Advancements in Recognizing Partial Discharge Root Causes through AI

Partial Discharge (PD) measurements have emerged as a crucial tool for evaluating the integrity of insulation in electrical systems. While PD levels traditionally guide the commissioning of new equipment, identifying the underlying root cause of PD and estimating potential risks remain challenging tasks that require the expertise of skilled professionals. Over the past 25 years, rudimentary artificial intelligence algorithms have been employed to assist operators in this endeavor, albeit with limited success, achieving a recognition rate of around 80-85%. However, recent advancements in machine learning and Artificial Intelligence (AI) have significantly enhanced these capabilities. This paper utilizes the latest AI toolboxes to improve recognition performance. The author investigated the results of 118 different AI algorithms using 3515 industrial GIS PD measurements and compared them with older algorithms. Among the various algorithms tested, the Light Gradient Boosting Machine algorithm (Light GBM) exhibited the highest performance, achieving a recognition rate of 97.3% on untrained datasets. Further optimization of the algorithm was conducted, followed by a discussion on the effectiveness of the 134 recognition features utilized and which ones can be omitted.