The application of Improved Back Propagation Neural Network on the determination of river longitudinal dispersion coefficient

Abstract

The Improved Back Propagation Neural Network (IBPN) model was developed to predict the longitudinal dispersion coefficient for natural rivers. The hydraulic variables [mean flow depth (H), flow velocity (u) and shear velocity (u*)] and geometric characteristic [channel width (B)] constituted inputs to the IBPN model, whereas the longitudinal dispersion coefficient (Kx) was the target model output. The model was trained and tested using 23 data sets of hydraulic and geometric parameters, of which first 20 data sets were used to train and validate the model, and the rest data to test. In this model, cross validation theory was applied. To overcome the shortage of the traditional BPN model, the network was designed to determine the optimal weights and thresholds by random sampling at the interval (−1,1) for 1000 times, which would generate an output as close as possible to the target values of the output. The training of the IBPN model was accomplished with the no error fitting and the prediction average relative error was 8.07%. The results indicated that both prediction accuracy and the generalization ability were significantly improved.