Intelligent Fault Diagnosis in Power Grids via Edge AI Analytics

Thesis Abstract

The modern electric power grid faces increasing fault incidence and complexity due to distributed generation, high penetrations of renewable energy, and evolving communication infrastructures, which together challenge conventional monitoring and fault diagnosis approaches. This study addresses the problem of timely and accurate fault diagnosis in power grids by leveraging edge artificial intelligence analytics to enable real-time decision-making at substation and distribution levels, reducing outage duration and improving system reliability. The aim is to develop an edge-enabled fault diagnosis framework that fuses heterogeneous sensor data from phasor measurement units, digital relays, and smart meters using lightweight deep learning models and uncertainty-aware inference. Specific objectives are (1) to design an edge-optimized neural architecture capable of manifold data fusion and robust fault classification under communication latency and bandwidth constraints; (2) to integrate physics-informed priors and domain-specific features into the learning process to enhance interpretability and resilience to cyber-physical threats; (3) to validate the framework on a comprehensive emulated grid dataset and a real-world microgrid testbed; (4) to quantify performance under varying load, topology, and noise conditions; and (5) to assess operational feasibility, including deployment considerations, energy efficiency of edge devices, and resilience to data loss. The methodology adopts a mixed-methods, multi-stage research design. The population comprises distribution networks with high renewable penetration and embedded microgrids, including a 100-bus synthetic test system and a 15-bus real microgrid testbed installed with PMUs, PMU-like sensors, and protective relays. A stratified sampling approach selects fault scenarios across short-circuit, open-circuit, and transient faults, yielding a labeled dataset of 20,000 time-series samples for training and 5,000 for testing. Data collection instruments include high-resolution sensor streams (sampling rates of 1 kHz to 5 kHz), event records, SCADA logs, and fault records provided by the testbed and synthetic simulations. The study employs a dual-model strategy (i) a lightweight edge neural network (e.g., a constrained convolutional-recurrent architecture with quantization) deployed on industrial-grade edge devices at substations, and (ii) a cloud-based ensemble for model calibration and long-term learning. Validity and reliability are ensured through cross-validation, hold-out testing across topology variations, and calibration against known fault signatures. Ethical considerations address data privacy, cyber-security implications, and adherence to grid operation safety standards. Analytical techniques include time-domain and frequency-domain feature extraction (transient phasor analysis, differential protective signals, wavelet coefficients), followed by supervised learning with stratified k-fold cross-validation. Model performance is evaluated using confusion matrices, precision, recall, F1-scores, and area under the ROC curve, with ablation studies to determine the contribution of edge inference versus cloud-based augmentation. Sensitivity analyses explore the impact of communication delays, data loss, and device heterogeneity. Interpretability is enhanced through SHAP (SHapley Additive exPlanations) values and feature importance ranking, complemented by a physics-informed neural network (PINN) component that enforces electrical constraints. The theoretical foundation integrates reliability-centered maintenance theory and distributed AI with edge computing paradigms, drawing on the Technology Acceptance Model to assess operator trust and adoption potential. Expected findings indicate that the edge AI framework achieves fault classification accuracy above 95% within sub-second latency on edge devices, resilience to 20% data loss, and improved detection of high-impedance faults compared to centralized analytics. The integration of physics-informed priors is anticipated to reduce misclassifications by 15% and provide interpretable indications of fault location and type. The study anticipates demonstrating substantial reductions in detection time and outage duration, with energy-efficient edge inference consuming less than 2 W per device under typical operating conditions. The contribution to knowledge includes a novel edge-centric fault diagnosis architecture for power grids, a validated methodology for real-time data fusion and uncertainty-aware inference in constrained environments, and empirical evidence on the trade-offs between edge and cloud analytics in cyber-physical power systems. The main conclusion posits that edge AI analytics, when coupled with physics-informed features and robust validation on realistic testbeds, can deliver timely, accurate, and interpretable fault diagnosis for modern grids, enabling proactive protection and maintenance strategies. Recommendations emphasize standardized data formats and interface protocols for edge deployment, guidelines for selecting edge hardware with sufficient compute and energy efficiency, and future work exploring adaptive learning to accommodate topology changes and evolving grid configurations.