🌬️ PM2.5 Pollution Forecasting

This demo visualizes precomputed predictions from saved .npy files. It does not run live model inference. Delhi and the surrounding NCR are marked on each map.

Test Window (0 – 21)

Each window is a different time period from the test set.

0 21

Forecast Hour (+1h – +16h)

How many hours ahead the model is predicting.

1 16

Input PM2.5 — Last Known Hour

Predicted PM2.5

Model: ConvLSTM encoder + Fourier Neural Operator (FNO) hybrid
Training: Kaggle T4 GPU · ANRF competition dataset · 4 months × 16 atmospheric features
Grid: 140 × 124 spatial points · WRF domain over India · Delhi / NCR highlighted Input: 10 hours of atmospheric data → Output: 16-hour PM2.5 forecast
Competition Rank: 2 · Final Score: 0.8795 (sMAPE-based)

PM2.5 guide: 0–15 Good · 15–35 Moderate · 35–55 Sensitive · 55–150 Unhealthy · 150+ Hazardous