Multiscale Support Vector Machine Model For Monthly Precipitation Prediction

Reliable rainfall forecasting,both short-term and long-term forecasting,can not only provide key parameters for early warning of flood disasters,but also have great reference value for water resources management,agricultural cultivation and crop yield increase.However,rainfall is a complicated nonlinear atmospheric process,which is space and time dependent.Thus,it has been challenging to understand the complexity and predictability of precipitation,and produce accurate forecasts of precipitation on monthly and seasonal time scales.In order to improve the prediction accuracy of monthly precipitation,in this paper we propose a hybrid least squares support vector machine(LSSVM)model with data decomposition for forecasting monthly precipitation data from 138 meteorological stations in Yangtze River Basin and 11 large scale climate indices from 1960 to 2012.The data in the 43-year period of 1960 to 2002 are used for the calibration of models whilst those in the 10-year period of 2003 to 2012 for the test of models.First,the optimal lag time is identified by calculating the partial information(PI)between the forecast predictors and the standardized monthly rainfall to select the best model inputs.Then,the LSSVM model is built based on the parameters which are optimized by difference evolutionary algorithm to forecast monthly precipitation.To evaluate the improvement of the performance prediction of LSSVM model by using decomposition method,the EMD-LSSVM and DWT-LSSVM monthly precipitation prediction model are constructed by coupling least squares support vector machine and empirical mode decomposition(EMD)or discrete wavelet transform(DWT),respectively.At the same time,multiple linear regression(MLR)is used as a reference model to evaluate the prediction performance of LSSVM,EMD-LSSVM and DWT-LSSVM models.By comparing the Nash-Sutcliff coefficient,correlation coefficient and relative absolute error between the predicted and the observed values and provide the spatial distribution pattern of the evaluation indicators,we can evaluate the prediction accuracy of the four proposed model.The results are showed as follows:(1)The forecast skill of all modes exhibits a similar west-east spatial pattern in the Yangtze River basin.The prediction skill of models is most powerful at stations in the western region of the basin and the prediction performance in the eastern region is generally low,which tells a strong correlation with the spatial distribution of the randomness of precipitation.It is also shown that the prediction accuracy of all the models has a significant decreasing tendency as the randomness of precipitation increases.The LSSVM model performs better than the MLR model.Compared with the LSSVM model,both the EMD-LSSVM and DWT-LSSVM can improve the forecasting precision of the LSSVM model at more than 88%of stations over the Yangtze River basin in forecasting monthly rainfall.However,the improvements of the EMD and DWT to the LSSVM do not always work.The use of EMD and DWT lowers the forecasting performance of the LSSVM at about 11%and 5%of stations over the basin,respectively.The prediction skill of the EMD-LSSVM is more powerful at about 60%of stations but has a higher degree of variability than the DWT-LSSVM.(2)These four models have a reasonable but different performance for monthly precipitation prediction.Wherein,the LSSVM model performs better than the MLR model,having the higher NSE and lower RAN values at approximately 60%of stations.Both the EMD-LSSVM and DWT-LSSVM can improve the forecasting precision of the LSSVM model at more than 88%of stations over the Yangtze River basin in forecasting monthly rainfall compared to LSSVM model.The use of EMD and DWT lowers the forecasting performance of the LSSVM at about 11%and 5%of stations over the basin,respectively.(3)The improvement is more significant by the EMD-LSSVM and DWT-LSSVM,respectively,at about 39%and 28%of stations with high randomness of precipitation relative to LSSVM model.And the prediction skill of the EMD-LSSVM is more powerful at about 60%of stations but has a higher degree of variability than the DWT-LSSVM.