Research On Neural Network Forecasting Of Temperature And Precipitation Probability In Southeastern My Country

Based on the emsemble forecasting data from ECMWF in TIGGE dataset,the ERAInterim reanalysis data and CMORPH precipitation products,the study of 2-m temperature and24-h accumulated probabilistic forecast experiment over southeast China(20°N-36°N,106°E-125°E)is carried on.Firstly,FNN(Feedforward Neutral Network)and CNN(Convolutional Neural Networks)are used to build the probabilistic forecast model for 2-m temperature,which is normally distributed.Then the model of probabilistic forecast for non-distributed precipitation is built.Considering the differences of the improvement capabilities of different area in precipitation forecasts,the geographic information of each point is added to the probabilistic forecast system to build the FNN-GI(Feedforward Neutral Network-Geographic Information)and CNN-GI(Convolutional Neural Network-Geographic Information)model.Thoese models are evaluated and compared in the study.In terms of the probabilistic temperature forecast,the forecast results have been improved obviously with bias correction by the FNN and CNN,The results show that CNN performs better than FNN.Although FNN,CNN and raw ensemble forecasts all have limited dispersion,the probability distribution of FNN and CNN is better than that of raw ensemble forecasts.Meanwhile,the forecast accuracy of FNN and CNN is higher than the raw forecasts,but the degree of improvement decreases with the increase of forecast lead time.CRPS of the CNN and FNN is 20.86% and 21.94% lower than that of the raw ensemble forecast for 24 h lead time,and for 168 h,it is 5.50% and 7.59%,repectively.In terms of the probabilistic precipitation forecast,FNN and CNN enlarge the dispersion of the probability distribution of precipitation from ECMWF,and the probability distribution of precipitation for each threshold is more reliable.The overall forecast accuracy of FNN and CNN has improved,and the improvement of CNN is greater than that of FNN,as the CRPS for24 h lead time decreases by 14.29% and 16.27%,respectively.However,for the precipitation with a larger threshold value,the forecasting skills of the FNN become lower than that of ECMWF because the sample size is relatively small,and the improvement of CNN is also smaller than the precipitation when the threshold value is low.At the same time,for different locations,the improvement effects of FNN and CNN are different.In the early stage of forecasting,the forecasting skills of the Changjiang Plain are relatively low.Compared with the probabilistic precipitation forecast of FNN and CNN,the prediction accuracy of FNN-GI and CNN-GI are improved.The forecast accuracy of precipitation with larger thresholds is also improved and the forecast skills at the 50 mm threshold from the FNNGI and CNN-GI are higher than those from the raw ensemble forecast.