CS(AR)-11/97-98 : Calibration of a catchment model using SCE-UA algorithm

Abstract

An important step in application of a conceptual model to a catchment is model calibration. The objective of a calibration is to determine the model parameters such that an acceptable match is obtained between the observed and the computed discharge hydrographs. Two approaches are followed for calibration of a conceptual model -- manual using trial and error and automatic using an optimization algorithm.

The aspects of the conceptual models which cause problems during automatic calibration are : 1) Interdependence between the model parameters. 2) Indifference of the objective function to the values of the inactive parameters, 3) Discontinuities of the response surface, and 4) The presence of local optima.

The degree of complexity of model plays a significant role in model calibration phase. The difficulties encountered during the calibration are closely connected to the number of parameters typical of the model and to the greater or lesser ease of visualizing the various parameters. There are differing views in the literature about what constitutes the adequate data for model calibration. Both the right kind (the data which activate all the model parameters) and right duration of data are needed for a good calibration. Besides ensuring that the data are error free, one has to be careful about the duration of the calibration period also, the periods of extreme events should be suitably incorporated.

Duan et al. (1992) have presented a new algorithm, named Shuffled Complex Evolution Method (SCE-UA). It has been claimed that this method has a very high probability of finding the global optima. The main aim of this report was to apply the SCE-UA algorithm for the calibration of a CRR model. The model reported by Jain (1993) was used to simulate the response of a basin of size 820 km². The results show that the algorithm is able to converge to the global optimum when the computations are started from a number of initial points. The main conclusion of the study is that the SCE-UA algorithm is a global optimization method and is able to converge to the global optimum parameters when different initial values of parameters are used. The computational requirements for calibrating a CRR model are quite reasonable and thus the algorithm is computationally efficient.