Research on the Construction of Fault Prediction Model for Flexible DC Converter Valve Submodule Network Based on Deep Learning

Abstract

Along with the wide use of DC power network in the modern electric power system, the fault of the sub-module network of converter valve has become a key factor which influences the stability of the power system. For the purpose of forecasting these failures and enhancing the stability of the power system, a new model of the deep learning fault forecast is presented in this paper, which is based on the Transformer and GCN. This model combines the global time series modeling capability of Transformer with the topological relationship modeling capability of GCN, and can simultaneously process the time series data of converter valve submodule and the spatial topological structure information between modules, thereby enhancing the accuracy of fault prediction. Experimental results show that this model has significant advantages over traditional machine learning methods (SVM, RF) and other deep learning models (LSTM, CNN + LSTM). The Transformer + GCN model achieved 96.2%, 94.7% and 95.4% in accuracy, recall and F1 score respectively, which greatly surpassed the performance of traditional models. It is proved that the combination of time series and topology characteristics of the deep learning model can increase the precision and robustness of the fault forecast.