Development of Deep Learning Model Based Resnet with SSD for Fault Detection in Distribution Network

Abstract

This research proposes the development of deep learning model based residual network with single short multi box detector (resnet-SSD) for fault detection in distribution network. Power infrastructures are components used in the power system to generate, transmit or distribute power for the benefit of the consumers. These components include conductors, insulators, transformer, cross arm and many more. However, disturbances such as symmetrical fault, asymmetrical fault, shattered disk, broken cross arm, and power line encroachment affects the constant flow of power in distribution network. Failure to restore the system in good time may lead to vandalization, loss of revenue, black out and even system collapse. Various method of inspection has been used to address these problems which include manual inspection which is tedious and inefficient, use of heuristic and meta-heuristic algorithm which are in accurate and less efficient and the used of machine learning that suffers the problem of computational complexity and convergence time. To address these challenges associated with the existing solutions, a deep learning model based resnet-SSD is proposed which has the ability of addressing overfitting, vanishing and exploding gradient thereby overcoming the issue of computational complexity and convergence time. The developed model resnet-SSD was compared with mobilenet-SSD using accuracy, precision, recall and F1-score as performance metric. The result recorded shows that resnet-SSD has outperformed mobilenet-SSD in all the aforementioned performance metric by recording accuracy of 97.7% as against 86.3% for mobilenet-SSD and F1-score of 96.5% as compared to mobilenet-SSD that recorded 87.6%, from the result presented it can be inferred that resnet-SSD has better prediction efficiency as compared to mobilenet-SSD this is due to its residual network layer that help in efficient error back propagation of the deep learning model