Spatiotemporal forecasting of PM2.5 from Aral Sea dust using a physics-guided CNN-BiLSTM-attention model

Fine particulate matter (PM2.5) from the Aralkum Desert poses serious healthrisks to populations in Kazakhstan, Uzbekistan, and Turkmenistan, where dustplumes routinely exceed the daily limits set by the WHO, yet no early warningsystem exists. The chemical toxicity of Aralkum dust, combined with the nearabsence of surface air quality monitoring in Central Asia, indicates that accuratemulti-day PM2.5 forecasting in the region is not only critically important but alsofundamentally more difficult than in regions with established monitoring networks.In this work, we introduce a Physics-Guided CNN–BiLSTM with Multi-Head Attention (CNN–BiLSTM–MHA) model trained on meteorological and physically deriveddatasets to produce three-day-ahead Health Index forecasts based on PM2.5 concentrations over a 39 × 49 spatial grid covering the Aral Sea region. We evaluateour model against a Transformer model and a persistence baseline, as the proposedmodel attains R2 = 0.676 on Day +1 and R2 = 0.341 on Day +3, outperformingthe Transformer while using 13 percent fewer parameters. The model’s relativeimprovement over persistence increases from R2 = +0.041 on Day +1 to +0.526 onDay +3, confirming that physics-informed features improve generalization at longerforecast horizons where purely statistical models fail. Ablation experiments showthat removing ERA5 data reduces R2 from 0.502 to 0.053, justifying the physicsguided feature. Spatial attention weights concentrate on the dust source region andmajor transport corridors, consistent with known dust climatology. A regionallycalibrated AQI classification layer produces zero misclassifications between Goodand Hazardous categories across all lead times, providing an immediately usefuldecision-support tool for regional health authorities and early warning systems