Depolarization Current Prediction of Transformers OPI System Affected From Detrapped Charge Using LSTM

The depolarization current is critical for the computation of insulation-sensitive parameters of the transformer. Prolonged polarization–depolarization current (PDC) measurement times, on the other hand, can increase the likelihood of obtaining distorted original data as a result of environmental factors and electromagnetic noise. Therefore, techniques for measuring PDC data over a short period of time, such as forecasting depolarization current, are required. In addition, unlike polarization current, which is easily predictable, depolarization current is affected by charging duration and detrapping current. Consequently, an extremely accurate prediction method is required. This article proposes a multivariate recurrent neural network (NN) model based on the long short-term memory (LSTM) network for depolarization current prediction using only polarization data. It reduces measurement time and treats the model with time-varying parameters (MTVP) elements as features. Therefore, the influence of the detrapped charge effect while forecasting depolarization current is also considered. The performance of LSTM is compared to that of linear, multidense, and 1-D convolutional NN (1-D-Conv) to determine the optimum model for forecasting depolarization current. Thereafter, a permutation-based explainability method is used in Python to describe model behavior by assessing the variable importance through dropout loss. The presented results demonstrate that forecasting with the LSTM model produces the lowest mean absolute error and maintains prediction consistency throughout the testing period.