Real-Time Fault Detection and Classification in Radial Power Distribution Network using DNNs

Author

Nimmi Regmi, Southern Illinois University CarbondaleFollow

Date of Award

8-1-2025

Degree Name

Master of Science

Department

Electrical and Computer Engineering

First Advisor

Anagnostopoulos, Iraklis

Abstract

Electric power distribution systems are evolving rapidly with increased integration ofdistributed generation (DG), flexible load, and bidirectional power flows. While this is a new norm of modern power systems and advantageous in many ways, it also comes with its own share of added complexity and problems that need due attention. One such example is of the protection systems in the power grid, which are conventionally designed to function with a fixed direction of current flow and pre-defined logic of operations. With added variability and uncertainties discussed above, the conventional protection and fault detection systems in the power grid sometimes fall short in identifying and classifying faults with the desired accuracy. Inspired by the wide range of applications of neural networks in complex operations, this thesis proposes a data-driven, lightweight Deep Neural Network (DNN) based solution methodology for accurate and real-time fault detection and classification in radial distribution networks with DG penetration. This involves simulating the modified IEEE 4 bus distribution network in ePHASORsim solver based on OPAL-RT hardware simulator with numerous fault scenarios to generate a dataset with respective fault level, train and optimize different DNN architectures (MLP and 1-D CNN) based on the dataset, and deploy the most efficient trained DNN for real-time fault detection and classification in the radial power distribution network. Different distribution network parameters like fault resistance, connected load, etc., were varied over a wide range in the simulated scenario to emulate the complexity and anomalies of real-world cases. Likewise, the feature extraction and point of detection or control were chosen such that they closely match the natural position of conventional power systems protection relays and sensing elements. The performance of the DNN models was closely monitored during the training in terms of accuracy and inference time by using accuracy percentage, confusion matrix, etc., as the metrics. Likewise, the hyperparameters of the DNN models were tuned using Bayesian Optimization while co-optimizing the accuracy and inference time using the Pareto front. Once trained, the models acted as independent black boxes capable of analyzing the power distribution network features—current and voltage—in real time and making inferences on the go. The results section highlights the proposed methodology’s performance in terms of accuracy and inference time, and confirms its effectiveness, robustness, and applicability in modern radial power distribution networks.