Spatiotemporal Forecasting
Overview
Spatiotemporal forecasting studies how patterns evolve across both space and time, supporting decisions such as routing traffic, preparing for floods, and tracking disease outbreaks. Common approaches include graph neural networks for connected locations, convolutional and Transformer models for gridded fields, and state-space or neural-operator methods for long sequences and physical dynamics. Researchers also use physics-informed constraints, multimodal data, and probabilistic or generative models to represent mechanisms and forecast uncertainty. A central challenge is making forecasts stable and calibrated over long horizons while generalizing to sparse observations, new regions, and changing conditions; realistic benchmarks often reveal weaknesses hidden by standard test splits.
Latest papers
429 papers59 papers in the last four weeks, up 48% on the four weeks before. 0.3% of all new papers.