Search papers, labs, and topics across Lattice.
The paper introduces the Multi-View Geo-Wind Guided KAN (MVG-KAN) model for short-term PM$_{2.5}$ forecasting, addressing the limitations of existing spatio-temporal methods that inadequately capture the complex interplay of meteorological factors and human activities. By integrating local periodic patterns, station-specific dynamics, and a Geo-Wind Graph that accounts for wind direction and speed, MVG-KAN effectively models pollutant dispersion and enhances forecasting accuracy. The results demonstrate significant improvements in prediction performance compared to traditional methods, highlighting the model's capability to better represent the heterogeneous factors influencing PM$_{2.5}$ levels.
MVG-KAN achieves superior PM$_{2.5}$ forecasting by leveraging a novel Geo-Wind Graph that captures wind-driven pollutant transport dynamics.
Accurate short-term PM$_{2.5}$ forecasting is important for public health protection, air-quality early warning, and urban environmental management. However, PM$_{2.5}$ variation is driven by multiple coupled factors, including stable periodic changes induced by human activities and meteorological regularity, station-specific short-term concentration evolution, and meteorology-driven pollutant dispersion among monitoring stations. Existing spatio-temporal forecasting methods may capture station relationships to some extent, but distance-only, correlation-based, or purely adaptive graphs are often insufficient to comprehensively represent these heterogeneous factors, especially wind-direction-dependent pollutant transport. To address this problem, we propose a Multi-View Geo-Wind Guided KAN model for PM$_{2.5}$ forecasting, named \textbf{MVG-KAN}, which models station-level PM$_{2.5}$ evolution from three complementary views: local periodic regularity, station-wise residual temporal dynamics, and meteorological-environment-guided spatial dispersion. Specifically, the periodic-residual forecasting backbone first separates stable daily and weekly patterns from non-periodic residual variations. A Geo-Wind Graph is constructed by combining geographic distance decay with wind-direction- and wind-speed-aware transport, providing a lightweight physically motivated directed spatial prior for residual propagation among stations. In addition, a temporal Kolmogorov-Arnold network (TKAN) residual head is then introduced to learn station-wise nonlinear autoregressive correction from de-periodized PM$_{2.5}$ residuals and historical multi-pollutant sequences, thereby enhancing the modeling of local residual inertia and pollutant co-variation.
