An approach based on machine learning and mel-frequency cepstral coefficients for locating faults in transmission lines

Faults in transmission lines may cause great loss to users and managers of electric power systems. Thus, it’s important making the process of locating these faults more efficient, in order to repair them as quickly as possible. In this study, mel-frequency cepstral coefficients were used for processing voltage signals collected on both transmission line terminals during faults, along with an machine learning (ML) model, responsible for locating faults. Different ML models were tested: artificial neural network (ANN), support vector machine and least squares support vector regression, among which was noticed that ANN had the best overall result, processing all simulations. A modeled line based on parameters of a real line was also used. The proposed method provided results with high precision in locating faults in environments without noise, with mean relative error of 0.00004% and mean absolute error of 0.13 ms. Subsequently, the influences of training dataset size, noise, fault’s types, fault’s resistances, fault’s angles and fault’s distances in the location method were evaluated through the results of the best ANN architecture. The proposed method was still able to detect the faults quickly and precisely, even with small size of the data set and/or different signal to noise ratio. These results indicate that the proposed procedure is a good alternative for fault location in LT in practical scenarios.