Numerical Prediction Of Summer Precipitation In East Asia Based On Multi-model Integration And Dynamic Downscaling

Short-to-medium-range precipitation forecasts with the lead times of 1–7 days are an important part of the traditional operational weather forecasting,which play a significant role in issuing early warnings and assisting decision-making for the governments.In this study,to improve the prediction skill of the short-to-medium-range summer precipitation forecasts,a variety of statistical postprocessing models are constructed to provide deterministic forecasts or probabilistic forecasts in East Asia based on the daily ensemble precipitation forecast outputs from multiple prediction centers of the TIGGE datasets and the observations from the Global Precipitation Measurement.Various verifications including the traditional "point-to-point" precipitation amount assessments and the evaluations for the spatial structure characteristics of the precipitation object are carried out to examine the forecast experiments.Aiming at the "18·8" continuous heavy precipitation events occurred in Guangdong Province,the regional ensemble system COSMO EPS is used to conduct dynamic downscaling forecast experiments,and the benefit of assimilation of radar data is investigated by applying the latent heat nudging(LHN)approach to the COSMO EPS.Additionally,the ensemble forecasts of COSMO EPS are compared with the results of multi-scheme multimodel ensemble forecasting to verify their forecast ability for extreme precipitation.The main conclusions obtained are as follows:(1)With regard to the Method for Object-Based Diagnostic Evaluation(MODE),the precipitation object identification and its corresponding attributes are influenced by the convolution radius and precipitation threshold.Both observations and model forecasts show that the number of detected objects decreases with increasing convolution radius and/or precipitation threshold.For all precipitation fields,the number of identified objects also decreases as the object area increases.A larger precipitation threshold will cause the area of the identified precipitation object becomes smaller.Meantime,a larger convolution radius will result in a larger object area through spatial smoothing.The verification results of the traditional "point-to-point" precipitation amount assessments and the evaluations for the spatial structure characteristics of the precipitation objects demonstrate that different models have different forecast capabilities for different aspects of precipitation.Thus,an object-based superensemble(OBJSUP)model is innovatively proposed in this study,which determines the weight of each contributing model by the similarity of the spatial structure characteristics between the precipitation objects from forecast and observed fields.On the short-to-medium-range timescale,the OBJSUP model has better prediction skills than the raw ensemble forecasts and the traditional gridpoint-based superensemble(GPSUP)model.The GPSUP model calculates the weights based on traditional "point-to-point" error analysis.The main reason why OBJSUP model provides better forecast quality is that it has a better prediction for the centroid loaction of the precipitation object.This research would provide a new idea for the selection of the weight metric for the multimodel ensemble techniques.(2)The multimodel ensemble precipitation probabilistic forecast experiments with 1–7lead days indicate that the standard Bayesian Model Averaging(s-BMA)and the standard Ensemble Model Output Statistics(s-EMOS),as two state-of-the-art approaches,have higher forecast skills than the raw ensemble forecasts.The s-EMOS model performs generally better than the s-BMA model in East Asia,and the s-BMA model has limited skill improvement for moderate and heavy precipitation events.Correspondingly,on the basis of the standard model,the categorized models including the c-BMA model and c-EMOS model are established in this study.The 24 h accumulated precipitation amount is divided into three categories by the ensemble mean forecast(i.e.,light precipitation below 10 mm,moderate precipitation from 10 to 24.9 mm,and heavy precipitation above 25 mm).Samples of these precipitation categories are selected to establish different BMA and EMOS models and their respective parameters are estimated from the training period.Finally,the most appropriate forecast model is chosen for the forecast period according to the forecast ensemble mean.Compared with the standard models,c-BMA and c-EMOS models further improve the prediction skills of the precipitation probabilistic forecasts.The improvement of the c-BMA model relative to the s-BMA model is greater than that of the c-EMOS model compared to the s-EMOS model.This is mainly due to the relatively high prediction quality of the s-EMOS model itself.The improvement of the cBMA model decreases with the extension of the forecast lead days,but it has improved the forecast ability of multiple precipitation categories throughout East Asia,especially for the moderate and heavy precipitation events.Moreover,c-BMA model also increases the probability reliability.This study would provide meaningful technical support for the short-tomedium-range precipitation operational probabilistic forecasts.(3)The dynamic downscaling forecast experiment of the "18·8" heavy precipitation event occurred in Guangdong Province shows that compared with other models,COSMO EPS have better predictions for the rainstorms taken place in eastern Guangdong Province.However,the location of the forecasted rainstorm is eastward in comparison with the observations and the rainstorm area is relatively small.Guangdong Province is affected by the monsoon depression,southwesterly and southerly winds continue to transport water vapor from the South China Sea to Guangdong Province.The southwest monsoon low-level jet gradually advances northward over time,which promotes the occurrence and development of continuous heavy precipitation in the coastal areas of Guangdong Province.As the monsoon depression gradually moves westward and weakens,the low-level jet tends to weaken accordingly and the precipitation in the Guangdong Province also gradually disappears.Additional experiments are carried out to investigate the impact of Latent Heat Nudging(LHN)data assimilation scheme on the forecast skill of COSMO EPS.In these experiments,the rainfall rate retrieved from the radar covering Guangdong Province are analyzed in the first three hours and then LHN is switched off to provide twenty-one hours free forecasts.The results indicate that the COSMO EPS with LHN data assimilation improves the forecast quality of the area and location of the heavy precipitation,but the improvement of the rainstorm forecast is not obvious.The initial condictions and lateral boundary conditions of each ensemble member are the same,contributing to a relatively small ensemble spread of COSMO EPS.Furthermore,the COSMO EPS dynamic downscaling results are compared with multi-scheme multimodel ensemble forecasts to show that the object-based superensemble(OBJSUP)performs better for the daily heavy precipitation and spatial distributions of the accumulated precipitation over 27–31August 2018 in Guangdong Province.The categorized Ensemble Model Output Statistics(cEMOS)forecasts are much closer to the observations for the heavy precipitation occurred in eastern Guangdong than COSMO EPS,and the categorized Bayesian Model averaging(c-BMA)predicts weaker precipitation.The comparison between the c-EMOS model and the OBJSUP model shows that the probabilistic forecast model is more likely to predict heavy precipitation than the deterministic forecast model.