Neural Network Based Fault Location in Power Distribution System.

Locating the fault in the power distribution system is a tedious process that involves manually searching along the line. While some methods rely on complete system parameters or centralized data processing, it has become difficult due to the bidirectional power flow in the active distribution system. This paper proposes a framework consisting of three steps: detecting, classifying, and locating the fault in the active distribution system. The proposed method reduces the search area by using the sending-end sample value of the line current. It can be implemented in a distributed form without knowledge of system parameters using pole-mounted data recording meters and relaying the approximate fault location and type to the control center. In the first step, the fault is detected using an isolation forest. Then, the fault is classified by the support vector machine, and finally, the fault is located using an adaptive weight convolutional neural network (CNN). The CNN weights are modified according to fault type information that uses time-frequency information extracted by a continuous wavelet transform (CWT). The proposed framework is tested using an IEEE 13-node test feeder with a solar photovoltaic system. Different training parameters for the CNN are also tested to analyze the proposed framework