Study On Prediction Of Eutrophication By BP Neural Network Based On Bayesian Arithmetic In West Lake

A Bayesian Back-Propagation (BBP) Neural Network was studied, whose generalization was improved by Bayesian Arithmetic. And the model was applied to predict the eutrophication of West Lake in Hangzhou. We have established 8 sampling spots in West Lake, and selected spot 7 as main study object to set up a BBP network, which can most represent the water quality status of the lake. Then we used spot 3 to test the performance of the network, at the same time, two other networks were built for comparation with the BBP network we have set. The results showed that BBP has better generalization than conventional BP arithmetic and other arithmetics (Early Stop for example) in the same net-scale and training-error, and higher speed in convergence. The correlative coefficients R could be 0.8867, with the Mean Absolute Error(MAE) and the Mean Square Error(MSE) being 15.07 and 377.4 respectively for the training set, while the predicting precision were 3.5% and 16.5% for the corroborant set. This network could also solve the uncertainty and non-linear among factors in eutrophication so as to get higher precision in prediction, and this will be useful in planning and carrying the work out to improve the water in West Lake. This indicates that BBP network is an effective method for forecasting the concentration of chlorophyll-a. And it can provide the scientific basis for the control of the eutrophication of West Lake.