Downscaling Forecast For Zhejiang Precipitation In Summer Based On The BP Neural Network Model

At present, the numerical forecasting products in our country are for a large scale. To be exact, they can just fine to the county station, which can not meet the demand of township forecasting. Therefore, fining product of meteorological elements to form a spatial resolution of10km×10km is the expected goal of fined weather elements forecasting. This paper adopts data of two types of resolutions based on a large scale observatory and high resolution of automatic weather station and is combined with three methods-artificial neural network technology, the weather typing techniques, empirical orthogonal function, establishing reliable large scale live field and high resolution analysis field, and generate nonlinear downscaling function.The downscaling is realized by utilizing the summer daily precipitation data of60weather stations and372musicales automatic stations in Zhejiang province from2007to2012, the daily500Hpa geopotential height data based on the historical reanalysis grid data of NCEP/NCAR global1.0°×1.0°observation data of three sounding stations. Using the mid-low level wind of sounding station summer precipitation data and the classification scheme developed by Jenkinson are further divided, in four main different wind and in four main different types, local single-station precipitation,the EOF principal component of local precipitation as prediction components as prediction object, associated with large-scale precipitation by0.1significant test of principal components as predictors, using artificial neural network method to build downscaling forecasting model, called as BP1model、BP2model. The linear and nonlinear relationship between large scale circulation and local scale precipitation field in different mid-low level wind direction and classification schemes are analyzed, the analysis effect of the scale model test prediction is compared with the commonly used spatial interpolation methods. Throughout the forecast during the test, we come to the following conclusion:(1) The analysis of mean daily precipitation and distribution in three regions which are north Zhejiang province, southwest Zhejiang and the coastal area of eastern Zhejiang, indicating that the precipitation in the southwest、west and northwest is more than other wind, and the regional differences distribution of daily precipitation is obvious.(2) From the selection of predictors based on BP neural network, we can find that the number of selected predictors based on weather classification is more than the ones based on the wind configuration.(3) By the analysis of the neural network model and the linear regression model with the same prediction object fitting and forecasting daily rainfall, we find the results showing that the nonlinear model is better than the linear model in terms of fitting and forecasting.(4) BP2model among three artificial neural network models, according to the analysis of the forecasting for amount precipitation、error evaluation, the results are better than the BP1model.(5) To further discuss the application scope of BP2model, electing original predictors from the result of the circulation classification, according to the Statistical indicators, which means absolute relative deviation(MRD), and the results show that the prediction of the downscaling model applied to rainstorm is the best of three types of rainfall amount after defining rainfall with different levels.