DRAN: A Distribution and Relation Adaptive Network for Spatio-temporal Forecasting
Researchers have developed a new network architecture called DRAN to improve spatio-temporal forecasting, particularly for systems with changing dynamics. DRAN addresses non-stationarity by dynamically adapting to shifts in data distribution and relationships over time. Key innovations include a Spatial Factor Learner (SFL) to preserve spatial relationships during normalization and a Dynamic-Static Fusion Learner (DSFL) to integrate both changing and stable relationships. The approach has demonstrated superior performance on weather and traffic forecasting tasks compared to existing methods. AI
IMPACT Introduces a novel architecture for more accurate spatio-temporal predictions in dynamic systems.