Gradient Boosting Decision Tree and Random Forest Based Partial Discharge Pattern Recognition of HV Cable

As a typical “white box” pattern recognition method, the traditional Decision Tree (DT) method, which is easy to be interpreted and trained, has been widely applied in the Partial Discharge (PD) pattern recognition of high voltage cables. However, the disadvantages of the DT method are significant, including the limited recognition accuracy, poor anti-interference ability and poor generalization ability, etc. In order to overcome the challenges, Gradient Boosting Decision Tree (GBDT) and Random Forest (RF) based PD pattern recognition methods are presented in the paper. Firstly, 3500 PD samples are obtained based on the PD testing of five types of artificial defects of ethylene-propylene (EPR) cables in the high voltage lab. Secondly, PD data pre-processing is carried out and 34 types of PD features are extracted from the raw data. Thirdly, the principles of GBDT and RF based PD pattern recognition are presented in details. Finally, the 3500 PD samples are applied to evaluate the GBDT and RF based PD pattern recognition methods, which are compared with the traditional DT method. The results show that the PD pattern recognition accuracies of RF, GBDT and DT are 89.7%, 90.4% and 84.3% respectively. GBDT and RF based PD pattern recognition methods show remarkable advantages, compared with the DT method, which are both applicable for industrial application of PD based condition monitoring of high voltage cables.