SARIMA-CEEMD-LSTM Model For Predicting Long-term Precipitation

Precipitation observation is one of the most important basic data in water resources scientific research.The temporal and spatial distribution of precipitation in many areas is uneven,and drought and flood disasters often occur.Therefore,in production and life,not only short-term precipitation forecast is needed,but also medium and longterm precipitation forecast are necessary.In order to establish an efficient long-term prediction model of precipitation time series,a new model named SARIMA-CEEMD-LSTM was constructed based on 30years(1990-2019)precipitation data of 36 meteorological observation stations in East and central China by using Complementary Ensemble Empirical Mode Decomposition,Seasonal Autoregressive Integrated Moving Average and Long Short-Term Memory neural network.The new model combines the advantages of SARIMA with LSTM neural network,and CEEMD decomposition makes LSTM neural network model information mining information more thoroughly.It makes the prediction efficiency and prediction accuracy of the new model further improved.By combining the new model with the interval prediction of quantile regression mean method,the point prediction and interval prediction of precipitation are realized simultaneously.The procedures of SARIMA-CEEMD-LSTM are as follows: First,36 meteorological observation stations were fitted with the SARIMA model,and then the residuals were decomposed by CEEMD.Next,based on the results of cluster analysis,LSTM models were established for two types of meteorological observation stations.Finally,the final prediction results were obtained by summation of LSTM fitting results and SARIMA fitting results.MSE,RMSE and adjusted R-square were used to evaluate the new model and other models,where other models include SARIMA,LSTM,ARIMA-BP,GM(1,1)-Markov and CEEMD-RF-KRR.The results show that the MSE and RMSE of the new model are the smallest,and the adjusted R-square are the largest.The results of interval prediction by quantile regression average method were evaluated by forecast interval coverage and forecast interval width,and compared with nonparametric kernel density estimation method and prohet algorithm.The results show that the difference of forecast interval coverage between the three methods is small,but the forecast interval width of quantile regression average method is the narrowest,with a maximum reduction of 24.43%.This study provides an effective prediction method for the future precipitation of each meteorological observation stations in East and central China.And the study can also provide reference for the government departments in the province to take efficient measures to deal with natural disasters and rationally plan urban water resources.