Joint Sensor Deployment and Physics-Informed Graph Transformer for Smart Grid Attack Detection

A research team led by Mariam Elnour has published a novel AI framework that simultaneously optimizes sensor placement and attack detection for critical power grids. The core innovation is a Physics-Informed Graph Transformer Network (PIGTN), a detection model that integrates the physical laws of AC power flow with graph neural networks to understand grid topology and behavior. This model is trained in a closed-loop with the Non-dominated Sorting Genetic Algorithm-II (NSGA-II), which explores the massive combinatorial space of possible sensor locations. The joint optimization aims to maximize detection performance while adhering to real-world constraints like sensor failures and costs, moving beyond traditional methods that treat sensor placement and detection as separate problems.

The results, validated across seven standard power system models (including IEEE 30, 57, 118, and a 200-bus system), show significant improvements. The PIGTN detector generalized well to unseen attacks, outperforming other graph-based models by up to 37% in accuracy and 73% in detection rate, while maintaining a remarkably low mean false alarm rate of 0.3%. Furthermore, the optimized sensor layouts drastically improved the accuracy of estimating the grid's operational state (a process called State Estimation), reducing the average error by 61% to 98%. This dual advancement in both 'seeing' (sensor deployment) and 'understanding' (AI detection) represents a major step toward resilient infrastructure, where AI doesn't just monitor but actively helps design more secure and observable energy networks.