MVG-KAN Improves PM2.5 Forecasting with Geo-Wind Guidance

Accurate short-term forecasting of PM2.5 concentrations is crucial for public health, air quality warnings, and urban environmental management. However, PM2.5 variations are complex, influenced by human activities, meteorological patterns, local concentration evolution, and wind-driven pollutant dispersion. Existing spatio-temporal forecasting methods often fall short in comprehensively representing these heterogeneous factors, particularly wind-direction-dependent transport. To address this, researchers propose MVG-KAN (Multi-View Geo-Wind Guided KAN), a novel model for PM2.5 forecasting. MVG-KAN models station-level PM2.5 evolution from three complementary perspectives: local periodic regularity, station-wise residual temporal dynamics, and meteorological-environment-guided spatial dispersion. It features a periodic-residual forecasting backbone that separates stable daily/weekly patterns from non-periodic variations. A key innovation is the Geo-Wind Graph, which combines geographic distance decay with wind-direction and wind-speed-aware transport to provide a physically motivated spatial prior for residual propagation. Additionally, a Temporal Kolmogorov-Arnold Network (TKAN) residual head learns station-wise nonlinear autoregressive corrections, enhancing the modeling of local inertia and pollutant co-variation.