Forecast bias correction through model integration: A dynamical wholesale approach

Abstract

Unlike the retail-like (for selected variables) statistical post-processing methods, a wholesale-like (for all variables) dynamical approach is proposed to correct forecast bias during model integration. Subtracting a bias tendency from the model total tendency is intended to de-bias all variables at once to better (i.e. more dynamically consistent) couple with downstream applications. Three experiments were tested using an ensemble prediction system since the method is intended for an ensemble model. The verification was carried out over China for a period of 31 days (1–31 July 2015). The verification of 500 hPa temperature indicates that all three experiments have significantly improved the raw ensemble forecasts with reduced bias error, a more accurate ensemble mean, a better spread-skill relationship, and more reliable and sharper probabilities. The performance is better than or comparable to the current operational statistical method. When the verification was expanded to include more variables, a summary scorecard shows that the three experiments also had a general positive or neutral impact on both upper-air and surface variables, especially the height and temperature fields. Precipitation forecasts remained relatively unchanged. There were only a few categories that were degraded. The comparison between the three experiments yielded a mixed result: the most sophisticated approach often performed the best for 500 hPa temperature, while the simplest approach worked the best when verifying a mixture of variables. The degradation of the wind forecasts by the third experiment was discussed. These are the two challenges: how to accurately describe the bias tendency and how to add internally coherent bias tendencies to multiple variables. Given its advantages, this approach could be a promising approach for correcting biases in a numerical model.