Abstract
Geomagnetic disturbances (GMDs) pose a significant threat to high-voltage power grids by inducing quasi-DC currents that can lead to transformer failure. While neutral blocking devices (DC blockers) can mitigate these risks, determining their optimal placement is a computationally intensive Mixed-Integer Non-Linear Program (MINLP). This paper presents a Physics-Informed Heterogeneous Graph Neural Network (PIHGNN) that leverages the structural properties of power systems and physical constraints to provide high-quality placement strategies with significant speedups over traditional optimization solvers.
Introduction
Solar storms, or geomagnetic disturbances, induce geomagnetically induced currents (GICs) in power lines. These currents flow through transformer neutrals into the ground, causing saturation, overheating, and potential grid collapse. To protect infrastructure, engineers install DC blockers. However, the high cost of these devices necessitates an optimized placement strategy. Traditional mathematical solvers struggle with the non-convex, non-linear nature of this problem at scale. While standard machine learning (ML) offers speed, it often fails to account for the heterogeneous nature of grid components and the underlying physics of power flow.
Methodology
The proposed PIHGNN architecture addresses these challenges through two primary innovations:
1. Heterogeneous Graph Representation
Unlike standard Graph Neural Networks (GNNs) that treat all nodes as identical, the PIHGNN models the grid as a heterogeneous graph. It utilizes tailored embeddings for distinct node types, such as AC buses, DC nodes, and generators, as well as specific edge types for transmission lines and transformers. A key feature is the use of "virtual relations" to link components that represent the same physical entity but operate differently in AC and DC models.
2. Physics-Informed Training
To ensure the model respects the laws of physics, a surrogate power flow evaluator is integrated into the training loop. The loss function combines standard classification loss with a physics-informed penalty. This penalty discourages predictions that violate physical constraints or result in poor power delivery. An adaptive training algorithm gradually introduces this physics-based penalty, allowing the model to refine its understanding of grid constraints over time.
Experimental Results
The model was evaluated on the EPRI21 (20-bus) and UIUC150 (150-bus) synthetic benchmarks using 400 GMD scenarios with varying electric field magnitudes and directions.
- Accuracy: PIHGNN demonstrated superior performance, improving average accuracy by 5.2% on EPRI21 and 1.3% on UIUC150 compared to standard HGNNs.
- Computational Efficiency: The speedup is substantial. On the UIUC150 system, a traditional heuristic solver required approximately 390 seconds per sample. In contrast, the PIHGNN completed the same task in under 4 seconds, with raw inference taking less than one second.
- Generalization: The model showed promising transferability; a PIHGNN trained on EPRI21 achieved 67.2% accuracy when tested on the significantly different UIUC150 topology.
Conclusion and Future Work
By combining the structural strengths of heterogeneous GNNs with physics-informed loss functions, the PIHGNN offers a robust and efficient solution for DC blocker placement. While currently focused on steady-state analysis, future research will aim to incorporate spatiotemporal data to account for the movement of solar storms and explore dynamic stability challenges using Reinforcement Learning.
Summary updated