Output updating of a physically based model for gauged and ungauged sites of the Upper Thames River watershed

Abstract

This study introduces a new ANN updating procedure of streamflow prediction for a physically based HEC-HMS hydrological model of the Upper Thames River watershed (Ontario, Canada). Besides streamflow and precipitation, the updating procedure uses other meteorological variables as inputs, which are not applied in calibration of the HEC-HMS model. All the results of performance measures on training, validation and test datasets for river gauges at Mitchell and Stratford revealed that the ANN updated models have performed better than the HEC-HMS model. The ANN model results were in excellent agreement with observed streamflow. The uncertainties can be associated with different input variables and different length of datasets used in the HEC-HMS model and the ANN model. The performance results suggest improvement in the RMSE values of the trained networks when additional meteorological data was used. The updated errors from the gauged sites of Mitchell and Stratford were used to update the streamflow values at the ungauged site of JR750 of the HEC-HMS model. While the underlying physical process in the ANN model consisting of interconnected neurons to map input-output relationships is not easily understood (in a form of mathematical equation), the HEC-HMS hydrological model can reveal useful information about the parameters of a hydrological process.

Keywords

• neural network
• bayesian
• hydrometeorology
• hydrological model
• hec-hms