Using Artificial Neutral Network Model To Predict Water Quality In Yellow River

This paper studied the diffusing and moving rules in the observation surfaces of Sanmenxia, Xiaolangdi and Huayuanke because of the river pollutant (COD_Cr and NH₄⁺-N are appraising genes). The studying area is from Wei River (Tong guan) to Huangyuanke in middle reaches of the Yellow River. Because of the randomcity in the Yellow River, the deferent uncertainty was brought in the model of water quality forecast. Using manual neural network method, the water quality model was built to reflect the relation of backward position's pollutant concentrations and upward position's influence factors. It was applied to open out the impersonal rules of upward and back ward's main water quality factors, analysis and validate the applicability of this method. The main results show that:(1) Because of complexity of the Yellow River's water circumstance, the actual measure data showed biggish inequality. If these data were forced to combine and built single model, it was not successful for very complicated water quality system in the Yellow River. In this paper, the studying area was carved up four typical spaces (from Long men to Tong guan, from Tong guan to Sanmenxia, from Sanmenxia to Xiaolangdi and from Xiaolangdi to Huayuanke) and different flow levels (including Q≤500m³/s, 500m³/s3/s, Q≥1000m³/s) in every space. A lot of lesser neural network models were built to replace single model. The advantage was showed by using this mean, according to the stability and forecast capability of model.(2) The main influent factors were selected about backward surfaces. They are upward flow (Q), interzone pollutant (W), sand quantity (S), reservoir water level (Z) and background concentration of upward surface (Co). The training mode was built, such as three layers framework of BP network model, six nerve cells of connotative layer, best inputting mode of 5-6-6-1 BP model, studying speed 0.01, most training times 2000¹0000, training error 0.0001⁰.001, forward disposal function "prestd", backward disposal function "poststd", and the function of "trainlm" was applied to train and test the network.(3) Matlab language was used to write training and applying programs. In the trainingprogram, the forecasting model of influent factors was built, according to twelve year's statistic data in representative studying reach. At the same time, the relative error of testing data was analyzed in the program. The affections of imitation and forecast were showed visually by using figures, when the model was trained well. In the course, the matrixes of weight and remainder were transferred from connotative layer to output layer. The output results of backward surface were calculated in the model.