CATCHMENT MODEL CALIBRATION SCHEMES

Abstract

Three catchment model calibration schemes were presented in this study. They are: (1) Fractional Factorial Design (FFD) based response surface calibration scheme, and its improved version, Fractional Factorial and Central Composite Designs (FFD-CCD) based response surface calibration scheme; (2) Neural Network (NN) based response surface calibration scheme; and (3) Genetic Algorithm (GA) based calibration scheme. The catchment model used in the study is the Storm Water Management Model, SWMM (Huber et al.,1984). The calibration schemes were tested on a catchment in Singapore, the Upper Bukit Timah catchment The number of calibration parameters considered in the study is 8. The results were compared with those from another catchment model calibration scheme, the Factorial and Central Composite Designs based response surface calibration scheme (Chua, 1992; Ibrahim and Liong, 1992; Liong and Ibrahim, 1994). The comparisons were made on the basis of: (1) the prediction accuracy in the peak flow and the runoff volume; and (2) the number of simulations or the time required for the calibration process. In the FFD and FFD-CCD based response surface calibration schemes, two response surfaces were constructed. One of them serves to relate the peak flow with the eight calibration parameters considered while the other response surface is for the runoff volume. An optimality search on each of these response surfaces yields the optimal values of the calibration parameters for the corresponding objective function. Results show that the peak flow and the runoff volume prediction errors are relatively low. Compared to their counterpart, the FD-CCD based calibration scheme, the prediction errors resulting from the FFD are slightly larger while those from the FFDCCD are almost identical. It should be noted, however, that FFD and FFD-CCD based calibration schemes require only 64 and 81 simulations, respectively, as opposed to 273 simulations required by the FD-CCD based calibration scheme. In the Neural Network based response surface calibration scheme, the 273 data sets generated by SWMM in the FD-CCD based calibration scheme were used to train the neural network (NN). The trained NN was then used to simulate the necessary 273 data sets required in the FD-CCD based calibration scheme. Although some of the prediction errors resulting from the FD-CCD with NN based calibration scheme are comparable with those of FD-CCD with SWMM based calibration scheme, in general the results are not satisfactory. Better rainfall representations for the nodes in the input layer have to be identified. The study, however, shows a promising application of the NN in this area. In the Genetic Algorithm (GA) based calibration scheme, the prediction accuracy is definitely comparable to, if not higher than, those of FD-CCD calibration scheme. In addition to the high prediction accuracy of GA, GA based calibration scheme requires only 50 catchment model simulations as opposed to the 273 simulations required in the FD-CCD based calibration scheme.