Wavelet Network And Its Application In Environment System Modeling

The research of Artificial Neural Network(ANN) and wavelet analysis is the forefront and focus of information science technology today, which has great value both in theory and application for Modeling and predicting of complex systems.Wavelet Neural Network(WNN) gets rapid development in recent years, and is widely used in time series analysis. WNN inherits both the virtue of wavelet in regionalizing of time-frequency field and self-learning ability of neural network and has strong capability in function approximation and error tolerance. There are two combinations between wavelet and neural network. The first one-called wavelet network with relaxed structure is to deform the input series into different levels of resolution scales by the wavelet transform. The other-called wavelet network with compact structure is to substitute the base function in neural network with scaling function (or wavelet mother function).To begin with, this paper gives the structure of compact Wavelet Neural Network from wavelet analysis theory. In order to avoid low efficiency in traditional algorithm, and to improve the learning performance, this paper gives the improved algorithm."Curse of dimensionality"has always been a difficult problem of traditional WNN in high dimensional cases. So this paper gives a WNN with new structure. The main feature of the proposed WNN is to multiply the reconstruction of each dimension in the output layer instead of adding them as usual. Thus the scaling functions will be generated automatically to cover the input space.It is difficult for uniform resolution wavelet network to obtain a good approximation when it learns form absent data or nonuniform distribution of training data. The algorithm of nonuniform resolution is developed to solve this problem, and it is applied to compensate weather data in the modeling of environment system. A Decomposition-Reconstruction-Prediction wavelet neural network is proposed, and it is applied into the construction of predicting model of nitrogen dioxide concentration in the winter of Harbin. The inputs of the predicting model are chosen, and the data are compensated, denoised, and pretreated. At last, the improved multidimensional wavelet network, relaxed structure wavelet network and BP network are respectively used in the construction of prediction model of nitrogen dioxide. The simulation shows that the algorithm of this paper is suitable and advanced.