Robust fault detection and uncertainty quantification in smart grids using graph neural networks

Due to the development of smart grids with the integration of the Internet of Things (IoT) devices, renewable energy sources (RES), and AI-driven solutions, the need for an advanced fault detection and uncertainty quantification framework has emerged. The vast and variable data generated by these networks result in the failure of traditional fault detection methods to manage the data, leading to delays in the fault identification and poor uncertainty estimates. Graph Neural Networks (GNN)-Smart Detect, a model for GNNs to improve accuracy and exceptionally robust uncertainty quantification in error of interconnected innovative grid components, is introduced in this study. GNN-SmartDetect exploits graph-structured information to capture topological dependencies among grid nodes and improves its anomaly detection and fault classification performance even under noisy and incomplete data situations. Our experimental data show that GNN-SmartDetect can detect 96.4 % of faults with outstanding performance, and two orders of magnitude lower false positive and false negative rates compared to the traditional models. Moreover, the built-in uncertainty quantification system provides dependable confidence levels by each detection and therefore allows operators to estimate risk and come up with proactive reactions in the face of threats whilst they are not known in grid conditions. It is also scalable and can be adapted to different grid configurations including microgrids and national grids, which shows that the model can be implemented in real time under Industry 4.0 operations smart grid operations. As a result, these results indicate that GNN easy-to-detect is a good and significant addition to the grid management, operational resilience, and grid stability development in the context of increasingly sophisticated energy markets.