Characterization and determination multiple partial discharge sources location based on neural network in medium voltage power cable

Partial discharge (PD) in a medium voltage distribution line (MVDL) is one of the issues in the power cable that should be prevented earlier to avoid failure in the electrical system. The sources of discharge activities may come from cable fabrication defects such as voids, corona discharges, surface discharges, and many more. The current issue of PD signals does not usually occur as a single source. The PD source that occurred on the MVDL can also present multiple signal sources. Poor workmanship during cable jointing, aging, or exposure to the surrounding environment is the most common cause of PD in HV cable systems. As a result, the location of the PD signals that occur cannot be classified without identifying the multiple PD signals present in the cable system. Thus, knowing the PD location is necessary to identify the cause of the PD, determine its type, and know the severity of the fault. Without detection and PD localization in the MV power cable, PD frequency and occurrence increase over time, leading to the failure of the MVDL, causing equipment damage and loss of revenue because of an unscheduled outage. The methodology of localization technique in multiple PD signals captured by Rogowski coil (RC) sensors at different locations of the cable has been presented in this thesis. Multiple PD signals in an 11 kV underground power cable have been modeled using Electromagnetic Transient Program-Alternative Transient Program (EMTP-ATP) software based on the location technique algorithm. In order to know the PD types of each location in MVDL, neural networks (NN) have been applied with additional statistical features to reduce the large data. The lowest error means the great performance of NN in classifying the PD types. Once the PD types have been revealed, the location of PD types can be known by using a three-point technique that calculates the location of the existence of multiple PD signals. This thesis found that the original signal for different locations can be identified by using NN, with an accuracy of classification is 96.3% and a percentage error for localization is 0.011%. This finding shows that the combination of NN and statistical features has been successfully implemented to classify different types of PD in an MVDL at once, giving high precision of the three-point technique to determine the location of the PD occurrence.