When Monitoring Channels Fail—Next-Hour PM2.5 Forecasting across Unseen Beijing Stations

Operational air-quality forecasts can lose local pollutant, neighboring-station, meteorological, or all current signals. Using public hourly data from 12 Beijing stations, this study stress-tested a persistence-corrected next-hour PM2.5 model in a leave-one-station-out, future-year design under eight matched channel-availability scenarios. On test rows with observed local PM2.5, residual gradient boosting improved equal-station MAE from 9.67 µg/m³ for persistence to 9.13 and improved all 12 held-out stations. Across all clean test rows its MAE was 9.39. Meteorological loss changed MAE only modestly (9.46–9.48), whereas masking local PM2.5 raised it to 11.51, masking neighboring PM2.5 to 33.28, joint PM loss to 38.72, and 24-hour all-signal blocks to 45.77. Coverage fell from 0.910 clean to 0.734 under joint PM loss, while interval score rose from 81.91 to 610.73. An availability-aware fallback limited MAE to 18.32 under joint PM loss and 17.64 under all-signal blocks. Robustness depended on which channels failed and on the fallback used, not merely the missing fraction.