Partial discharge detection using convolutional neural network and k-nearest neighbor algorithm

A partial discharge (PD), according to IEC 60270, is a localized electrical discharge that only partially bridges the insulation between conductors. Medium voltage power lines transport electricity over long distances; as a result, officials have a difficult time monitoring and early detecting the PD of the cables in these areas. The challenge is to detect PDs early enough to prevent long-term damage; hence, early PD detection is essential to prevent costly repairs and substantial power outages. There are numerous approaches for measuring PDs online, but the most common method is done offline and includes an expert manually identifying acceptable features to classify PD type and severity, thus, it is essential to have a detection mechanism that can extract features and classify them automatically to predict PDs rapidly and accurately. The dataset was collected from kaggle.com. The Technical University of Ostrava (VSB) has designed a specialized meter to determine the voltage signal of stray electrical fields and insulated overhead cables. The fast Fourier transform technique was used to de-noise the signals as a preprocessing step before building the classification models. Four classification models were built and compared with each other; the models are CNN-KNN, CNN, CNN-LSTM, and KNN. The Adam optimization algorithm was used to update the weights to reduce the error between actual output and predicted output, and the sigmoid function was used in the output layer to keep the output within the range 0–1, and used dropout technique between layers to prevent overfitting during training of the model. Model performance was measured using several metrics: F-score, recall, precision, accuracy, and confusion matrix; the results show that the hybrid model (CNN-KNN) gives higher performance than the other models (CNN, CNN-LSTM, and KNN).