Statistical Post-processing Research Of Regional Precipitation Ensemble Forecast In Fujian Based On GBDT-BMA Method

Precipitation is a meteorological element formed under the influence of a series of complex physical processes and affected by a variety of weather systems.Therefore,improving its prediction accuracy is the focus and difficulty of meteorological business and scientific research.In this paper,Bayesian model averaging(BMA)method is applied to research on statistical post-processing of precipitation ensemble forecast over Fujian and its surrounding areas based on CMAP hourly precipitation fusion data and 24-hour cumulative precipitation ensemble forecast composed of six models in different forecast periods provided by THORPEX Interactive Grand Global Ensemble(TIGGE).Aiming at the different structural modifications of BMA method,we study its impact on the accuracy of precipitation prediction,and optimize the BMA prediction scheme.Then,based on the optimization scheme,the GBDT-BMA precipitation classification prediction model is proposed to carry out the research on classification precipitation prediction.Further,based on GBDT-BMA model,we conduct precipitation prediction experiments in each forecast period,and test and evaluate the prediction results.This research has farreaching significance for improving the accuracy of precipitation ensemble forecasting and the operational application of precipitation forecasting.(1)Based on the selection of different parameter estimation methods(regional and grid scale modeling),the selection of different distributions(normal distribution and gamma distribution)and the application of different data conversion methods(logarithmic transformation and power transformation),we carried out the experimental design of BMA method.On this basis,we carry out the research on the influence of BMA structure modification on precipitation prediction results,and optimize BMA scheme from the perspective of certainty and probability prediction and hierarchical precipitation prediction accuracy.The results show that:The BMA scheme based on the parameter estimation method of grid scale modeling has better prediction effect;BMA weight can reflect the forecast ability of each model to a certain extent,and the poor model has a lower weight in BMA;the selection of gamma distribution can bring better clear-rainy forecast and light rain forecast effect,and the forecast accuracy is nearly 50% higher than that of the original model;BMA-D2-T2 scheme with 1/3 power data conversion and gamma distribution is the best scheme.(2)In order to improve the precipitation forecast ability of moderate rain and heavy rain,We introduce gradient boosting decision tree(GBDT)and classified BMA method,propose GBDT-BMA classified precipitation prediction model based on the optimal BMA scheme,and test it at three representative stations in Fujian.The results show that: The prediction accuracy of the model can be further improved by selecting the prediction result of 0.52 quantile in the heavy rain prediction;compared with the optimal model,the moderate rainfall forecast accuracy of the model has been significantly improved,the TS of moderate rain at Fuzhou Station has increased from 0.55 to 0.82,gaining an improvement of 50%;at the same time,the heavy rain forecast ability of each representative station is the same as or slightly improved from the best model.(3)We extended the GBDT-BMA model to Fujian and its surrounding areas,carried out 24 h cumulative precipitation prediction experiments in different forecast periods,and compared and evaluated the prediction results of different schemes from the perspectives of certainty and probability prediction,hierarchical precipitation prediction and spatial distribution of precipitation prediction.The results show that: The forecast effect of each scheme decreases slightly with the increase of forecast period;The TS of moderate rain in GBDT-BMA precipitation forecast is more than 70% higher than that of the original model,and the forecast ability of heavy rain is also slightly enhanced;GBDT-BMA can more accurately depict the scope of the main rain belt and the area without rain,and has better prediction performance.