Science Inventory

Improving surface PM2.5 forecasts in the U.S. using an ensemble of chemical transport model outputs, part I: bias correction with surface observations in non-rural areas

Citation:

Spero, T., B. Murphy, H. Huanxin Zhang1,2, Jun Wang1,2, Lorena Castro García1,2, Cui Ge, J. Wang, L. Castro García, C. Ge, AND T. Plessel. Improving surface PM2.5 forecasts in the U.S. using an ensemble of chemical transport model outputs, part I: bias correction with surface observations in non-rural areas. JOURNAL OF GEOPHYSICAL RESEARCH: ATMOSPHERES. American Geophysical Union, Washington, DC, 125(14):e2019JD032293, (2020). https://doi.org/10.1029/2019JD032293

Impact/Purpose:

Air quality forecasting plays an important role in informing the general public and decision-makers on reducing exposure to air pollution. Air quality models simulating atmospheric constituents such as particulate matter with a diameter less than 2.5 µm (PM2.5) are often used to provide daily forecasts. However, these models are subject to large error and uncertainty as a result of the simplified representation of the real atmosphere. Here, we develop a computationally efficient framework to improve model forecasts by performing bias correction on model outputs. We focus on non-rural areas in the continental United States and show that our technique improves model forecasts of surface PM2.5.

Description:

This work is the first of a two-part study that aims to develop a computationally efficient bias-correction framework to improve surface PM2.5 forecasts in the U.S. Here, an ensemble-based Kalman Filter (KF) technique is developed primarily for non-rural areas with approximately 500 surface observation sites for PM2.5, and applied to three (GEOS-Chem, WRF-Chem and WRF-CMAQ) chemical transport model (CTM) outputs for June 2012. While all CTMs underestimate daily surface PM2.5 mass concentration by 20-50%, KF correction is effective for improving each CTM forecast. Subsequently, two ensemble methods are formulated: 1) the arithmetic mean ensemble (AME) that equally weights each model and 2) the optimized ensemble (OPE) that calculates the individual model weights by minimizing the least-square errors. While the OPE shows superior performance than the AME, the combination of either the AME or the OPE with a KF performs better than the OPE alone, indicating the effectiveness of the KF technique. Overall, the combination of a KF with the OPE shows the best results. Lastly, the Successive Correction Method (SCM) was applied to spread the bias correction from model grids with surface PM2.5 observations to the grids lacking ground observations by using a radius of influence of 125 km, which further improves the prediction of surface PM2.5 at the national scale. Our findings provide the foundation for the second part of this study that uses satellite-based aerosol optical depth (AOD) products to further improve the prediction of surface PM2.5 in rural areas by performing statistical analysis of model output.