Enhancing Air Quality Forecasts Across the Contiguous United States (CONUS) During Wildfires Using Analog-Based Post-Processing Methods

Abstract

With our growing understanding of the risks of air pollution to human health, air quality forecasting has become a very important tool to enable decision-makers to take preventive and corrective measures for current and future policies. In addition, accurate predictions of air quality can help predict and mitigate the impacts of wildfires on human health, which have an increased risk due to anthropogenic climate change. However, errors in air quality forecasts limit their value in decision-making processes. Thus, increasing the accuracy of air quality forecasts is of significant importance. In this study, we have utilized the Community Multiscale Air Quality (CMAQ) modeling system with a 12 km horizontal grid resolution to generate daily 48-h fine particulate matter (PM2.5) forecasts for the Contiguous United States (CONUS) domain for June 1st through September 29th (major wildfire season) during 2015–2021. We conduct CMAQ offline simulations using meteorological inputs generated by the NOAA’s Unified Forecast System (UFS) numerical weather prediction model. We analyze the performance of the CMAQ PM2.5 and UFS meteorological forecasts over seven years of simulations for Environmental Protection Agency (EPA)-defined ten regions using the Air Quality System (AQS) ambient air pollution data from over a thousand monitoring sites across the CONUS. We have found that on average, the CMAQ model performs better in the eastern CONUS with the lowest Root Mean Square Error (RMSE) (10–30 μg/m3) while in the west, where wildfires are prevalent, the model has the highest RMSE of up to 120 μg/m3. To address these substantial model errors, we advance the field by applying Analog Ensemble (AnEn), a state-of-the-art statistical-dynamical method, to enhance PM2.5 forecast accuracy, with a particular focus on wildfire events, when producing an accurate forecast is highly challenging. We further refine this approach by incorporating a Carbon Monoxide-FIRE (CO-FIRE) tracer to precisely monitor fire smoke. In parallel to our objective and for the first time, we introduce two novel predictors, i.e., a CO−FIRE tracer in the model, and the maximum observed PM2.5 concentrations from the previous day (PM2.5−preday−max), and evaluate their potential significance in improving forecast accuracy, which aligns with the primary goal of our study. Despite the challenges of using AnEn for wildfires, we demonstrate that it has the potential to improve the CMAQ model forecast over the CONUS. We find that AnEn decreases the model RMSE by up to 25%, including additional 7% and 15% reduction by CO−FIRE and PM2.5−preday−max predictors, respectively, at different forecast lead times. In addition, the correlation between AnEn forecasts and observations is 20%–40% higher than that between CMAQ and observations. The Mean Bias Error (MBE) for AnEn forecasts is consistent and approximately −0.5 μg/m3 whereas CMAQ MBE varies between −1 and +1 μg/m3 between 0–48 forecast hours. AnEn significantly improves the PM2.5 forecast results during its highest episodes. During the initial days of wildfires, AnEn performs similarly to CMAQ. However, it soon catches up and decreases the error significantly.