PM2.5 forecasting under controlled sensor outages in the Beijing Multi Site archive

Abstract

Monitoring outages can change both the amount and the temporal structure of the information available to an air-quality forecaster. This study compares ten forecasting systems on the Beijing Multi-Site archive using point dropout, station-block outages and episode-dependent station-block outages at nominal rates of 10%, 30% and 50%, with 6-hour and 24-hour prediction horizons. The design contains nine outage settings per horizon and eighteen horizon-specific settings overall; learned systems share mixed-mechanism training augmentation. In the reported 24-hour aggregates, interpolation followed by a distance-graph neural network has the lowest Symmetric Mean Absolute Percentage Error (SMAPE), 0.900, compared with 0.911 for the joint mask-informed AirBridge architecture. Removing AirBridge’s observation masks increases mean SMAPE by 0.0123; differences associated with its adaptive graph and auxiliary imputation head are smaller. The pooled episode-dependent minus point-dropout difference increases from 0.0006 at 10% missingness to 0.0181 at 50%. Persistence has the lowest reported 6-hour SMAPE, 0.661. These are descriptive results under a compact training budget. Root mean squared error gives a different ordering, and shared augmentation without an unaugmented control cannot establish its causal contribution. The comparison motivates explicit reporting of masks, imputation, outage generation and forecast horizon, with replication and statistical verification needed before operational recommendations.