FLOOD ROUTING USING ARTIFICIAL NEURAL NETWORKS

摘要

Artificial neural networks (ANNs) provide a quick and flexible means for simulating flood flows. However, for ANNs to have some relationship with physical processes in flood flow transmission, their network architecture should correspond to well-established numerical models. This paper presents an ANN based approach for flood routing through river reaches. An ANN architecture was formulated resembling the finite difference solution of the Muskingum method. The network weights are equivalent to routing coefficients C_1, C_2 and C_3 and bias accounted for lateral inflows. The network was trained for the inflow-outflow time series data using the backpropagation Levenberg-Marquardt algorithm. The ability of the network to replicate the Muskingum scheme was analysed in terms of coefficients K and X. The generalisation capability of the trained network was tested using untrained data sets. Means of improving the network performance such as passing of inflows through a non-linear transfer function were explored. A standard multiplayer perceptron network was also trained for the comparative analysis of the performances. A case study from the Neckar River in Germany demonstrates the application using historical flood data sets. The results of the study show that the network can be trained to resemble the Muskingum scheme. The trained network was found to provide highly acceptable performance, comparable to standard neural networks.