MVG-KAN: Multi-View Geo-Wind PM2.5 Forecasting (Jun 2026)

MVG-KAN, a new multi-view model for short-term PM2.5 forecasting, was submitted to arXiv on 23 Jun 2026 by Cheng Huang and eight co-authors (arXiv:2606.24347). The architecture models station-level PM2.5 using three complementary views: stable periodic regularity, station-wise residual temporal dynamics, and meteorology-guided spatial dispersion.

How does MVG-KAN work?

MVG-KAN separates predictable periodic patterns from unpredictable residuals, then models residual propagation with a wind-aware spatial prior and corrects station-level residuals with a temporal Kolmogorov-Arnold network. The model first de-periodizes PM2.5 to extract daily and weekly patterns, then applies a periodic-residual forecasting backbone to isolate non-periodic residual variations.

After de-periodization, MVG-KAN constructs a Geo-Wind Graph that combines geographic distance decay with wind-direction- and wind-speed-aware transport, providing a directed spatial prior for how residuals propagate among monitoring stations. A TKAN residual head learns station-wise nonlinear autoregressive corrections from the de-periodized PM2.5 residuals and historical multi-pollutant sequences, capturing local residual inertia and pollutant co-variation.

What components make MVG-KAN different from prior spatio-temporal methods?

MVG-KAN adds two concrete innovations to typical spatio-temporal forecasting pipelines: a physically motivated Geo-Wind Graph, and a temporal Kolmogorov-Arnold network residual head. The Geo-Wind Graph goes beyond distance-only, correlation-based, or purely adaptive graphs by encoding wind-direction and wind-speed into directed transport edges, while the TKAN head specifically targets station-level nonlinear autoregressive corrections using multi-pollutant histories.

The paper frames these features as responses to three coupled drivers of PM2.5 variation: stable periodic changes from human activity and meteorology, station-specific short-term concentration evolution, and meteorology-driven pollutant dispersion among stations. The Geo-Wind Graph supplies the meteorology-informed spatial prior, and the TKAN models station residual dynamics after periodic effects are removed.

Why it matters

Short-term PM2.5 forecasts power public health alerts and urban environmental management. MVG-KAN directly addresses two common limitations in air-quality forecasting: insufficient modeling of wind-dependent transport and weak separation between periodic and residual dynamics. By combining a de-periodization step, a Geo-Wind Graph that explicitly uses wind direction and speed, and a TKAN residual learner, the model targets both regional dispersion and local station inertia, which should improve actionable short-term predictions.

What the submission shows now

The submission metadata identifies the work as arXiv:2606.24347 and gives the upload date as 23 Jun 2026. The submission file size recorded in the archive is 4,949 KB. The author list begins with Cheng Huang and includes Muyao Guan, Jairus Yougui Railey, Ning Xu, Honghui Xu, Changjiang Zhang, Zhen Zhang, Shiqing Zhang, and Cong Bai.

What to watch

Follow whether the authors release code or evaluation results linked from the paper, and whether the Geo-Wind Graph obtains measurable gains over distance-only and adaptive graphs in held-out station forecasts. The next concrete signals will be a code or dataset release cited in the paper or a public benchmark comparing MVG-KAN to existing spatio-temporal baselines.