Fault Classification and Section Identification in Distribution Networks Using Convolutional Neural Networks
Research Poster Engineering 2025 Graduate ExhibitionPresentation by Mohammadreza Mirjafari
Exhibition Number 167
Abstract
Power outage is one of the inevitable events in the distribution networks. With the increasing time of blackouts, costs and dissatisfaction will increase exponentially, and this is the reason that electric utilities are exploring new approaches to reduce the outage time. A critical aspect of outage management is fault classification and location, which has become more complicated with the increasing number of distributed generation (DG) units. Since the traditional methods are inadequate for handling fault classification and location in this modern structure, this study suggests using a Convolution Neural Networks (CNN) model. An effective data augmentation approach is used in this study to address the scarcity of fault data. To generate a fault dataset considering uncertainties in load and fault resistance, we use a co-simulation of MATLAB-OpenDSS to simulate the IEEE 13-node test feeder. The results have demonstrated the effective performance of the proposed model with a classification accuracy of 97.3% and localization accuracy of 94.4%, outperforming the Artificial Neural Network (ANN) model. We also evaluated the model's performance under varying levels of PV penetration, and the results demonstrate its robustness across different scenarios.
Importance
Distribution systems are crucial for delivering reliable electrical power to consumers, but they are susceptible to faults. These faults can be caused by equipment failures, weather, vegetation or animals, which can lead to power outages in part of the system. These outages can have severe consequences on society, including security, health, education, and public services. To manage the time and costs of power outages, fault studies is essential. Traditional methods for fault classification and location involve post-event mathematical computations, while AI-based methods use pre-event knowledge. These reasons have led us to focus on a Deep Learning (DL) model for fault classification and location in the distribution networks that offers higher accuracy with reduced computational requirements.