A threshold artificial neural network model for improving runoff prediction in a karst watershed

Artificial neural network model (ANN) has been extensively used in hydrological prediction. Generally, most existing rainfall-runoff models including artificial neural network model are not very successful at simulating streamflow in karst watersheds. Due to the complex physical structure of karst aquifer systems, runoff generation processes are quite different during flood and non-flood periods. In this paper, an ANN model based on back-propagation algorithm was developed to simulate and predict daily streamflow in karst watersheds. The idea of threshold was introduced into artificial neural network model [hereafter called Threshold-ANN model (T-ANN)] to represent the nonlinear characteristics of the runoff generation processes in the flood and non-flood periods. The T-ANN model is applied to the Hamajing watershed, which is a small karst watershed in Hubei Province, China. The network input, the previous discharge, is determined by the correlative analysis, and the network structure is optimized with the maximum Nash coefficient as the objective function. And the precipitation and previous discharge are chosen as the threshold factors to reflect the effect of specificity of karst aquifer systems, respectively. By using the T-ANN, the simulation errors of streamflow have been reduced, and the simulation becomes more successful, which would be helpful for runoff prediction in karst watersheds.