Smart Grid Fault Detection in a Metropolitan Distribution Network: A Case Study

Thesis Abstract

The rapid evolution of smart grid technologies in urban distribution networks presents a critical challenge for timely fault detection, restoration, and reliability assurance in metropolitan contexts. This study addresses the problem of latent fault signatures, intermittent disturbances, and data-driven misclassification that impede rapid fault isolation in a large-scale distribution network operated by a metropolitan utility. The aim is to develop and validate an integrated fault detection framework that leverages synchrophasor measurements,Weather-aware feature extraction, and hybrid machine learning to improve detection accuracy and mean time to repair (MTTR). Specific objectives are to (i) characterize fault signatures under different network states using PMU/PMU-like phasor data, (ii) design a data fusion architecture that combines SCADA, PMU, and smart meter streams for robust feature engineering, (iii) develop a hybrid anomaly detection model that fuses supervised and unsupervised learning to identify faults in near real-time, (iv) evaluate the framework on a real-world metropolitan distribution feeder with diverse fault scenarios, and (v) provide actionable recommendations for switching coordination and restoration prioritization. The methodology adopts a mixed-methods research design combining quantitative data-driven modeling with qualitative validation from grid operators. The population comprises three concentric layers (a) the metropolitan distribution network operator’s feeder-level data (approximately 120 feeders, 5,000 PMU samples per day over 12 months), (b) historical fault records and restoration logs, and (c) operator interview transcripts to capture decision workflows. A stratified random sample of 60 feeders with diverse loading and topology characteristics is selected for in-depth analysis. Data collection instruments include high-fidelity PMU and SCADA data streams, fault event catalogs, and structured interview protocols. Instrument validity and reliability are ensured through cross-validation with historical fault classifications and test-retest procedures, with measurement reliability indices (Cronbach’s alpha for survey items) maintained above 0.8 and PMU data quality checks exceeding 99.5% uptime. The analytical framework integrates time-series preprocessing, feature extraction, and classification. Key techniques include (i) wavelet packet decomposition and Shannon entropy for transient fault signature extraction, (ii) graph-based feature representations of network topology, and (iii) a hybrid classifier that combines a supervised long short-term memory (LSTM) network with an unsupervised isolation forest to detect anomalies and classify fault types. Model validation employs k-fold cross-validation (k=10) and out-of-sample testing on quarterly fault events. Hypothesis testing uses paired t-tests to compare detection latency before and after framework deployment, complemented by non-parametric Mann-Whitney tests where distributional assumptions are violated. The conceptual foundation draws on resilience theory and the control theory of networked systems, with the Information Processing Theory guiding feature relevance and interpretability. A conceptual model illustrating data fusion, fault signature extraction, and decision-support outputs is developed and refined iteratively with operator feedback. Expected findings indicate a measurable reduction in fault detection latency by 25–40% and a corresponding MTTR improvement of 15–30% for major fault types (phase-to-ground, line-to-line, and symmetrical faults). The hybrid model is anticipated to achieve an overall fault classification accuracy above 92% on held-out fault instances, with robustness to missing data and varying communication delays. The study also expects to uncover practical insights into operator decision workflows, revealing how real-time indicators influence switching strategies and restoration prioritization. The study contributes to knowledge by delivering a scalable, data-driven fault detection framework tailored to metropolitan distribution networks, integrating phasor data analytics, topology-aware features, and hybrid learning to enhance reliability engineering practice. It provides empirical evidence on the benefits and limitations of multi-source data fusion for fault detection and offers a transferable methodology for utilities implementing smart grid resilience programs in dense urban environments. The main conclusion posits that an integrated, topology-informed hybrid learning approach substantially improves both the speed and accuracy of fault detection and classification. Recommendations include deploying the framework with incremental rollout on critical feeders, integrating operator feedback loops for continual model recalibration, and aligning restoration prioritization with the framework’s probabilistic confidence outputs to optimize outage management strategies.