OmniPMNet fuses station and grid forecasts, cuts PM10 error by 30%
New neural process model beats GNNs and CAMS while producing continuous spatial fields.
OmniPMNet tackles a persistent challenge in air-quality forecasting: reconciling point-based station predictions with continuous gridded model outputs. Chemical transport models like CAMS provide broad spatial coverage but suffer from local biases, while graph neural networks excel at station-specific accuracy but cannot produce gridded fields. The new model uses a Convolutional Conditional Neural Process (ConvCNP) backbone with a terrain-aware Gaussian set convolution to lift irregular station forecasts onto a regular grid. A multi-scale Spatial Source Attention module then blends those station-derived features with CAMS forecasts. An omni-query readout decodes this unified representation into consistent PM10 predictions at either stations or grid cells over a 108-hour horizon.
Evaluated across all 1,618 air-quality monitoring stations in China for the full year of 2024, OmniPMNet matches the station-level accuracy of the stronger GNN baseline (mean absolute error 21.14 vs 22.00 µg/m³) while slashing CAMS error by 30%. Its largest gains occur in high-concentration scenarios: the 90th-percentile MAE drops 9% relative to the GNN and 25% relative to CAMS. During severe dust storms, the model improves categorical detection skill and accurately tracks evolving spatial trajectories. By delivering both discrete and gridded forecasts from a single architecture, OmniPMNet offers a practical pathway for operational air-quality monitoring systems.
- OmniPMNet uses a ConvCNP-based fusion architecture to unify discrete station forecasts (from GNNs) and gridded CAMS forecasts into one shared spatial representation.
- Tested on 1,618 stations across China in 2024: matches best GNN station accuracy (MAE 21.14 vs 22.00 µg/m³) and cuts CAMS MAE by 30%.
- Especially effective in high-concentration extremes — 90th-percentile MAE down 9% vs GNN and 25% vs CAMS — and during dust storms.
Why It Matters
Enables operational air-quality systems to get both local accuracy and full spatial coverage from a single model.