Parameter estimation in distributed hydrological modelling: comparison of global and local optimisation techniques

Abstract

Much research has been spent in the last three decades in developing more effective and efficient automatic calibration procedures and in demonstrating their applicability to hydrological problems. Several problems have emerged when applying these procedures to calibration of conceptual rainfall -runoff and groundwater (GW) models, such as computational time, large number of calibration parameters, parameter identifiability, model response surface complexity, handling of multiple objectives and parameter equifinality. All these are expected to be much more severe for more complex models, for which comprehensive calibration studies have not so far been conducted. The scope of this paper is to investigate the performance of a global and a local optimisation technique, respectively, the Shuffled Complex Evolution algorithm and the gradientbased Gauss -Marquard -Levenberg algorithm, in calibration of physically based distributed models of different complexity. The models considered are a steady-state GW model, a transient GW model and a fully integrated model of the same catchment. The calibration is conducted in a multi-objective framework where two different aspects of the model response, the simulated runoff and the groundwater elevation are aggregated and simultaneously optimised. Different aggregated objective functions are used to give different weights to the calibration criteria. The results of the calibration procedures are compared in terms of effectiveness and efficiency and demonstrate the different performance of the methods. Moreover, a combination of the global and local techniques is investigated as an attempt to exploit the advantages of both procedures, while overcoming their drawbacks.

Solution of the antenna placement problem by means of global optimization techniques

“…They generally follow populationbased search strategy. Therefore, they start searching by generating an initial population rather than a single solution as in the case of local optimization [10].…”

Uncertainty assessment of integrated distributed hydrological models using GLUE with Markov chain Monte Carlo sampling

“…Global calibration techniques, such as the Shuffled Complex Evolution (SCE) algorithm by Duan et al (1992), have been successfully applied in some calibration studies of distributed models (Madsen, 2003;Mertens et al, 2004;Blasone et al, 2007a). To reach convergence, global methodologies require a larger number of model runs than local gradient-based techniques, which have also been employed in conjunction with this type of models (Sonnenborg et al, 2003;Blasone et al, 2007a). On the other hand, it has been demonstrated that local procedures have a high probability of converging to suboptimal solutions when they are applied to integrated, distributed models (Blasone et al, 2007a).…”

“…To reach convergence, global methodologies require a larger number of model runs than local gradient-based techniques, which have also been employed in conjunction with this type of models (Sonnenborg et al, 2003;Blasone et al, 2007a). On the other hand, it has been demonstrated that local procedures have a high probability of converging to suboptimal solutions when they are applied to integrated, distributed models (Blasone et al, 2007a).…”

Inverse Parametrization of a Regional Groundwater Flow Model with the Aid of Modelling and GIS: Test and Application of Different Approaches

“…Historically and under consideration of the available computational power, model calibration was manually performed by adjusting model parameters with a more or less trial and error process. This approach has been widely used for several complex regional models, e.g., [11][12][13]. However, manual calibration is time-consuming, tedious and very subjective in nature [14].…”

“…However, manual calibration is time-consuming, tedious and very subjective in nature [14]. The accuracy of results from this approach requires good experience of the modeller and thorough understanding of the system and is also characterized by the strategies employed to adjust the model parameters [13]. In recent times, manual calibration has been partly substituted by automated approaches, which are also recognized as nonlinear parameter estimation techniques [15].…”