Analog–Kalman Filter−Based Post-Processing of Surface Ozone from the CMAQ Model

A Kalman filter- Analog (KFAN) method has been developed and tested for post-processing surface PM_2.5 predictions from the NOAA Community Multiscale Air Quality (CMAQ) forecasting system used in the National Air Quality Forecasting Capability (NAQFC). The KFAN method has been used since 2015 by the NOAA/NWS/National Centers for Environmental Prediction (NCEP) and is now part of their operational PM_2.5 forecasting system. More recently a similar KFAN method has been tested for reducing systematic errors in ozone forecasts generated by the operational CMAQ forecasting system, and we will report on this new development.

The post-processing scheme first creates a set of best-matching analog ozone forecasts, formed as the weighted average of the ozone observations that corresponded to the N closest historical analog forecasts to the current CMAQ prediction, and then applies the Kalman filter to the weighted analog ensemble mean (KFAN bias correction). Both analog search and Kalman filter runs use the same training period, which normally spans many months and for best results should include the same season as the forecasting period. Several model parameters (called analog predictors) are used in the search for analogs, including ozone itself, ozone precursors NOx and NOy, and meteorological predictors including wind speed and direction, surface temperature, surface solar radiation, and planetary boundary layer height. An optimal weighting scheme is applied to the analog predictors. The ensemble means of the selected analogs at each observation site are used to compute site-dependent bias corrections for each ozone forecast. These corrections are then spread across the model domain by an iterative objective analysis technique and used to compute a grid-point specific correction applied to the current forecast, resulting in a new bias-corrected forecast over a 2D grid.

The performance of the post-processed model forecasts of ozone is analyzed as a function of geographic location, season, and time of day. In addition, the spatial and seasonal variation of the predictor weights will be discussed.