Network System Forecasting Despite Topology Shift

Many real-world dynamical systems have an underlying network structure and evolve over time, such as transportation networks affected by road closures, or power grids undergoing reconfiguration and sensor failures. Forecasting their dynamics is challenging not only due to high dimensionality, but also because the interaction topology is often partially observed, noisy or changes after deployment. Such structural uncertainty induces a form of distributional drift that can significantly degrade predictive performance. Understanding when and how forecasting remains reliable under such structural drift is therefore essential. We study the predictability of graph time series under random topology perturbations and characterize how structural noise impacts forecasting accuracy in large-scale networks. In the limit of large networks, we identify distinct regimes ranging from stable predictability to intrinsic unpredictability under increasing topology perturbation. Motivated by this analysis, we propose a scalable forecasting framework based on a probabilistic low-dimensional representation of network dynamics. Our method leverages Bayesian coreset approximations of graph embeddings to obtain representations that are more robust to topology misspecification, and couples them with light temporal modeling for efficient prediction. Experiments on synthetic and real-world networks demonstrate competitive forecasting accuracy under structural uncertainty, while significantly reducing computational cost.