Fault Diagnosis System Using CNN-LSTM Networks

Abstract

Prompt classification and detection of faults along transmission lines is critical to the smooth operation of energy companies, the maintenance of their power systems, and efficient transmission. Traditional methods are characteristically limited in their ability to handle large volumes of data thus, the need for intelligent networks such as neural networks. This research proposes the use of a hybrid of two of such networks - the Convolutional Neural Network (CNN), and the Long Short-Term Memory (LSTM), to improve the fault diagnostics capabilities for a three-phase transmission line. Fault data for twelve possible scenarios are obtained from a SIMULINK model. The CNN-LSTM model is trained and tested using this dataset in 80-20 training-testing split. The features are four current values - three phases’ currents and ground current with targets labeled as 0 (no fault) and 1 (faulty). The targets enable the model to classify faults into the twelve already defined labels. The CNN-LSTM model was trained using normalized values to prevent overfitting. The CNN-LSTM model which is robust and good for time-series prediction and adapting to changing load patterns. The model's performance was evaluated using confusion matrix, accuracy, precision, recall, and F1-score. The test accuracy was 94.56%, precision 95.89%, recall 95.40%, F1 score 94.88%, and the confusion matrix showing 13 faults were misclassified out of 239, representing approximately 5%. The CNN-LSTM model is saved for real-time fault diagnosis.