The Application Of Wavelet Analysis And Neural Network In Air Pollutant Forecasting

Wavelet analysis and Artificial Neural Networks(ANNS) has made great progress in many applied fields of natural science. This paper has deeply researched the application of wavelet analysis and ANNS in air pollution prediction, widened its application scope in air pollution prediction, and established firm base for further air pollution analysis and prediction.The major research results in this paper are as follows:Firstly, the yearly-changing trend of concentration on air pollutant are analyzed by using the wavelet reconstruction of low frequency signals of the highest layer, the jump features of variations are analyzed by using the wavelet reconstruction of high frequency signals of the last two lowest layer.Secondly, a new prediction model based on wavelet time series are developed. By wavelet decomposing, air pollutant concentration series is decomposed into many wavelet coefficient series according to scale, then by building time series model, prediction results of every wavelet coefficient series are acquired, Subsequently, by reconstructing the result, we get the forecasting of original series. This wavelet time series prediction model is adapt to nonlinear and nonstationary time series, and can make multi-step prediction into reality.Thirdly, the paper has studied the issue on the input samples of ANNS model. Input samples of ANNS model are analysed and determined on the basis of air pollution meteorology, and dimensionality of input variables are reducted using Principal Component Analysis(PCA).Fourthly, a new BP ANNS prediction model with divided pattern is constructed, and each pattern are pertinently designed. Generalization ability is improved by using early stopping, neural network ensembles and Bayesian regularization. The results showthe divided ANNS model has good quality in terms of prediction precision, calculation speeds and generalization.Finally, a Decomposition-Reconstruction-Prediction wavelet neural network is proposed. Through the wavelet transform air pollutant concentration series are decomposed into wavelet coefficient series on different scales, then these wavelet coefficients series are reconstructed and predicted by using appropriate BP neural networks, thus, after the synthesis of the predicted results of wavelet coefficients series, the final prediction result of the air pollution concentration is obtained. The predicting results show that the Decomposition-Reconstruction-Prediction wavelet neural network has good quality in terms of prediction precision and generalization.