A Data-Driven Framework for Fault Diagnosis and Electrical Load Prediction

This paper presents a machine learning framework for fault diagnosis and load forecasting in power systems. As modern power grids become increasingly complex, intelligent methods are essential for reliable operation and energy efficiency management. This work utilizes historical electricity data and system information to develop smart models capable of predicting future electricity demand and identifying faults in power systems. For load forecasting, the XGBoost Regressor algorithm learns patterns in electricity consumption across different time periods, enabling better supply planning and overload prevention. For fault diagnosis, the XGBoost Classifier analyzes voltage, current, and other system parameters to detect faults before they escalate into serious problems. The methodology employs real-world smart grid data with features including temperature, humidity, solar/wind power generation, and transformer conditions. The load forecasting model achieves an R² score of 0.8997 with RMSE of 0.1129 kW, demonstrating high prediction accuracy. The fault diagnosis model achieves 96.11% overall accuracy, with 98% precision in normal condition detection, 95% accuracy in transformer fault identification, and 87% recall in overload condition detection. Compared to traditional methods like ARIMA, the proposed XGBoost approach reduces prediction errors by 67.7% and improves fault detection accuracy by over 5% compared to existing WSN-based methods. The framework provides power companies with a simple, effective tool for faster decision-making, improved reliability, and reduced operational risks in smart grid systems.