Machine Learning Methods Applied on Long Term Data Analysis for Rain Detection in a Partial Discharge Sensor Network

Abstract — Partial discharges (PD) on the surface of high voltage insulators are directly related with the accumulation of pollution. A complete partial discharge sensor network was previously developed and has been in operation for approximately three years. This system records the PD activity, classifying it into four levels. As the PD activity is influenced by the weather conditions, each sensor system in the network also measures the one-hour average temperature and relative humidity. Also a fuzzy inference system was developed to extract the flashover occurrence risk level based on the partial discharge activity recorded. However, a strong rain event can wash insulators almost instantaneously, in turn decreasing the risk level. For a correct interpretation of the results it is important to properly analyze the weather data to detect the rain occurrence. This paper presents a comparison among three machine learning techniques for rain detection from humidity and temperature data, namely, Naïve Baye