Development of an AI-enabled Smart Grid Fault Detection System

Thesis Abstract

The increasing complexity and demand for reliable power transmission in contemporary electrical systems necessitate advanced fault detection mechanisms to ensure uninterrupted service and system resilience. Traditional fault detection methods often rely on manual monitoring, rule-based systems, or simplistic signal analysis techniques, which are often inadequate for the dynamic, real-time operational environment of modern smart grids. This study aims to develop an AI-enabled fault detection system that leverages machine learning algorithms and real-time data analytics to enhance the accuracy, speed, and reliability of fault identification within smart grid infrastructure. The specific objectives include (1) designing a comprehensive data acquisition framework that captures high-frequency electrical parameters from various segments of the grid; (2) developing machine learning models, including convolutional neural networks (CNNs) and ensemble classifiers, to classify fault types and locations; (3) evaluating the effectiveness of the proposed models in different fault scenarios using historical and simulated data; and (4) integrating the developed system into a prototype smart grid environment for real-time testing and validation. The research adopts a cross-sectional quantitative design grounded in the systems theory and the anomaly detection framework. Data will be collected from a smart grid laboratory comprising a network of 50 simulated feeder lines with embedded sensors to record voltage, current, and frequency variations under both normal and fault conditions. A stratified sampling technique will be used to select 200 fault events from this dataset, with data augmentation techniques employed to increase the robustness of model training. Instrumentation includes high-precision digital oscilloscopes, synchronized phasor measurement units (PMUs), and custom data loggers, ensuring high fidelity and temporal resolution. The validity and reliability of the data collection instruments will be assured through calibration and repeated measurements, while data preprocessing will involve noise filtering, normalization, and feature extraction. Analytical procedures will include supervised machine learning model training and testing, employing techniques such as support vector machines (SVM), CNNs, and random forest classifiers, with model performance evaluated through metrics like accuracy, precision, recall, and F1-score. Cross-validation will mitigate overfitting, and hyperparameter tuning will optimize model effectiveness. Additionally, comparative analysis using receiver operating characteristic (ROC) curves will identify the most effective models for specific fault types. The study will explore the integration of real-time simulation frameworks using MATLAB/Simulink to validate system performance under variable environmental conditions. Expected findings indicate that AI-driven models, particularly CNNs combined with ensemble methods, will significantly outperform traditional threshold-based systems in fault detection accuracy and response time. The system is anticipated to identify fault types at a rate exceeding 95% accuracy within milliseconds, facilitating prompt restoration and minimizing system downtime. Furthermore, the study expects to demonstrate the model’s robustness across different fault scenarios, including short circuits, open circuits, and equipment failures, under varying load conditions. This research contributes to existing knowledge by advancing intelligent fault detection methodologies tailored for smart grids, highlighting the efficacy of deep learning techniques in power system protection. It provides a scalable, real-time fault detection architecture adaptable to diverse grid configurations, thereby reinforcing grid resilience and operational efficiency. The findings will inform utilities and system operators about the integration of AI tools into standard fault management protocols and support policy development towards smarter, more resilient power networks. The main conclusion underscores the potential of AI-enabled systems to transform traditional grid protection strategies, emphasizing their role in sustainable and reliable energy distribution. The study recommends further research on integrating additional data sources such as satellite imaging and environmental sensors, exploring unsupervised learning for anomaly detection, and implementing large-scale pilot projects to evaluate system performance in operational environments. Overall, this thesis aims to set a benchmark for intelligent fault detection systems, fostering innovation in smart grid management through cutting-edge artificial intelligence solutions.