Impulse Discharge Voltage Prediction of Complicated Engineering Gaps Based on Machine Learning of Spatial Electric Field Features

Discharge voltage prediction of engineering gaps with complicated arrangements is a long-term goal in high voltage studies. This paper proposes a spatial electric field feature set for transmission line - tower gaps. A strong correlated electric field zone was defined between the energized bundled conductor and the grounded tower body or the cross arm. This zone is a conical region with an assumptive cone angle and a circular arc boundary delimited by the equipotential surface. After electric field simulation, 13 features were extracted from this conical zone, and 15 features were extracted from the shortest path. These spatial electric field features were input to a machine learning model based on support vector machine (SVM). Trained by the features and the experimental data of 15 gaps used in 500 kV, ±660 kV, and 750 kV transmission lines, the SVM model was used to predict the 50% discharge voltages of ±800 kV and 1000 kV engineering gaps. Six prediction cases were carried out with two cone angles and three equipotential surfaces used in definitions of the conical feature extraction zone, and the mean absolute percentage errors (MAPE) are within 10%. When the cone angle is 90° and the equipotential surface is 30 percent of the applied potential, the MAPE is only 4.85%, and the relative errors are within 7.02%. In addition, the predicted
$U_{50}-d$
curves have similar variation trends compared to the experimental curves, which demonstrates the feasibility of the proposed method.