Smart Grid Fault Detection Using Machine Learning Techniques

Thesis Abstract

The increasing integration of distributed energy resources and the evolution of smart grid technologies have heightened the imperative for rapid and accurate fault detection to ensure reliability, efficiency, and security of electrical power systems. Traditional fault detection techniques are often limited by their reliance on static thresholds and manual monitoring, which can lead to delayed response times and potential system failures. This research aims to develop an advanced, machine learning-based fault detection framework that enhances the responsiveness and accuracy of monitoring in smart grids. The primary objectives include identifying suitable machine learning algorithms for fault classification, designing a scalable data processing architecture, and evaluating the system’s performance in realistic operational scenarios. A mixed-methods research design was adopted, combining quantitative and qualitative approaches to ensure comprehensive system development and validation. The quantitative component involved empirical data analysis, while qualitative insights informed system usability and implementation strategies. The population targeted comprised operational data from a regional smart grid comprising 150 distribution feeders, with a specific focus on fault incidents recorded over a two-year period. A stratified random sampling technique was employed to select 300 fault records, ensuring representation across various fault types, including short circuits, ground faults, and line-to-line faults. Data collection was carried out through existing SCADA (Supervisory Control and Data Acquisition) systems, supplemented by historical fault logs and real-time sensor data. Custom data acquisition interfaces were developed to extract high-resolution voltage, current, and frequency measurements at sampling rates of 1 kHz, enabling detailed feature analysis. To enhance data quality, preprocessing steps involved noise filtering, normalization, and feature extraction, including time-domain, frequency-domain, and wavelet transform-based features. Machine learning models such as Support Vector Machines (SVM), Random Forests, and Convolutional Neural Networks (CNN) were trained and validated using a 7030 training-test split. Model performance was evaluated based on accuracy, precision, recall, F1-score, and confusion matrices, with cross-validation applied to prevent overfitting. Additionally, the study employed the Receiver Operating Characteristic (ROC) curve and Area Under the Curve (AUC) metrics to compare classifier robustness. Feature importance analysis was conducted using permutation feature importance and SHAP (SHapley Additive exPlanations) to interpret model predictions. The research also integrated the Theory of Self-Organizing Networks and the Complex Adaptive Systems theory to underpin the adaptive learning capabilities of the proposed models. Expected findings include a significant improvement in fault detection accuracy, reduction in false alarms, and faster fault localization times compared to conventional threshold-based methods. It is anticipated that the CNN model outperforms other classifiers in handling complex fault signatures, with over 95% accuracy in identifying fault types. The study also expects to demonstrate the feasibility of deploying the developed models within existing smart grid infrastructures, supported by scalable data processing architectures. The contribution of this research lies in establishing a robust, machine learning-driven fault detection framework that enhances the situational awareness and operational resilience of smart grids. It advances the understanding of applying deep learning and ensemble techniques in power system monitoring, providing a foundation for real-time fault diagnosis and predictive maintenance strategies. The study concludes with recommendations for integrating the proposed system into grid management protocols, emphasizing the importance of continuous learning and adaptive model updating to cope with evolving grid conditions. It also highlights future research pathways, including the application of transfer learning and semi-supervised techniques to further improve detection capabilities in data-scarce environments.