PowerGraph: Cascading Failure Prediction

Cascading failure propagation across the IEEE-24 bus grid

Built a Graph Neural Network (GNN) pipeline on the NeurIPS 2024 PowerGraph benchmark to predict cascading blackouts in power grids. Graph Isomorphism Network with edge features (GINe) achieved 0.989 balanced accuracy and 0.999 AUROC on IEEE-24. General Powerful Scalable (GPS) (a hybrid local Message Passing Neural Network (MPNN) + global Transformer architecture) matched GINe at 0.9908 ± 0.0019 with half the seed variance, demonstrating that global attention captures long-range cascade propagation patterns local message passing misses. A GNN+XGBoost embedding ensemble reached 0.9954, the new best across all metrics. Edge features were very important and zeroing them collapses balanced accuracy to 0.746.

Directed GINe with learnable direction weighting reaches 0.9919 ± 0.0011 balanced accuracy (the best single-model result) by separately aggregating messages along and against physical power flow direction. The learned α converges to ~0.5 across all seeds, revealing that backward propagation (load → generator) carries equally useful cascade signal as the forward direction. This is the first application of directed GNNs to any power grid task.

ONNX INT8 quantisation: 53.8% smaller, 2.25× faster

Open Neural Network Exchange (ONNX) INT8 quantisation reduces the model from 173 KB to 80 KB (53.8% smaller) and achieves 0.349 ms CPU inference, 2.25× faster than FP32, well under the 100 ms real-time budget for N-1 contingency screening in energy management systems. 99.5% output agreement with the full-precision model confirms negligible accuracy loss.

GINe vs XGBoost balanced accuracy under feature noise injection

At ε = 0.05 feature noise, GINe holds at 0.926 balanced accuracy while XGBoost drops to 0.798 i.e. a 0.128 gap that widens with noise intensity. Graph topology acts as a denoising signal: neighbourhood message passing averages out corrupted sensor readings that flat-feature models cannot recover from. Targeted edge dropout (removing highest-loading lines first) degrades both models faster than random dropout, confirming the structural importance of high-load transmission paths.

Integrated Gradients explainability identifies which transmission lines are causally responsible for a predicted cascade, validated against ground-truth cascade labels from the benchmark.

Integrated Gradients edge attribution — top causal lines highlighted

Edge-level failure prediction reformulates the task from “will a cascade occur?” to “which lines will fail?”, achieving 0.9976 AUROC and 0.763 PR-AUC with network-centric features (betweenness centrality, load ratio). Near-perfect ranking confirms GINe can identify vulnerable transmission lines which is a qualitatively different capability that flat-feature models like XGBoost cannot replicate.

DNS regression with Kirchhoff physics-informed regularisation targets PowerGraph’s stated open problem (best published R² = 0.43). The result is an informative negative: multi-task training degrades classification by 0.9 pp while Kirchhoff regularisation improves regression (R² from −2.14 to −0.39 on cascading graphs) but collapses classification by 12 pp. The gap to Physics-Informed Graph Neural Jump Ordinary Differential Equations (PI-GN-JODE) R² = 0.951 confirms that DNS regression on static snapshots requires fundamentally different architectures i.e. temporal modelling and dedicated physics-informed designs rather than a simple auxiliary head.

GNNs can capture complex interdependencies in power grids, and are well suited for early warning systems that predict cascading failures before they happen, allowing crucial early interventions to prevent blackouts. The PowerGraph benchmark provides a realistic testbed for developing and evaluating these models, and my results show the effectiveness of GNNs for improving grid resilience.