task Spectral Graph Convolutional Network (MSGCN) is proposed for power system branch parameter identification tasks. This method utilizes a novel aggregation method called Simplified Spectral Graph Convolution (SSGConv) to simplify graph convolution operations. It introduces graph adaptive normalization and a learnable skip-connection mechanism to enhance the model’s robustness and scalability. Additionally, the method incorporates a graph attention mechanism, enabling our model to automatically learn the power grid branch topology, reducing the influence of data contamination on branch parameter identification accuracy. It adopts a self-balancing loss function of the multi-task model based on homoscedastic uncertainty to simultaneously identify multiple branch parameters, improving both the accuracy and training speed of the model. Experimental results demonstrate that this method outperforms traditional approaches and other graph neural network methods in terms of efficiency and accuracy in branch parameter identification.