Comparative Study On Monthly Rainfall Forecasts In Dongting Lake Basin Based On Multiple Neural Network Models

Precipitation is a complex atmospheric process, which is space and time dependent. Monthly rainfall forecast can provide key parameters in terms of water resource management, flood and drought disaster early warning, hydrological forecast analysis. Because of the complexity and diversify of meteorological conditions, the influence of rainfall information distributed in different time scales of predictors and the complicated non-linear/linear relationships between the predictors and monthly precipitation,the monthly precipitation forecasting accuracy obtained by the traditional method without data processing and selection is really low. To improve the forecasting accuracy of monthly precipitation and deal with the problems about predictors processing and selection, this paper introduces a wavelet mutual information neural network model. The standardized monthly precipitation anomaly and large-scale climate indexes are first decomposed into subseries components with different time scales, and determine the best predictors by mutual information algorithm. The standardized monthly precipitation subseries are then forecasted, respectively, under different time scales by using the cascade-forward(CF) neural networks in which the number of hidden neurons isoptimized by particle swarm optimization(PSO). Finally the monthly precipitation is forecasted by combining all predicted subseries and using the inverse transform of standardized monthly precipitation. In this study, we analyzed monthly rainfall data from 27 observation stations in the Dongting Lake district together with large-scale climate indexes which include Pacific Decadal Oscillation, North Oscillation Index, Global Mean Temperature Anomalies, Indian Ocean Dipole, Antarctic Atmospheric Oscillation, North Oscillation Index, North Atlantic Oscillation during 1961- 2012. The neural network, wavelet neural network, wavelet mutual information neural network are developed for effectively forecasting monthly rainfall in the Dongting Lake district with training data from 1961 – 1992, validating data from 1992 – 2002 and testing data from 2002 – 2012. The content and results are showed as follows:(1) From the Nash-Sutcliffe efficiency and Relative absolute error viewpoints, the wavelet mutual information neural network provides the highest forecasting accuracy among these three models and wavelet neural network outperforms neural network.(2) The wavelet mutual information neural network not only improves the monthly rainfall forecast accuracy but also can provide more precise prediction accuracy of extremely monthly rainfall.(3) The forecast results of neural network, wavelet neural network and wavelet mutual neural network show that the mean Nash-Sutcliffe efficiency is 0.64,0.69 and 0.71 and the relative errors is 0.46,0.35 and 0.33, respectively.(4) The implement of wavelet transform and mutual information correlation analysis to process and select input variables can effectively select the best predictors, and improve monthly rainfall accuracy in Dongting lake district.