Deep Learning Tools Analysis for Automatic Partial Discharge Detection Based on PRPD Patterns

Smart grid monitoring is growing with large scale deployments including technologies such as partial discharge sensors for assessing insulation condition. This paper describes the application of permanent monitoring solutions supplemented with artificial intelligence systems to generate reliable alerts for insulation defects based on partial discharge pattern recognition. Three different data models, initially with a high level of accuracy, have been implemented and compared using several types of architecture and training data sets. A characterization of the three models has been designed to show the performance in real conditions where data is: more complex than the training data set, noise is mixed with the defect, sensitivity is lost, and clustering techniques were needed to separate multiple defects. The analysis of the results is performed using a criticality matrix that helps assess which data models accurately identify which failures are true.