A Novel Approach for Fault Location in Distribution Systems Based on Bayesian Optimised CNN-SVM

This work proposes a novel, optimized hybrid fault location finding technique in power distribution systems with radial and ring topologies. The transient current signals recorded at each node of the IEEE 13-node test feeder are processed using various convolutional neural network (CNN) designs, including AlexNet, GoogleNet, SqueezeNet, and ResNet. Support vector machines (SVMs) (which identify the nodes and determine where the fault occurs in the IEEE 13-node test feeder) are trained using these features extracted by CNN architectures. Two SVM hyperparameters—the box constraint and kernel scale—are adjusted using Bayesian optimization (BO) to increase the accuracy of fault location. The success of the suggested method is evaluated using many characteristics, including fault type, resistance, inception angle, and nodes. An analysis is done on how four different Kernel functions—polynomial, RBF, Gaussian, and linear—affect the performance of the SVM method. In terms of fault location accuracy, this model's performance is evaluated. By obtaining 100% fault location accuracy for radial topologies and 97.35% for ring topologies of power distribution systems, the simulation results show the utility of the proposed model.