Post-Processing of CMAQ forecast for Improving Air Quality Predictions

We will show some preliminary results of our NOAA Joint Technology Transfer Initiative (JTTI) 2-year project with the goal of applying a machine learning (ML) post-processing to improve the Community Multi-scale Air Quality (CMAQ) model operational air quality forecasts issued over the US by National Air Quality Forecasting Capability (NAQFC) at NOAA/NCEP. Specifically, we have tested an extension of the analog ensemble (AnEn) model currently implemented at the NAQFC from point-based to 2D gridded predictions.The AnEn utilizes a training dataset comprising predictions from CMAQ and corresponding observations of the quantity to be predicted (i.e., O3 or PM2.5) to generate future ensemble predictions based on past observations. The ensemble is constructed for a given deterministic CMAQ forecast by collecting past observations corresponding to the best matching past CMAQ forecasts (called analogs) to the current CMAQ prediction.We have conducted a preliminary application of the AnEn to reduce the errors of CMAQ PM2.5 and ozone surface gridded concentrations using a combination of past gridded chemical reanalysis from the Copernicus Atmosphere Monitoring Service (CAMS) Near-Real-Time model with measurements from AirNow stations. The analog method requires a continuous training dataset of hourly values of observed chemical concentrations obtained by merging the CAMS surface PM2.5 and ozone fields with the respective observations from the AirNow network using the Satellite-Enhanced Data Interpolation technique (SEDI) (Dinku et al. 2015). SEDI removes the bias from CAMS analysis and short-term forecast fields while preserving the AirNow-measured values at the station locations.We will first show validation of the SEDI bias-corrected CAMS concentrations against AirNow PM2.5 and ozone-measured concentrations from stations not used in the SEDI correction process. Then, we will verify the performance of the whole forecasting system in some regions of the contiguous United States in the 0-72 hours lead time range.