An Evaluation of Baseline-Adjusted Ensemble Hourly PM2.5 Forecast for 2023 Wildfire Season

Abstract

PM2.5 is a harmful pollutant linked to many illnesses. Accurately predicting elevated PM2.5 levels remains challenging, and many forecasts may underestimate PM2.5 during wildfires. This research incorporates a baseline, assuming an AirNow observation at 00:00 UTC remains constant for 24 hours, into hourly forecasts by applying the predicted intraday changes from 00:00 UTC of multiple existing forecasting systems to the baseline. The forecast systems used were the Copernicus Atmosphere Monitoring Service Global Atmospheric Composition Forecasts (CAMS), the National Air Quality Forecast Capability AQMv6, and the Goddard Earth Observing System Composition Forecast v1. This research first tested equal weighting of the three systems, then a dynamic weighting updated once per day based on the previous day’s forecast accuracy measured by RMSE. The latter model achieved an RMSE of 13.49 μg/m³ on all days, a 12.32% improvement over the baseline, and an 18.63% improvement over CAMS, the most accurate forecasting system by RMSE. Mean Excess Proxy Outdoor Exposure (MEPOE) measures the additional PM2.5 exposure incurred when a person goes outside based on a forecast rather than the actual lowest-pollution hour. The dynamic weighting achieved MEPOE of 3.61 μg/m³ on all days and 8.86 μg/m³ on days with an hourly PM2.5 peak over 35 μg/m³, outperforming the baseline by 49.80% and 62.91%, respectively, and CAMS by 7.58% and 11.31%, respectively. This model demonstrates a potential approach that may improve PM2.5 forecasting during wildfire season.