Bias adjustment techniques for improving ozone air quality forecasts

In this study, we apply two bias adjustment techniques to help improve forecast accuracy by postprocessing air quality forecast model outputs. These techniques are applied to modeled ozone (O-3) forecasts over the continental United States during the summer of 2005. The first technique, referred to as the hybrid forecast (HF), combines the most recent observed ozone with model-predicted ozone tendency for the subsequent forecast time period. The second technique applied here is the Kalman filter (KF), which is a recursive, linear, and adaptive method that takes into account the temporal variation of forecast errors at a specific location. Two modifications to the Kalman filter are investigated. A key parameter in the KF approach is the error ratio, which determines the relative weighting of observed and forecast ozone values. This parameter is optimized to improve the prediction of ozone based on time series data at individual monitoring sites in the Aerometric Information Retrieval Now network. The optimal error ratios inherent in the KF algorithm implementation are found to vary across space; however, comparisons of the resultant KF-adjusted forecasts using a single fixed value of this parameter with those using the optimal values determined for each individual site reveal similar results, suggesting that the uncertainty in the estimation of this parameter does not have a significant impact on the final bias-adjusted predictions. The KF postprocessing is also combined with the Kolmogorov-Zurbenko filter, which extracts the intraday variability in the observed ozone time series; the results indicate a significant improvement in the ozone forecasts at locations in the Pacific coast region, but not as much at locations across the rest of the continental U. S. domain. Both HF and KF bias adjustment techniques help significantly reduce the systematic errors in ozone forecasts. For most of the global performance metrics examined, the KF approach performed better than the HF method. For the given model applications, both methods are effective in reducing biases at low ozone mixing ratio levels, but not as well at the high mixing ratio levels. This is due in part to the fact that high ambient ozone levels occur much less frequently than low to moderate levels for which the current model exhibits a systematic high bias. Additionally, the 12 km model grid structure is often unable to adequately capture the magnitude of peak O3 levels. Thus, these extreme events at discrete monitor locations are more difficult to predict.