Évaluation de méthodes d'optimisation pour le calage efficace de modèles hydrologiques coûteux en temps de calcul

Abstract

Many studies have shown the usefulness of blackbox optimization algorithms for the calibration of lumped conceptual hydrological models with low computational costs. Among these algorithms, the « Shuffled Complex Evolution method developed at the University of Arizona » (SCEUA) is a very popular one. However, when it comes to calibrating distributed and/or physically-based models, computational efficiency becomes an issue. A single simulation with this type of model may take several minutes and the optimization process may require more than thousands of simulations. Therefore, the efficiency of other optimization algorithms needs to be studied. Two recently developed and potentially more efficient optimization methods, « Dynamically Dimensioned Search » (DDS) and « Mesh Adaptive Direct Search » (MADS), are more closely examine. This work aims to verify the computational efficiency of DDS and MADS for the calibration of the HYDROTEL model (distributed and physically-based). Two versions of the model are used, one with 10 parameters and one with 19 parameters, and they are both applied to two different watersheds located in the province of Quebec (Canada). A second, lumped conceptual model (HSAMI) is also applied to both watersheds to examine the impact of model structure and spatial discretization of the basins on the results. Each combination of model-watershed is calibrated with each one of the optimization algorithms: DDS and MADS. Different functionalities available with the use of MADS are also examined. A third algorithm, SCEUA, is also used as the benchmark for comparison. The objective function uses the Nash-Sutcliffe Efficiency criterion, and is computed between simulated and observed streamflows at the outlets of the watersheds. For every combination of ‘modelwatershed-algorithm’, calibrations are repeated 32 times and the mean results are shown. The results show that the DDS algorithm offers significant potential for reducing the number of model evaluations (computational cost), and this is observed for each ‘model-watershed’ combination. DDS has the ability to globally explore the parameter space and this characteristic makes it a dominant optimization method in comparison to other approaches, in terms of efficiency. MADS comes second for the calibration of the 10 parameters version of HYDROTEL, but seems to have difficulty when the number of parameters increases (19 parameters HYDROTEL and 23 parameters HSAMI). SCEUA then outperforms MADS in terms of efficiency for these two models with higher numbers of parameters. This study also shows that the type of hydrological model has an impact on the behavior of the optimization algorithms. In addition, two configurations of MADS show a good potential for reducing the number of model evaluations compared to the default version of the algorithm. Finally, although the parameter sets found by MADS can satisfy the optimality conditions, DDS provides better quality parameter sets. This characteristic of MADS is not necessary for efficient calibration, but it remains interesting from an optimization point of view.

Date	2 Sept 2014
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Annie Poulin (Supervisor) & Stéphane Alarie (Co-supervisor)