Research On Prediction Model Of Roughness Coefficient Of Open Channel Based On Neural Network

Roughness coefficient is the most important hydraulic factor in open channel hydraulic calculation.It reflects the extent to which the fluid is obstructed by the boundary wall when moving in the riverbed.In the process of studying the roughness coefficient,the complexity of natural river course and the sensitivity of roughness coefficient make the study of the law and value of roughness coefficient received a great obstacle.At present,the value of roughness coefficient depends on lookup table method and practical experience,etc.,which can not well meet the practical needs in engineering,and it is often necessary to analyze and demonstrate again or model test to get a reasonable roughness coefficient.In order to better predict the roughness,the artificial roughening rectangular open channel is taken as the research object,and the method of neural network is applied to model predict the roughness coefficient in the range of0.004-0.024,and the following four prediction models are established:BP neural network model based on L-M algorithm,BP neural network model based on BFGS algorithm,RBF neural network model and GRNN neural network model.The optimal neural network model was established by comparing and evaluating the prediction results of each model,and the multiple nonlinear regression model was used to fit the relationship formula of roughness coefficient.The following related studies were carried out and the following conclusions were reached:(1)Evaluate and compare the four neural network prediction models from the perspective of prediction accuracy and computational efficiency.The results show that the prediction effect of the four neural network models on the roughness coefficient is ideal.The BP neural network prediction model based on L-M algorithm has the best performance.Among the other neural network models,the accuracy of RBF neural network model,BP neural network model under BFGS algorithm and GRNN neural network model decreases successively.(3)By comparing the neural network prediction model with the previous prediction model of support vector machine,it is concluded that in the training process of BP neural network L-M algorithm and BFGS algorithm,the network is more sensitive to the independent variable flow Q,and the accuracy decreases when the independent variable is average water depth h.RBF neural network and GRNN neural network with radial basis function as transfer function are more sensitive to the average water depth of independent variable.(4)BP neural network prediction model under L-M algorithm is used to predict the roughness coefficient of 0.2mm?1mm,which complements the data blank when the Froude number is less than 4.It is found that the change rate of roughness coefficient changes gradually with the increase of Froude number if the slope increases gradually.(5)Use the method of multiple nonlinear regression to fit the empirical relationship between relevant hydraulic elements.The relative errors are compared with four kinds of neural network models.It is found that the prediction model of BP neural network roughness coefficient under L-M algorithm is more stable and the error fluctuation is very gentle.The superiority of BP neural network prediction model based on L-M algorithm in predicting roughness coefficient is further explained.