OmniPM-Net Improves PM10 Air Quality Forecasts with Fusion Model.

Researchers have developed OmniPM-Net, a novel Convolutional Conditional Neural Process (ConvCNP)-based fusion model designed to enhance PM10 particulate matter forecasting. This model addresses the challenge of integrating both precise station-level measurements and broad gridded forecasts, which are typically provided by chemical transport models (CTMs). OmniPM-Net works by lifting irregular graph neural network (GNN) station forecasts onto a regular grid using a terrain-aware Gaussian set convolution. It then blends these with CAMS (Copernicus Atmosphere Monitoring Service) forecasts via a multi-scale Spatial Source Attention module. A shared omni-query readout then decodes this combined representation into consistent PM10 predictions for both stations and grid cells over a 108-hour horizon. Evaluated across 1,618 stations in China, OmniPM-Net matched the accuracy of strong GNN baselines at the station level and reduced CAMS mean absolute error by 30%. Its most significant improvements were observed in high-concentration scenarios and during dust episodes, demonstrating enhanced detection skill and better tracking of evolving spatial trajectories.