Résumé:
The existed hydrological models that were used to describe the hydrological processes varied between the pure mathematics analysis and equations to relations that describe the process empirically based on the physics of the watershed. An oscillation exists between the model’s performance and robustness but it was approved that such a hydrological model is chosen according to the study objective and perspectives. The models that are based on purely physical or conceptual, face the problem of calibrating the large number of parameters that contain. Furthermore, the major watershed measurement is scarce, especially in developing countries. To further understand and excavate this concept, we engaged in this study and went deeper into the hydrological models. We choose different models from both conceptual and Data-driven to analyze. The study area was selected inorder to verify the applicability of the NARX-NN model in small drainage areas, containing theconditions of the Semi-urbanism and the low intensity of rainfall, and then it will be compared to the conceptual HEC-HMS model. We investigate the NARX-NN model to simulate the basin response to real storm events of a small urban watershed. The NARX-NN steps are described in detail, from data representation and insertion, and then calibration and testing.To reach the target hydrograph by the conceptual HEC-HMS, we must pass through several stages. Conceptually, thehole R-R process is a combination of several sub-processes; we need to treat each one separately. We proposed a new method to carry out the best parameters of calibration using the weightedaverage function ofHEC-HMS models and that gave the best performance. The steps involvedin each model to reaching the hydrograph are dramatically different; the significance is to presentan analytical comparison step by step of each model. The two models are extremely differentin terms of data requirements. The conceptual and black box models are compared statisticallyand graphically. The statistical assessment shows that during calibration, the two models are closely the same in terms of NSE and MAE with a slight superiority of NARX-NN. Except for the PEP benchmark which indicates the supremacy of the NARX-NN over HEC-HMS in mapping the peak of the hydrograph, which is agreed as one of the most important elements in the short storm events. In the testing phase, statistical parameters indicate that the NARX-NN outperforms modeling the basin response
The findings drawn from the results demonstrate the high capability of the NARX-NN model to capture the R-R process, although the short time-series data is used in the training and testing phases. Moreover, the result of the testing is more accurate than that of the training phase, which demonstrates the strength of the generalization feature of NARX-NN. It has more strength to produce the shape bending of the hydrograph; in contrast, the rising and falling limb in the HEC-HMS model presents straight curvature. Consequently, the NARX-NN model is better to highlight the curvatures resulting from the local peaks of rainfall.
