Abstract: In order to describe and interpret hydrological processes in more detail, and at the same time to construct a more accurate distributed hydrological model, we took the Karup watershed in Denmark as an example and calibrated three parameters of MIKE SHE model, namely, saturated hydraulic conductivity, saturated horizontal hydraulic conductivity, and leakage coefficient of river bank, and simulated the daily runoff process in the watershed. Results demonstrate that the root mean square error (RMSE) obtained by the method of parameter calibration based on BP neural network is smaller than that by automatic parameter calibration in MIKE SHE model, with the model efficiency coefficient Ens closer to 1. Having been treated by parameter calibration by BP neural network, the values of RMSE of daily runoff of three test samples are 0.04 m³/s, 0.03 m³/s, and 0.08 m³/s, respectively, and the value of Ens is 0.99. As the simulated runoff displays a trend in agreement with the real runoff, the back analysis method of parameter calibration based on BP neural network is of certain value in runoff simulation.

Key words: runoff simulation, parameter calibration, MIKE SHE model, BP neural network, back analysis, uniform design, Karup watershed in Denmark