Edge-Computing Accelerated Fault Diagnosis in Power Grids Using AI Edgelets

Thesis Abstract

The integration of edge computing with intelligent edge analytics offers a transformative approach to real-time fault diagnosis in modern power grids, addressing latency, bandwidth, and reliability constraints inherent to centralized cloud-based solutions. This study addresses the persistent challenge of timely and accurate fault detection in heterogeneous distribution and transmission networks, where conventional SCADA and PMU-based methods suffer from delayed response and limited local decision support. The aim is to develop an edge-centric fault diagnosis framework that leverages AI edgelets—compact, deployable AI modules running on near-field edge devices—for accelerated, interpretable fault classification and isolation under diverse operating conditions. Specific objectives include (1) to design and implement an AI edgelet architecture that enables distributed feature extraction, model inference, and collaborative decision fusion across substation gateways and regional edge nodes; (2) to formulate a scalable fault taxonomy and build a labeled dataset capturing transient, persistent, and cascading fault scenarios across 132/33 kV and 11 kV feeders; (3) to develop lightweight, hardware-aware machine learning models (e.g., pruned deep learning classifiers, graph neural networks, and ensemble methods) with provable bounds on inference latency and energy consumption; (4) to evaluate robustness under noise, communication delays, and data loss using a hybrid testbed that combines real-world archival recordings with synthetic perturbations; and (5) to provide guidelines for deployment, security, and governance of AI edgelets in operational grids. The methodology adopts a mixed-methods research design combining quantitative experimentation with qualitative stakeholder feedback. The population includes utility-grade substations and feeder sections within a regional transmission organization’s network. A stratified sampling technique selects 20 substations and 60 feeders, with data collected from synchronized PMU streams, fault records, relay logs, and ambient sensor readings over a 12-month period, supplemented by an open-access synthetic fault dataset to ensure rare-event representation. Data collection instruments comprise high-resolution time-stamped measurement collections, edgelet software decoders, and diagnostic labels corroborated by operator annotations. Validity and reliability are ensured through cross-validation, k-fold testing, and repeated-measures assessments of edge performance under variable network loads. The analysis uses a multi-tier framework (i) descriptive statistics to characterize fault incidence and edge latency distributions; (ii) regression-based and time-series analyses to quantify the relationship between feature quality, network conditions, and diagnostic accuracy; (iii) comparative model evaluation employing cross-validated precision, recall, F1-score, and area under the ROC curve for classifiers such as lightweight convolutional neural networks, graph neural networks, and gradient-boosted trees; (iv) ablation studies to identify critical edgelet components; (v) latency-energy trade-off assessment through profiling on representative edge hardware (e.g., ARM-based gateways with AI accelerators); and (vi) Monte Carlo simulations to propagate data losses and communication delays through the fusion architecture. A theoretical underpinning is provided by the Combine-and-Conquer framework for distributed inference and the Information Bottleneck principle to balance feature sufficiency against resource constraints, complemented by governance-informed security models addressing adversarial robustness and data integrity. Expected findings indicate that AI edgelets can reduce average fault diagnosis latency by 40–60% and improve classification accuracy by 5–12 percentage points relative to centralized baselines, while maintaining sub-50 ms inference times for common fault types and demonstrating resilience to up to 20% data loss in edge-to-edge communications. The study contributes to knowledge by delivering a validated edge-centric fault diagnosis architecture, a benchmark dataset for AI edgelets in power grids, and guidelines for deployment, safety, and interoperability. It concludes that distributed AI edgelets, when aligned with hardware-aware model design and secure communication protocols, can significantly enhance situational awareness, reduce outage durations, and improve grid reliability. Recommendations include standardizing edgelet interfaces, expanding the dataset with extreme weather faults, and pursuing pilot implementations with utility partners to assess long-term operational benefits and regulatory compliance.