Development of an AI-Enabled Smart Grid Fault Detection System

Thesis Abstract

The increasing complexity and growing reliance on electrical power systems underscore the critical need for real-time fault detection and rapid response within smart grid infrastructures. Traditional fault detection methods often face limitations related to latency, accuracy, and scalability, thereby compromising grid reliability and potentially leading to costly outages. This study aims to develop an advanced Artificial Intelligence (AI)-enabled fault detection system tailored for smart grids, with the objectives of enhancing detection accuracy, reducing response time, and increasing system resilience. Specifically, the research seeks to design and validate an AI framework capable of identifying and classifying faults with high precision, integrating machine learning algorithms such as convolutional neural networks (CNNs) and support vector machines (SVMs), and benchmarking their performance against conventional methods. The research adopts a mixed-methods approach grounded in a quantitative paradigm to facilitate robust measurement and evaluation of model performance, supplemented by qualitative insights for system interpretability. The population for the study encompasses data collected from a regional smart grid operated by an integrated utility company, comprising approximately 20,000 operational data points including voltage, current, and frequency measurements captured over a 12-month period. A stratified random sampling technique was employed to select a representative subset of 2,000 data points, ensuring balanced inclusion of normal operations and various fault conditions such as short circuits, line-to-ground faults, and equipment failures. Data collection instruments include embedded sensors, Phasor Measurement Units (PMUs), and historical fault incident reports, which are pre-processed and annotated to serve as training and testing datasets. Data analysis involves the application of supervised machine learning techniques, with model training conducted using 70% of the data, and validation using the remaining 30%. Feature extraction methods such as principal component analysis (PCA) are employed to optimize input variables, while hyperparameter tuning is performed via grid search to enhance model accuracy. The study compares the performance of CNNs and SVMs through metrics such as accuracy, precision, recall, and F1-score, employing ANOVA tests to determine statistical significance in model differences. Additionally, the study applies the Theory of Adaptive Systems and the Information Processing Theory as conceptual frameworks to underpin the AI models' adaptability and decision-making processes. Expected findings include a significant improvement in fault detection accuracy—anticipated to exceed 95%—and a reduction in detection latency by approximately 40% compared to traditional threshold-based systems. The models are projected to demonstrate high robustness across different fault types and load conditions, with CNNs likely outperforming SVMs due to their superior feature extraction capabilities. The analysis is expected to reveal insights into the most salient features influencing fault classification, thereby informing future model enhancements and practical deployment strategies. This research contributes to the body of knowledge by integrating advanced AI techniques into smart grid fault management, thus advancing understanding of machine learning applications within power systems engineering. It bridges the gap identified in previous literature regarding real-time and scalable fault detection solutions, offering a practical framework for utility operators seeking to enhance grid stability and reliability. The study also extends the theoretical application of adaptive and information processing models within the context of predictive maintenance and fault management. Based on the findings, the study recommends adopting the developed AI-based system for real-time monitoring in operational smart grids, with suggested pathways for further research including real-world pilot implementations, integration with IoT-enabled infrastructure, and exploration of more complex deep learning architectures. It concludes that the deployment of AI-driven fault detection systems has the potential to significantly transform power system reliability, reduce maintenance costs, and support the transition towards fully autonomous smart grid networks.