Analysis Method of Dielectric Strength in Electric Power Apparatuses based on Electrostatic Features using Machine Learning

Abstract

Power equipment is becoming increasingly compact. As the spatial distance between current-carrying conductors decreases, the risk of electrical breakdown (BD) in power devices including energy storage systems or gas insulated medium-voltage switchgears increases. Therefore, the prediction of BD voltages is essential. Many studies have evaluated BD voltages based on various gas present and the geometrical shapes of electrodes. According to prior studies, Paschen’ law can evaluate the dielectric strength of insulation mediums themselves, and streamer (inception and propagation) criterion can predict the dielectric strength in non-uniform electric field distribution. However, these methods are only useful under certain conditions, because the breakdown are influenced by environmental factors, impurities, electrode shapes and discharge processes. Thus, machine learning is utilized to find out relation between these multiple parameters and BD voltages. In prior studies on machine learning-predicted voltages, known design variables are mainly used, and too many input features are applied to enhance the prediction accuracy, so optimization methods are employed. However, the optimization method is not useful in case that such features do not reflect the physical properties of studied electrodes. Also, deep learning can only evaluate some of problems that can emerge from among known features affecting the phenomenon of interest. Therefore, feature design is very important. For this, features were suggested based on background electric fields and electrical properties of discharge channels to reflect electrical characteristics of the studied electrodes. In this study, sphere-to-plane, or sphere-to-barrier-to-plane electrodes were utilized as test electrodes to describe current-carrying conductors of the power equipment. Especially, breakdown voltages of sphere-to-barrier-to-plane scenario are rarely predicted via machine learning. Thus, breakdown voltages of such studied electrodes were predicted by support vector regression (SVR), bayesian regression (BR), multilayer perceptron (MLP). The predicted breakdown voltages were compared with voltages derived using both the streamer propagation criterion and experiment. Through this study, the proposed method can provide more exact breakdown voltages and physically explain the important electrical properties affecting discharge under such conditions.