C-DNR framework cuts power grid prediction error by 73% using causal graphs

A team of researchers from multiple institutions, led by Ibrahim Shahbaz, has developed a novel framework called Causal Dynamic Network Representation (C-DNR) for online transient stability trajectory prediction in power systems. Traditional GNN-based approaches either rely on fixed admittance-based topologies that miss state-dependent coupling, or purely data-driven methods that ignore directional causal influences. C-DNR addresses this by fusing two complementary representations of inter-generator interactions: a dynamic structural graph inferred from real-time measurements, and a directional causal graph obtained via nonlinear causal discovery—specifically using the Peter–Clark Momentary Conditional Independence (PCMCI) algorithm, which isolates directional dependencies from misleading oscillatory correlations.

The fused graph is propagated through a Gated Recurrent Unit (GRU) to predict post-fault trajectories. Evaluated on the IEEE 39-bus test system, C-DNR reduces autoregressive prediction error by 73% compared to a dynamic structural baseline. Additionally, the learned edge-wise fusion weights offer interpretable diagnostics that align with the electrical topology, giving operators transparent, pairwise insight into why certain generator couplings are prioritized. The paper will be presented at IEEE SmartGridComm 2026, highlighting a practical step toward more resilient grid operations using causal AI.