The Application Of Several WA-ANNs In The Long-term Forecasting For The Hydrological Time Series

The accurate and efficient long-term hydrological forecasting can provide scientific evidence for water and meteorological departments when working on the planning and decision making of flood control and drought relief,dispatching and running of hydraulic project and water resources exploitation and utilization.In order to investigate the method for building and designing the artificial neural networks suitable for the long-term forecasting of the hydrological time series,different types of artificial neural networks based on the wavelet analysis(WA-ANNs)are built for hydrological information,including annual maximum water level,annual precipitation and flood season precipitation of Huangpugongyuan hydrologic station,Wusong hydrologic station and Xujiahui weather station in Shanghai.The wavelet analysis is the first step of WA-ANNs,which is used to identify and separate the deterministic and stochastic component.Then,different types of artificial neural networks are built for subcomponents respectively,including BP neural network,radial basis function neural network,general regression neural network,wavelet neural network and Elman neural network.The composition of deterministic and stochastic forecasting results is the final results.In the process of designing the artificial neural networks,multiple methods are applied to determine parameters such as chaotic characteristic analysis,autocorrelation analysis,empirical formulas,trial-and-error method and genetic algorithms.The performance of WA-ANNs is evaluated through several indexes such as mean absolute percentage error,mean absolute error,root mean square error and qualified rate.The main conclusions are as followed:(1)WA-ANNs are reasonable and practical when applied in the long-term forecasting of sample hydrological time series in the study area.The forecasting results show that WA-ANNs outperform ANNs due to the higher precision.(2)The methods for determining parameters are reasonable and practical.The number of the nodes in input layer is set through chaotic characteristic analysis and autocorrelation analysis.The number of the nodes in hidden layer is set through empirical formulas and trial-and-error method.The initial weight and threshold of BP neural network are optimized through genetic algorithms.WA-ANNs based on the above-mentioned designing methods perform well and meet the requirements of long-term hydrological forecasting.(3)ANNs are more suitable for the long-term forecasting of sample hydrological time series in the study area than autoregressive model,grey model and threshold autoregressive model given the higher precision.The general regression neural network performs best among the ANNs both in the precision and stability.The mean absolute percentage error of general regression neural network applied in the sample hydrological time series is 3.8%,2.8%,11.8%and 19.7%,and the qualified rate is 93.8%,100%,77.3%.68.2%.The performance of general regression neural network for the flood season precipitation of Xujiahui weather station meets the minimum requirements merely,which needs to be improved in the henceforth research.