Accuracy of HEC-HMS and Artificial Neural Network models in simulating runoffs in upper valley of the Medjerda-Tunisia

Abstract

Study region

The Ghardimaou-Jendouba section of the upper Medjerda valley watershed is located in the extreme north-west of Tunisia. The upstream section of this river has special topographical and hydrographical features that make it particularly vulnerable to flooding.

Study focus

This study aimed to replicate daily streamflow historical records using two distinct modeling approaches: the HEC-HMS and Artificial Neural Network (ANN) models. The effectiveness of both models was rigorously evaluated during their calibration and validation phases using key statistical metrics, namely Root-Mean-Square Error (RMSE), coefficient of determination (R2), and Nash–Sutcliffe efficiency (NSE).

New hydrological insights for the region

Model performance varied significantly. Post-validation, the HEC-HMS model yielded R2, NSE, and RMSE values of 0.3668, 0.573, and 0.664, respectively. In contrast, the ANN model ([2−4−1] architecture) showed substantially superior calibration performance: R2 of 0.978, NSE of 0.979, and RMSE of 8.46. These statistics unequivocally point to the ANN model's superior predictive capability. Further analysis revealed HEC-HMS overestimates low flows and underestimates high flows. Conversely, the ANN model accurately estimated both extreme and general flow conditions. This highlights the ANN model's strong potential for precise streamflow forecasting and water resource management in the Ghardimaou-Jendouba Basin. Future studies should compare other advanced machine learning models against HEC-HMS to refine daily streamflow forecasts.