The Research And Application Of Improved Genetic Neural Networks On Time Series Prediction

With the development of Information Technology, The artificial intelligence technology has been used many application areas, such as Pattern Recognition, Gaming Intelligence, robot technology, and so on. As a glorious industry, artificial intelligence technology has been attrached more and more attention. Someone even thought that artificial intelligence technology will be one of the most important development direction in computer science in the 21st century.Neural network is an important branch of artificial intelligence in the field because of the good nonlinear approximation, it was applicated in-depth in many engineering applications. However, the neural network models'performance was poor because of some items, such as training time too long, easy to access to the local optimizations, network initial parameters hard to define, and so on. Genetic Algorithm is based on the view of Darwin's theory of evlotion in the survival of the fittest point. GA transferred the actual problems that need to solve into evolution processes of the biosphere and ultimately find the optimal solution. The ability of global opitimization of GA I particularly suited to compensate for neural networks to prevent into local optimum situation. The structure and initial parameters of the neural networks can be optimized by GA. Wavelet transform is now increasingly used as a kind of mathematical tool, its multi-resolution characteristics, with the reputation of'mathematical microscope'. Its time-domain analysis ability that Fourier transform be not was especially suitable for time series processing. Use wavelet transform to the time series for pre-processing, helping to improve the accuracy of neural network training and reduce the training time.Fusion application of the neural networks, genetic algorithm and wavelet transform was be used. Using the nonlinear approximation capability of the neural networks, the good ability of global optimization of genetic algorithm, as well as the time-frequency processing capacity of the wavelet transform to establish the Wavelet denosing Genetic Artificial Neural Network (WGANN) to predict the time series. This thesis has been get the following results:1. The initial structure and parameters of the Neural Network have optimized by the GA. Considering the limits and deficiency of current methods to optimize neural network using GA, combing correspoding datasets, Dual-Phase Optimization(DPO) is proposed to optimize the neural networks'initial structure and parameters.2. A novel model, termed WGANN, is developed. Experimental results show that the WGANN has the better performance in the prediction accuracy for the upper reaches of Minjiang River runoff comparing with the existing approaches. Neural network is the main framework of this model. The Dual-Phase Optimization (DPO) is proposed to first determine the structure and then optimize the initial parameters of the neural networks. The wavelet transform is used to the data's pre-processing. At the end of this thesis, used a time series of runoff of the upper reaches of Minjiang River, predicted its shor-term, mid-term and long-term runoff. Then, compare with traditional neural networks model. Finally, the computer simulation had demonstrated that the performance of proposed WGANN was much better than any widely-uesd traditional neural network models. This model provided a new way for analysis of time series.