On Use Of LHN Method To Assimilate The Surface Precipitations For GRAPES_Meso Model Initialization

With the finer and finer resolution of numerical weather prediction(NWP) models, the traditional observations fail to provide the small-scale features for the high-resolution model initial conditions. One important way to meet this demand is to apply the radar and surface observations, such as the surface intensified precipitation data, with high spatial and temporal resolutions into the initialization procedures. There is a large amount of intensified surface precipitation data with very high temporal and spatial resolutions in our country provided by existing surface observed network. However, it hasn’t been quantitatively used in NWP models. Therefore, how to assimilate the surface intensified precipitation data to enhance the initialization and further improve the quantitative precipitation forecasts(QPF) will be studied in this thesis.Based on the high-resolution Global/Regional Assimilation PrEdiction System(GRAPES) model, some revisions have been made to the traditional latent heat nudging(LHN) scheme in GRAPES model, to assimilate the intensified surface precipitation data more effectively. The techniques include:(1) the “Warm-Latent Heat Nudging”(W-LHN) scheme is proposed to reduce the initial spin-up time;(2) the method of multi-scale initial analysis is proposed to improve the short-memory problem of initial conditions;(3) the probability matching(PM) method is used to further improve the QPF in very-short-range.Firstly, for the spin-up problem in initial integration, especially the time lagged problem in the prediction of rain belt, a group of LHN experiments were conducted to find the difference between the traditional “Cold-Latent Heat Nudging”(C-LHN) and the revised W-LHN methods. Preliminary conclusions were as follows:(1) the distribution of initial temperature, specific humidity and wind fields were modified reasonably by adjusting the model latent heating profiles, and thus the convective instability in the heavy rain area was increased;(2) compared with the C-LHN method, the revised W-LHN method produced more warm and moist air in the middle and low atmosphere, and was better to generate more precipitation at the initial time;(3) compared with the traditional C-LHN method, the W-LHN experiments achieved more favorable simulation of rainfall location and intensity. The disadvantage lies in that the rainfall intensity was a little over-estimated;(4) when comparing the results of different W-LHN methods, conclusions are that the W6-LHN precipitation results agreed more closely with the observations, while the W12-LHN precipitation was remarkably over-estimated. However, the limited impacts of the assimilation of precipitation data based on LHN methods were within 6 hours.Secondly, the initial “triggering” uncertainties and the sensitivities of the very-short-range QPF to multi-scale initial conditions are considered. Based on the coupling of LHN and the three-dimensional varational data assimilation(3DVAR), five experiments are designed in this thesis, such as 3DVAR method, the traditional LHN method(VAR0LHN3), the cycling LHN method(CYCLING), the spatial filtering(SS) and the temporal filtering(DFI) LHN methods. The numerical simulation results showed that:(1) in comparison with the control run, the CYCLING run could generate smaller-scale initial moist increments and was better at reducing the spin-up time and triggering the convection in a very-short time;(2) however, in the CYCLING fields the intensity of heavy precipitation was substantially weakened and narrowed after about 6 hours into the forecast;(3) the DFI runs could generate the initial analysis fields with relatively larger-scale initial increments and trigger the weaker convections at the beginning time(0-3h) of integration, but enhance them at latter time(6-12h). Therefore, the above five runs have their respective temporal and spatial advantages.Furthermore, the probability matching(PM) method was used to generate the QPF in very-short range by combining the precipitation forecasts of the five runs. PM method is a method working by defining the probability distribution function(PDF) of data with less accuracy from individual members to match the more accurate data, thus different prediction results with different temporal and spatial properties derived from multi-scale initial perturbations can be blended. The experiments for one month showed that: by combining the five runs with different convection triggering features, the PM forecasting skills were higher than any single run, the TS scores were higher and the Bias scores were closer to 1; in addition, the forecasting errors of the monthly averaged 6h accumulated precipitation derived from PM method were smaller than the control run.Therefore, to reduce precipitation uncertainties caused by the initial uncertainties, it is feasible to establish a convective-scale ensemble prediction system by means of PM method. In addition, the QPF with a small size of combining members proposed here is quite prospective in operation for its lower computation cost and better performance.