Water Simulation Method Based on BPNN Response and Analytic Geometry

With three-year monthly monitoring data (from Jan. 2007 to Dec. 2009) of ten given sections of one river in Beijing, a BP neural network of DO and principle components are built by taking the concentration of DO as the evaluation criterion of the water quality. Three orthogonality principle component vectors are exacted through correlation analysis and principal component analysis with other monitoring parameters (water temperature, pH, CODCr, CODMn, BOD, NH3-N). After the evaluation on the availability of the net with root mean square error (RMSE) and the average relative error of the net's response of the test set, the 3-D figure is built by the net's response and the three PC vectors. For the continuous output response of the BPNN, the errors of the untrained samples closed to the trained ones are trivial. The status of water quality and the location in the 3-D space are represented by a sample selected with the cluster analysis. As the center of sphere, the value of PC vectors is used to find the shortest vector for raising the DO concentration inside the sphere. According the contribution of the monitoring parameters to each PC vector, the most sensitive parameter to the DO concentration can be found to meet the demand of the required DO standard. This method finds the empirical function between DO and temperature, BOD, NH3-N, which can be used to make up for the missing data and reduce the monitorial indexes, also could serve as a reference for the water quality management.