Research On The Correction Method Of Numerical Weather Forecast Precipitation Products Based On Deep Learnin

Due to the complexity of meteorological systems,the spatial and temporal distribution of precipitation is highly uncertain,which poses a great challenge to the occurrence and prevention of natural disasters.First of all,the topography of mainland China is complex with a three-step distribution and variable climate,and there are many factors affecting precipitation changes.Satellite remote sensing technology has great advantages in obtaining precipitation information,but there are differences in the accuracy of different products in different regions.Second,numerical weather prediction is the main method for short-and medium-term precipitation prediction in China.The parameterization scheme in numerical weather prediction plays a decisive role in the prediction accuracy of precipitation products,however,the parameterization scheme involves some inherent uncertainties,and there are inevitably certain errors in practical applications.Therefore,in order to more accurately understand the spatial and temporal variation patterns of precipitation as well as to improve the accuracy of numerical prediction precipitation products the research work in this paper is divided into the following three parts:(1)In order to apply multi-source precipitation products efficiently and accurately,four sets of precipitation products(MSWEP,PERSIANN,ERA5,and TRMM)are selected as the reference benchmark in this paper to evaluate the accuracy on the daily and spatial scales(RMSE,MAE,CSI,POD,ETS,FAR,etc.)in mainland China and seven geographical subdivisions The results show that the four sets of precipitation products have been evaluated at the daily and spatial scales to provide a more accurate understanding of precipitation patterns.The evaluation results show that each of the four precipitation products has its own advantages and disadvantages and has a high detection capability in ”with or without” rain.(2)To address the problem that a single precipitation product can hardly meet the precipitation estimation in all climate zones,this paper firstly applies four machine learning algorithms(Lasso,RF,MLP,and XGBoost)for precipitation merging study of multi-source precipitation products in different geographical partitions;secondly,the optimal interpolation method is used to merge the merged precipitation products with ground observations for grid station merging,and the results show that the correlation coefficients(CC)of the four fusion methods in different geographical The correlation coefficients(CC)of the four merging methods are improved in the range of 0.682-0.739,and the precipitation near the stations of the assimilated products tends to be consistent with the ground observation data,which provides effective ”true value”data for further improving the accuracy of the precipitation products of numerical prediction.(3)To address the problem of bias between numerical weather prediction precipitation products and the actual ones,this paper firstly evaluates the importance by RF and selects different sets of physical parameters that have influence on precipitation;secondly,we construct a bias correction model DOR-CGAN based on deep learning,which is a mixture of physicaldriven and data-driven,with fused precipitation as the ”true value”.The model takes historical physical quantity data and numerical prediction precipitation products as input,and reshapes the precipitation regression task into an ordinal regression problem.The results show that DORCGAN can effectively improve the performance of precipitation prediction,especially for heavy rain(≥ 50mm)and heavy rain(≥ 100mm)by about 14% compared to the risk score(TS)of the original prediction results.However,the model’s ability to improve precipitation prediction gradually decreases as the prediction time duration increases.