A Study On Multimodel Ensemble Forecasts Of Meterological Elements Using LSTM

Based on the 1-7 days ensemble forecasts of the Europcan Center for Medium-Ranger Weather Forecast(ECMWF),the National Centers for Environmental Prediction(NCEP),and the Japan Meteorological Agency(JMA),the UK Met Office(UKMO)and the Korea Meteorological Administration(KMA)in the TIGGE datasets,the multi-model ensemble forecasts of surface air temperature,dew point temperature,U wind component and V wind component in China's regional during the period from 1 January to 30 September 2015 have been combined with ERA-Interim re-analyze the daily data that used as "observation" data to train and forecast.First,using the "observation" data of different elements and different models,understand theirs forecast errors and evaluate their forecasting capabilities.Then perform post-processing forecast experiments on these models with long-term memory neural network(LSTM),neural networks(NN),Random Forest(RF),Bias-Removed Ensemble Mean(BREM)with running training period and Multi-model Superensemble(SUP)with running training period to evaluate the prediction effects of different post-processing methods.Finally,by traditional multi-model integration and machine learning we explore the integrated forecast for regional high temperature weather,which is the third process(July 22-August 21)in southern China of high temperature weather in the summer of 2013.Taken together,the best model for forecasting different meteorological elements in China is ECMWF.The prediction effects of each model are quite different,and there are obvious differences in geographical distribution.In the selected period,there are seasonal differences in forecasting capabilities.The surface air temperature has large errors in winter and spring;the U wind component and V wind component have relatively large errors in spring and summer;the dew point temperature has large errors in spring and winter and small errors in summer.The forecasting capabilities of different forecasting methods in China in order from good to bad are LSTM,SUP,NN,ECMWF,BREM and RF.There are some differences in the six kinds of results in geographical distribution.The difference between surface air temperature and dew point temperature is greater than U wind component and V wind component.But for the error reduction rates compared with the LSTM is exactly opposite.With the increase of lead time,the improvement effect of LSTM is more obvious.LSTM has the smallest prediction error for the high temperature process in 2013 of the six kinds of results.Compared with ECMWF,NN and RF,the error reduction of LSTM is highly obvious;compared with SUP and BREM,the error reduction is not large.For the increase of lead time,the root-mean-square errors of the six results increase grow with different degrees.And they are not stable for high temperature forecasting.