Causal--Structural Dynamic Graph Learning for Online Transient Stability Trajectory Prediction in Power Systems

Ibrahim Shahbaz, Omar Al-Refai, Isaac Lagoy, Ahmad Al-Khateeb, Sathvik Sankaranarayanan, Eman Hammad

Power systems consist of dynamically coupled generators, motivating the use of Graph Neural Networks (GNNs) for online transient stability prediction. Traditional GNN frameworks are often constrained by fixed admittance-based topologies that fail to capture state-dependent coupling, or by data-driven methods that neglect directional influences. This paper proposes Causal Dynamic Network Representation (C-DNR), a novel framework that fuses two complementary representations of inter-generator interactions prior to temporal modeling: a dynamic structural graph inferred from measurements and a directional causal graph obtained via nonlinear causal discovery. An end-to-end learned edge-wise fusion mechanism adaptively weights these representations for each generator pair, and the resulting graph is propagated through a Gated Recurrent Unit (GRU) to predict post-fault trajectories. Evaluated on the IEEE 39-bus system, C-DNR reduces autoregressive prediction error by 73% compared to a dynamic structural baseline. Among the evaluated causal methods, only Peter--Clark Momentary Conditional Independence (PCMCI) achieves consistent improvements, owing to its ability to isolate directional dependencies from misleading oscillatory correlations. The learned fusion weights further provide interpretable diagnostics aligned with the electrical topology, offering transparent, pairwise insight into the prediction process.