Dependence of Model Performance on Vertical Resolution for Operational Prediction

In view of the growing expectation in the fine weather service, the horizontal resolution of numerical weather model is higher. Simultaneously, the vertical resolution is supposed to increase in coordination with the horizontal resolution in order to achieve the optimal effects. It is of great importance to match the horizontal resolution with appropriate vertical resolution to attain the best forecasting accuracy in the conditions of limited computation resources. on account of that excessively thick layers would enlarge the effects of topography leading to false gravity wave and unstable integration and excessively thin layers need smaller time-step which consume much more extensive calculation.

The effects of varying vertical resolution on the explicit simulation of Meiyu Season (2016) for 31 days are studied using the Weather Research Model (WRF) with the finest grid size of 3 km. Meiyu period in 2016 is very special owing to it’s high frequency, wide spectrum and rich types of rainfall. The experiments designed mainly include the one with vertical resolution of 51 levels which is the operational configuration currently in use and the other with vertical resolution of 42 levels which primarily reduces the levels almost by half above the height of 9 km with the rest layers remaining unchanged.

It is shown that decreasing the upper-level vertical resolution has little impact on the effects on the forecasts of precipitation and temperature. In general, reducing vertical resolution tends to produce a little higher scores of precipitation. Specifically, 42 levels show better performance in both Threat Score and Equitable Threat Score over different rainfall intensity of cumulative precipitation of 24 h except for light rain and produce higher correlation coefficient and lower root mean squares compared with that of 51-level design. Besides, as for temperature prediction It is found that, 51 levels show slightly higher accuracy with errors within 1℃ for both daily maximum and minimum temperature at 2 m height whereas the difference between two designs is less than 5 thousandths.

Note that there exit some systematic bias in temperature prediction specifically with warm bias of 1.35 ℃ in daily maximum T2m and cold bias of 0.7 ℃ for daily maximum and minimum T2m respectively, which maybe the crucial reason for generally lower prediction accuracy of daily maximum T2m than that of daily minimum T2m. Therefore, physical processes of radiation, cumulus parameterization and boundary layer are supposed to be responsible for the temperature bias and should be the important work to be focused on in the next stage.

It is concluded that both designs of vertical resolutions are desirable to model realistically driven weather systems and the rainfall intensity and distribution structures. Especially, the use of 42 levels vertical resolution is able to grasp the key process like convergence of water vapor and latent heating as well as the original design of 51 levels and even give rise to some advantage on the aspect of 24h cumulative precipitation, which is mainly attributed to the reservation of the vertical levels at lower height against 51-level design. In fact, 42 levels design is a more economical choice for operational running for the sake of both less computation cost by 17.4% and smaller data storage demand by 15.7% against the 51 level design.