| As the second largest tributary of Xijiang River in the Pearl River basin,with abundant average annual total of water,large amount of flood,and largest gross of flood in the all major tributaries,Guangxi Liujiang River is easily affected by the warm moist air flows from southwest to rain storm which runs quickly together to form flood in.Runoff prediction plays an important role in the synthetical development and utilization of water resources,such as flood and drought control,etc.Analysis and prediction of Liujiang River is of great significance to social development.However,as a multi-objective and multi-level huge system,Liujiang River’s high nonlinearity in space and time makes it hard to describe and resolve the proposed models,and the satisfactory solution is still unavailable so far.So it is very necessary to use new theories and methods to study Liujiang River hydrology system in other space and different ways.In this paper,both the complexity of runoff forecasting and the existing works on runoff prediction are analyzed.Then some of the main physical quantities of Liujiang River runoff,such as the annual total of runoff,water level,flow velocity,are analyzed and predicted by using some new ideas and methods.The main contribution are shown as follows:(1)Noise reduction of Liujiang River runoff series.The current methods based on wavelet threshold de-noising usually assume that noise is distributed in high frequency,therefore de-noising is only applied to the high frequency and the effect of noise in low frequency of the signal is ignored.With analysis for the noisy characteristics of the Liujiang River runoff,combining with the wavelet decomposition coefficient of different noises which has different changing rules in the wavelet decomposition scale,the recognition method of noise categories based on wavelet de-correlation is proposed.Then,according to the identified types of noise,the optimal layers of wavelet decomposition and the noise intensity,etc.,an adaptive determination method of threshold is put forward to de-noise in both low and highfrequency.The method can be used for not only the noise reduction of the Liujiang River runoff signal,but also the common noise reduction work.(2)Chaotic characteristic analysis of the Liujiang River runoff.A weighted periodic by power and weighted average method is proposed to calculate the average period of Liujiang River runoff with chaotic characteristic.The chaotic characteristics of the daily water level time series,which is in different area of the Liujiang River basin,are analyzed in different seasons and environment.Then the influence of seasons and environment on the chaotic characteristics of the daily water level of the Liujiang River is summarized.The chaotic characteristics and wavelet characteristics of the annual total of the Liujiang River runoff series with finite length is also analized,the conclusion is obtained that the annual total runoff of the Liujiang River also has chaotic characteristics.The result is applied to predict the annual total runoff of the Liujiang River and to optimize the performance of the model.(3)Wavelet characteristic analysis of the Liujiang River runoff.The subseries of the annual total runoff time series of the Liujiang River runoff generated by the wavelet decomposition remain the chaotic characteristics of the original runoff series.Experiments show that if the original series has chaotic characteristics,the corresponding wavelet coefficients also have chaotic characteristics.Liujiang River runoff cycle change under time scales is revealed by using wavelet analysis: the scale cycle of annual total runoff series is the least,that of monthly maximum flow series is larger and that of daily water level series is the largest.In addition,the wavelet scale cycle of water level series in wed season is less than that in dry season.Thus,when analyzing Liujiang River runoff series,the wavelet scales should be selected according to the units,time period,etc.,and then the scale cycle and other laws of the runoff series can be obtained through comprehensive analysis of real,modous,modous square and wavelet variance of the runoff series.(4)Prediction of the Liujiang River runoff time series.BP neural network model with double hidden layers based on Levenberg-Marquardt algorithm is presented to forecast daily water level time series of the Liujiang River;wavelet transform is used to analyze the annual total runoff and the annual maximum flow series,which significantly improves the performance of the original prediction model.Thus severalrelevant models are proposed: wavelet transformation and optimal subset regression model is proposed for the annual water level series of the Liujiang River.The predictive value of the model can be extrapolated to many steps with good accuracy.A neural network predicting model optimized by genetic algorithm based on wavelet transform for the annual total runoff of the Liujiang River is proposed,so is the model of neural network model optimized by genetic algorithm based on wavelet transform and chaos theory.The latter is the improvement of the former.All of the above conclusions are verified by experiments.Results show that these models can not only fully reflect the trend of Liujiang River runoff time series,but also have good stability when forecasting runoff accurately.They provide effective modeling methods for runoff time series of the Liujiang River and other rivers. |
