Research On Day-by-Day Precipitation Forecast Model Based On Support Vector Machine

As one of the common meteorological factors,precipitation has always been the key and difficult issue of weather forecasting due to its non-linear and discontinuous nature.The short-term,concentrated precipitation has led to heavy rain,floods and other meteorological disasters,which have adversely affected the development of the national economy and people's daily lives.How to improve the accuracy of precipitation forecast has always been one of the goals pursued by researchers in the field of meteorology.With the breakthrough of the basic theory of numerical model and the development of key technologies,the spatial and temporal resolution of numerical forecast products has been continuously improved,making the numerical forecasting model an important reference for daily weather forecasting.In order to obtain more accurate weather forecast results,this paper uses Support Vector Machine(SVM)to study the forecasting method of numerical forecast products.This study selected 36 ground stations in Guangdong Province as research sites,based on the daily output data of the ECMWF(European Centre for Medium-Range Weather Forecasts)numerical forecast model and the precipitation data from the ground observation stations.The forecasting study was conducted on the daily precipitation from April to June in the flood season in the period from 2013 to 2017.The research content of the dissertation mainly includes:According to the climatic characteristics of precipitation at the research site,the precipitation of the skewed distribution was normalized,and the precipitation was divided into two modules,namely the sunny rain and precipitation level.With regard to whether or not there is precipitation at the research site for 24 hours,the relevant field surveys were conducted on the element fields of the numerical products of the European Medium-Term Forecasting Center,and the forecasting factors were initially selected.Then,the principal component analysis was used to further reduce dimensionality of the primary screening factor,and the principal component factor with high correlation was extracted as the forecasting factor.A prediction model based on the combination of principal component and support vector machine was established.36 observational stations were established to model the 24-hour weather forecast.Firstly,ECMWF model forecast products are correlated with forecast objects,relevant factor fields are extracted from them,and bi-linear interpolation method is used to interpolate the element fields to the corresponding site positions,thus constructing a forecast station for observation sites.Then it will be used as a feature input model of the support vector machine to obtain the forecast results of the model and compare the result with the model results established by the principal component and the support vector.The comparison results show that the prediction accuracy of the model combined with principal component and support vector machine is about 10%~15% higher than that of the support vector machine model.According to the relevant provisions of meteorology,the daily average precipitation of 24 hours at each observation site was divided into six levels of no rain,light rain,moderate rain,heavy rain,heavy rain,and heavy rain.Taking the 24 h average precipitation level as the forecast object,a precipitation grade model combining principal component and support vector machine was established.The performance of this model is compared with the forecast performance of ECMWF numerical forecast output products.The results show that such a numerical forecast product release method has a better forecasting effect on precipitation levels than numerical products.Compared to previous work,the paper has made some innovative attempts in the following areas:(1)A site-by-day precipitation forecast model was established to implement the study of the local precipitation forecast model.For the study of the release forecast of numerical forecast products,most scholars use single station or subarea methods for the interpretation of numerical forecast products.This paper establishes a site-by-station precipitation forecasting model for the study area,and achieves a refined regional forecast of the objective meteorological elements.This is a useful attempt.(2)Principal component analysis method is used to further reduce the dimension and denoise the input factors of the precipitation forecast model,which greatly improves the prediction performance of the support vector machine model,and improves the prediction accuracy of the support vector machine model with some new efforts.