| Forecast is absolutely necessarily important tache and premise before decision-making and layout. Time series forecast is a important research aspect in forecast fieldo Research on Forecast of nonlinear time series takes on important practical significance because practical system is mostly nonlinear. However, it is complicated for practical system, and people have difficulty correctly choosing model to forecast, which causes application of forecast methods based on visible model being restricted. So, people turn to forecast methods based on phenomenon. Research on forecast based on artificial neural networks and chaotic theory is research hotspot at present, which is attached importance to and lots of fruits are got about ito Even now, there are many problems which should be researched ulteriorly when artificial neural networks and chaotic theory are applied to forecast and there is much research room. Systematic and embedded research is done in this dissertation. By all appearances, the research in this dissertation has clear academic significance and great practical significance.The contents of this dissertation are listed below:1. Based on reconstruction of phase space of dynamical system, we research the divinable capability of time series from the point of view of nonlinear dynamics by constructing recurrence plot.2. By dint of the concepts of entropy and redundancy of information theory, we get expression of computing redundancy showing at correlative dimension form based on the definition of Renyi a entropy, and we provide qualitative and quantitative methods detecting time series nonlinear property.3. We deduce frondose algorithm of three layers BP neuralnetworks which is used in common, and discuss several important issues in designing neural networks which is used to forecast, for example, number of hidden layer, nerve cell number of hidden layer, epoch of learning, embryonic power value, decision of node number about input and outputo At the same time, this dissertation sums up things that conventional BP algorithm is improved on considering disadvantages of it. In addition, we improved on conventional BP algorithm from the point of transfer function and networks structure.4. People put forward radial basis function networks considering the conventional BP algorithm problems of slow convergence speed and easily getting into local dinky value. The difference between BP networks and radial basis function networks is pointed out, and we discuss the decision of two important parameters of radial basis function networks.Also, we improved on HCM algorithm which is often used to forecast by combining input vector with output vector and getting extended vector when we decide key parameter of radial basis function networks----center vector.5. We discuss the forecast method which based on wavelet neural networks by combining good time and frequency local analysis ability which wavelet analysis possesses with learning ability which neural networks possesses, and bring forward a frondose, banausic algorithm in this dissertation0 Also, a essential thinking of combined forecast based on wavelet neural networks is described and a essential trait of combined forecast based on wavelet neural networks is pointed out.6. We describe the meaning of chaos> future idea of chaotic theory and influence on forecast; introduce the character of chaotic time series, and point out the problem and shortage of the methods already existed computing character value which are fractal dimension and the largest Lyapunov exponent andimprove on it; present the forecast principle of forecast method based on chaotic attractor, and point out the shortage of local field forecast method based on chaotic attractor and bring forward improved on methodo At the same time, we put forward a banausic algorithm and compare two models using practical example.7. We provide a few rules which should be considered when people choose forecast method, and a compact comment on this dissertation forecast methods...
|
PDF Full Text Request