Smart Grid Anomaly Detection Using Explainable Graph Neural Networks

Project Abstract

Smart Grid Anomaly Detection Using Explainable Graph Neural Networks presents a robust, scalable framework for real-time identification and interpretation of anomalies in modern electrical power systems. This work addresses the critical need for reliable monitoring in smart grids, where high-dimensional, dynamic data from phasor measurement units (PMUs), smart meters, and distributed energy resources converge, creating complex interdependencies that are difficult to model with traditional methods. We propose an explainable graph neural network (GNN) architecture that captures spatial–temporal relationships among grid components by representing the grid as a graph with nodes corresponding to buses, transformers, and DERs, and edges encoding electrical connections and operational dependencies. The model integrates temporal attention mechanisms and multi-relational edge features to detect anomalies such as sensor faults, cyber-physical intrusions, line outages, and load imbalances, while maintaining interpretability through post-hoc and intrinsic explanations. The methodology comprises data fusion from multiple grid data streams, preprocessing for noise reduction and alignment, and the construction of a heterogeneous graph that reflects network topology and operational constraints. A novel loss formulation combines reconstruction-based anomaly scoring with a contrastive learning objective to enhance discrimination between normal and anomalous states under varying operating conditions. The explainability component leverages graph-level and node-level saliency, attention weight visualization, and counterfactual reasoning to identify contributing factors and affected subsystems, enabling operators to validate and act upon detected anomalies. To ensure robustness, we incorporate adversarial training and domain adaptation techniques to generalize across different grid configurations and seasonal load patterns. We evaluate the proposed framework on a comprehensive, multi-year real-world grid dataset augmented with simulated fault scenarios to cover rare but critical events. Baseline comparisons include state-of-the-art statistical methods, deep autoencoders, and non-explainable GNNs. Evaluation metrics encompass detection accuracy, precision-recall balance, false-positive rate, detection latency, and interpretability scores assessed through operator studies and qualitative assessments of explanation fidelity. Experimental results demonstrate that the explainable GNN approach yields superior detection performance, with notable improvements in early anomaly detection and reduced misclassification during high-variance periods. The explainability module provides actionable insights by tracing anomalies to specific network components and operational contexts, facilitating faster diagnosis and targeted mitigation strategies. From a practical standpoint, the framework supports online deployment with incremental updating to adapt to topology changes and evolving DER penetration. We also discuss integration with existing supervisory control and data acquisition (SCADA) systems, cybersecurity considerations, and the potential for automated corrective actions guided by interpretable anomaly explanations. The study contributes to the advancement of transparent, data-driven anomaly detection in smart grids, enabling improved reliability, resilience, and situational awareness for grid operators.

Project Overview

What This Project Is About

---

A plain-language overview of the topic and what the project investigates.

---

---

The Problem It Addresses

---

There are frequent faults and unusual patterns in power grids that can disrupt service or cause damage. Detecting these anomalies quickly and explaining why they are flagged helps operators respond faster and with better decisions.

---

---

Objectives of the Project

---

1. Identify common anomaly patterns in smart grid data.
2. Develop a model that detects unusual behavior using graph-based data representations.
3. Make the model's decisions understandable to non-experts (explainability).
4. Evaluate accuracy, speed, and usefulness on real or simulated grid data.
5. Provide practical guidelines for deployment in control rooms.

---

---

What You Will Do Step by Step

---

1. Study basic ideas of grids, sensors, and data types collected from them.
2. Collect or simulate grid data with normal and abnormal events.
3. Turn data into a graph structure where nodes are grid components and edges show interactions.
4. Build a graph neural network model to detect anomalies.
5. Incorporate methods that explain why a decision was made (which parts of the graph mattered).
6. Test the model against different scenarios and compare with simpler methods.
7. Analyze performance and interpretability results.
8. Document how to deploy the approach in a real system.

---

---

Expected Outcome

An interpretable anomaly detection system for smart grids that performs well on test data and can be explained to operators. The project should yield a working prototype, evaluation results, and practical guidance for real-world use.