Establishing a time‐varying flood‐wave impulse function combined with a dynamic machine‐learning technique in response to the disturbance of boundary conditions

Abstract

AbstractTo seek other alternative approaches besides numerical methods, the linearized Saint Venant equations were utilized to derive the channel response functions for simulating streamflow. This study advocates developing a new approach to make the temporal distribution of response functions more flexible by introducing time‐varying reference parameters which depend on both the upstream inflow and the downstream boundary conditions. Moreover, to expand the model applicability in natural channels with irregular cross‐sectional shapes affected by the unsteady inflow, lateral flow, and the variation of tide level, a dynamic neural network algorithm is jointly applied to determine more appropriate time‐varying hydraulic parameters which are adopted in the channel flow response functions respectively derived from upstream and downstream boundary conditions. The tidal level at the estuary and the upstream inflow discharge are simultaneously considered influential factors to determine an optimal set of reference parameters by applying a machine learning technique, and this prediction can further be provided for updating the channel response functions. The novelty of this study is to propose a complete methodology to combine channel response functions with the machine learning algorithm for ameliorating the accuracy of channel flow simulation and resolving the uncertainty of model parameters. Therefore, by using the proposed method, not only the physical mechanism of Saint Venant equations can be preserved, but also the optimal hydraulic parameters can be specified. The problem of numerical instability during channel flow routing can be eliminated by using the proposed new model and therefore the reliability of real‐time flood forecasting can be reinforced.