Improved precipitation forecast by correcting phase and intensity errors of a meso-scale numerical model

Radar-based nowcasting utilizes observations in an optimal way to generate accurate precipitation forecast. Recent advance in data assimilation techniques facilitates the use of observations in numerical models and leads to the improved precipitation forecast. However, they still suffer from phases and intensity errors, possibly due to imperfect parameterization of various physics. Small phase errors can easily undermine the use of the traditional point-by-point skill scores. Thus, the separate evaluation of phase and intensity errors is extremely important to better represent the model performance. In addition, their consistency should be evaluated to better understand dynamics and parameterization in numerical models. In this work, we demonstrate the systematic evaluation of model errors (phase and intensity errors) from a meso-scale model and their consistency in space and time. Furthermore, we apply a correction algorithm to improve the accuracy of model precipitation forecast. This correction algorithm assumes the consistency of model errors in time and utilizes nowcasting techniques to advect model errors into the future.

A meso-scale, real-time, four-dimensional data assimilation and short-term forecasting system (RTFDDA) has been built upon a high-resolution MM5 and the Newtonian Relaxation (nudging) scheme. This MM5-RTFDDA incorporates radar data to modify the latent heat and is cycled every three-hours. A field demonstration of MM5-RTFDDA was conducted around Illinois and Indiana areas during May 15 – August 31, 2006. We used precipitation forecast from 0 h to 12 hours at the temporal resolution of five minutes and at the grid spacing of 5 km by 5 km in the domain of about 1000 km by 750 km. NSSL's hybrid surface rain maps are used to quantify the model errors.

The verification shows the existence of significant phase errors even at the model initial time, indicating that MM5-RTFDDA with frequent cycling is not sufficient to correct the model background phase errors. These initial phase errors slightly increase with time. The correction of model errors significantly improves the accuracy up to 9 hours and the corrected forecast performs better than extrapolation of radar echoes. We show the various steps of correction and their performance. Furthermore, we demonstrate how model information can improve radar-based nowcasting.