Partial Discharge Detection in Power Lines Using Automated Machine Learning

Electrical grids must be reliable, and maintenance plays a vital role in this matter. A common application of artificial intelligence in the industrial sector is predictive maintenance, alongside condition monitoring. In power systems, this means that maintenance can be scheduled for the right time so electrical faults can be avoided before they occur. In this way, to evaluate the health of insulation in power systems, partial discharges (PD) are used as indicators of insulation degradation in electrical equipment. Detecting a PD at an early stage can prevent catastrophic consequences. Machine learning (ML) and deep learning (DL) techniques have been used in PD diagnostics for detection, localization, and recognition. Domain knowledge on the matter improves feature engineering, but manually creating pipelines for ML models is a tedious task. DL is being used to not depend on human expertise, but these models are hard to configure. On the other hand, automated machine learning (AutoML) allows domain experts to implement robust ML models without having a background on statistics and ML. However, AutoML has not been widely addressed in the scope of PD diagnostics. For this reason, in this paper AutoML is used for the detection of PD faults. A public dataset of PD measurements on covered conductors from power lines is used to implement models using AutoML. The data was preprocessed, and several models were developed using various resampling techniques. An AutoML model with no resampling technique consisting of multiple algorithms obtained the best results, outperforming a LightGBM model.