Power System Fault Diagnosis Based on Logistic Regression Deep Neural Network

Though already being one of the standard modules of SCADA for years, the fault diagnosis software still exists the problem of high misrecognition rate. To mitigate the issue, this paper proposed a new fault diagnosis model and algorithm based on logistic regression deep neural network (LRDNN) to improve the diagnosis accuracy. The basic process of the proposed method is as follows: 1) construct a deep fully connected neural network for each power equipment. The input of the neural network is the vector of the fault signals of equipment, including switching signal, protection signal and etc., and the output is the failure probability; 2) train the LRDNN based on RMSprop method. Since the proposed method can dig out the complicated relationship between the fault signals and the fault probability, it owns high diagnosis accuracy. Moreover, this paper designed a sample expansion method based on random sampling strategy to successfully realize the training of the LRDNN, skillfully overcoming the defect of the lack of sample. The basic process of which is: 1) extend the historical fault records of equipment by including the fault records of equipment at nearby areas, whose type is the same as the one to be diagnosed; 2) generate adequate number of samples by probability statistics and random sampling operation. A comprehensive comparison of the proposed method and the method based on expert system was performed on practical power systems, which shows that the proposed method reduces the misrecognition rate from 31.25% to 0 compared with the one based on expert system.