Long Short-Term Memory (LSTM) Network for Classification of Partial Discharges

Partial discharges (PD) are critical indicators of insulation degradation and potential faults in electrical equipment. The timely detection and classification of PDs lead to reliable system operation. In this study, we explore the application of Long Short-Term Memory (LSTM) networks to classify partial discharge (PD) types. Two different types of data inputs are used to train two LSTM models. The two inputs to the LSTM models are derived by subjecting the partial discharge data to discrete wavelet transform (DWT) and wavelet packet transform (WPT). The DWT and WPT are both used to decompose the PD signals using low-pass and high-pass filters, which yield a set of approximation coefficients representing the lower frequency―general trends in the signal―and detail coefficients that capture the higher frequency―fast-changing features in the PD signals. Statistical metrics such as mean, variance, standard deviation, skewness, kurtosis, rms, and entropy are computed from these coefficients. For each transform, the computed statistical features are converted into a vector, which is used as inputs to train each model. The results indicate slightly higher PD classification accuracy in the LSTM model trained with statistical features from WPT (91%) than that trained with features extracted from DWT (89%), even though it takes longer to train.