Short-horizon PM2.5 forecasting from multi-station networks: a mask-aware, leakage-safe framework

Accurate short-horizon forecasts of fine particulate matter (PM₂.₅) at station resolution are critical for advisories, clinical preparedness, and episodic controls. We offer a framework for multi-station PM₂.₅ forecasting over operational hourly networks that is safe against leaks and cognizant of masks. The pipeline puts stations on the same time axis, fills in only small gaps with linear interpolation, records a label mask where PM₂.₅ is actually seen, and limits outlier capping and normalization to the training split with per-station scalers. We make 12-h input windows from the cleaned tensors and estimate targets for all stations at the same time, one hour in the future. We compare a persistence baseline, per-station Random Forest and XGBoost, per-station LSTM/GRU, a multi-station RNN that shares information across sites, and spatio-temporal graph models (ST-GCN and a Chebyshev-GCN followed by LSTM) using a distance-informed station adjacency. Neural models are trained using masked MSE, and predictions are scaled back to their original values for each station before being tested. The GCN-LSTM model gets RMSE 22.90 μg/m³, MAE 12.29 μg/m³, R² 0.545, and CPCB band accuracy 64.6% on held-out periods. This is about 21% better than ST-GCN (29.03 μg/m³) and about 17% better than Random Forest (mean 27.69 μg/m³). It also matches XGBoost on RMSE (22.83 μg/m³, Δ≈0.3%) while improving MAE and band accuracy. The multi-station GRU also does better than the LSTM (RMSE 23.37 vs. 24.35 μg/m³). The framework is model-agnostic, reproducible, and deployable with only station CSVs and site coordinates. The framework requires simply station CSVs and site coordinates to be deployable, reproducible, and model-agnostic.