Application Of BP Neural Network Based On Wavelet Analysis To Predict The Trend Of Eutrophication In West Lake

Artificial neural network has a powerful nonlinear approximation andfault-tolerant capability, while wavelet analysis has both a good time-domain and agood frequency-domain multi-resolution analysis capability. This thesis combines thetwo advantages, and applies a wavelet back propagation neural network based onmulti-resolution analysis (Wavelet neural network, WNN) to model the aquaticecosystem of West Lake and predict the short term trend of eutrophication in it.This thesis adopts the aquatic data of West Lake from January 2000 to July 2001by routine measurement and filtrates appropriate objects for modeling by principalcomponent analysis. Using wavelet multi-resolution analysis with proper waveletfunction and scale, the load series of chlorophyll-a are decomposed into alow-frequency and several high-frequency sub-series, which show the differentfrequency characteristics of the load. On the basis of this, two alternatives areconstructed which are called Modelâ… and Modelâ…¡.For Modelâ… , the low-frequencysub-series are used as an input parameter of the network to predict the concentrationof chlorophyll-a; For Modelâ…¡, the major (low-frequency sub-series) and minor(high-frequency sub-series) are predicted with different sub-networks respectively.The final result is achieved by reconstructing all the predicted results of thesub-networks together. The results show that Modelâ…¡combines the localizationcharacteristic of wavelet analysis, so it helps to overcome the defects of BP neuralnetwork such as confusing the principle and secondary information when it isoperating. Modelâ…¡not only enhances the stability but also avoids the phenomenon ofover-fitting of the network. In addition, this method possesses higher forecastingaccuracy and good adaptability. Therefore, it can be essential to eutrophicationmanagement and decision-making for West Lake.