Smart Grid Fault Detection and Localization Using Phasor Measurement Units and Deep Learning

Project Abstract

Smart grid reliability hinges on rapid and accurate identification of faults and their precise locations within complex transmission networks. This study presents a novel fault detection and localization framework that integrates synchronized Phasor Measurement Unit (PMU) data with deep learning models to enhance situational awareness and decision-making in real-time protection and restoration. The proposed approach leverages high-fidelity PMU measurements (voltage, current, frequency, and phase angle) collected from multiple substations to capture dynamic system behavior under transient and steady-state fault conditions. A two-stage deep learning architecture is employed (i) a fault classification module that distinguishes fault types (short-circuits, line-to-line, line-to-ground, three-phase faults) and estimates preliminary event parameters, and (ii) a localization module that computes accurate fault locations using a graph-structured neural network (GNN) that exploits network topology, line impedances, and PMU-derived feature embeddings. To address data scarcity for rare fault events, a physics-informed data augmentation strategy is implemented, embedding circuit laws and network constraints to generate physically plausible synthetic training samples, thereby improving model generalization. The models are trained on a comprehensive synthetic dataset generated from a detailed 118-bus IEEE test feeder and validated on a real-world utility dataset to ensure robustness across diverse operating conditions, including varying loading levels, renewable generation penetration, and PMU placement densities. The framework incorporates time-series analysis with attention mechanisms to capture temporal patterns preceding and during faults, enabling earlier detection and reduced false alarms. A roadmap for real-time deployment is outlined, including synchronization, communication latency considerations, and cyber-physical security measures to mitigate data integrity risks. Performance evaluation emphasizes detection latency, accuracy, and localization error under multiple fault scenarios, with benchmarks against conventional protection schemes and machine learning baselines. Sensitivity studies examine the impact of PMU placement, measurement noise, and missing data, while ablation studies quantify the contribution of each component—PMU data, temporal modeling, topology-aware learning, and physics-informed augmentation. The results demonstrate that the proposed method achieves sub-cycle detection and localization with high accuracy, significantly reducing service interruption times and improving fault isolation effectiveness. Practical implications for system protection coordination, maintenance planning, and accelerated restoration are discussed, along with considerations for scalability to larger networks and integration with energy management systems. The research also addresses interpretability by mapping model inferences to physically meaningful features, enabling operators to validate and trust automated decisions. Potential limitations include dependence on PMU availability and data quality, which are mitigated through robust preprocessing and redundancy planning. This work contributes to the advancement of intelligent grid analytics by combining high-fidelity phasor data with deep learning-driven inference, offering a viable pathway toward enhanced resilience and reduced outage costs in modern power systems.

Project Overview

What This Project Is About

---

A practical study that looks at how electrical grids can automatically detect and locate faults using measurements from specialized sensors and smart computer methods. It combines real-time data from the grid with simple, readable explanations to help understand when and where problems occur and how to fix them quickly.

---

---

The Problem It Addresses

---

Electrical grids are large and complex, so faults can happen anywhere and spread quickly. Traditional methods may be slow or inaccurate. This project explores a modern approach that uses precise electrical measurements and computer learning to improve speed and accuracy in fault detection and pinpointing.

---

---

Objectives of the Project

---

1. Identify what signals from the grid indicate a fault.
2. Understand how to locate the fault using measurement data.
3. Explore a simple learning method that can classify fault types and locations.
4. Show how the method performs on simulated grid data.
5. Discuss practical limitations and real-world usability.

---

---

What You Will Do Step by Step

---

1. Learn basic grid concepts and the sensors used to collect data.

2. Create or obtain labeled fault data for testing.

3. Build a simple detection-and-localization model using easy-to-understand machine learning ideas.

4. Validate the model with test data and measure accuracy.

5. Analyze what affects performance and how to improve it.

---

---

Expected Outcome

A clear, student-friendly method that can detect faults and say where they occurred in the grid, with discussion of how this could be useful for grid operators and safety improvements.