Robust Fault Detection and Classification in Power Systems via Physics-Informed and Data-Driven Learning

Electrical faults in power transmission systems can severely affect grid stability, equipment safety, and operational reliability. Traditional protection schemes, particularly distance relays, depend on apparent impedance computation that changes with error, creating a risk of misclassification. The results from relay overreach, underreach, or complete maloperation due to CT/PT saturation lead to developing problems in high impedance situations. These limitations highlight the need for adaptive, data-driven alternatives. This paper proposes an intelligent fault detection and classification model based on supervised machine learning techniques that overcome these challenges. The system’s robustness was validated under different training sizes and Gaussian noise levels, demonstrating consistent accuracy and generalization across diverse learning conditions. The presented approaches learn the complex nonlinear mapping between three-phase voltage/current patterns and the associated fault type, without assuming fixed impedance paths like traditional protection schemes. This method extracts a high set of derived features to represent the distinguishing characteristics of six fault categories by utilizing line voltages and currents. Different models such as Artificial Neural Networks (ANN), Support Vector Machines (SVM), Random Forests, XGBoost, Long Short-Term Memory (LSTM), and Physics-Informed Neural Networks (PINN) are developed and processed on SMOTE-balanced datasets. These models classify the errors without fixed thresholds or fault loop assumptions, improving sensitivity and robustness. The supervised machine learning approaches bridge the gap between traditional impedance-based protection and smart, scalable data-driven grid analytics that are implemented into a wide area monitoring and control system. The PINN achieved the highest fault detection accuracy of 99.86% while sustaining 99.79% multiclass classification accuracy on the clean dataset. The PINN maintains high accuracy under 2–5% noise and 1–60% training data, providing millisecond-level inference by embedding power system equations, which enables accurate real-time protection by understanding the missing simple data-driven parameters.