การประยุกต์ใช้การเรียนรู้ของเครื่องสำหรับการบำรุงรักษาเชิงพยากรณ์ของหม้อแปลงไฟฟ้าระบบจำหน่าย

Abstract

The cellulose paper degradation is a criterion for the lifetime of the distribution transformer . During the entire operation time, the distribution transformer has to withstand numerous stresses. These stresses are high temperature, electromagnetic field and mechanical nature which can result in various problems, such as insulation degradation, hotspots, partial discharge etc. The remaining useful life (RUL) of cellulose paper can be caused by many reasons, but in the most cases it is because of the winding hotspot temperature (HST) related to electric current flow through the transformer winding which around the neighborhood of the cellulose paper and mineral oil. Both of them deteriorate under the HST or electrical stress of distribution transformer in service which are not directly monitored, and they are implied from other measurements. However, measurement errors influence prediction models and if undecided variables are not taken into account this can lead to erroneous maintenance decisions. In this thesis present the time series prediction using long short-term memory (LSTM) technique for predict the degradation of cellulose paper insulation in distribution transformers which approach combining uncertainty modeling, data driven predicting models. The proposed approach utilizes the weight and bias values to adjust the optimized network parameter through learning algorithms. The goal of the proposed framework is to predict the daily HST state given inspection data up to now and predict the likely future remaining lifetime given hypothetical future profiles. After that, to calculate the RUL of cellulose paper by the lifetime modeling which lead to condition assessment of distribution transformer.