Assessing model calibration adequacy via global optimization

An assessment of the application of varying levels of optimization on model simulation performance and parameter identification was done using the genetic algorithm (GA) and a 10-parameter version of the MODHYDROLOG rainfall-runoff model. Four levels of optimization were obtained through the use of two GA formulations, the traditional and an improved GA, and by varying the optimization parameters with each formulation. Sixteen years of data from a 27 km2 Australian catchment was used. With each level, ten randomly initialized optimization runs were made. The differences in simulation performance quantified by the coefficient of efficiency, bias, absolute deviation and a residual mass curve coefficient were not considerable although the performance improved as the level of optimization effort increased. Superior parameter identification, and consequently a better detection of parameter correlations was achieved with the higher optimization levels. Based on the objective function values, the highest level of optimization practically located the global optimum in all the ten runs. The second level achieved this in nine of the ten runs while the lower two levels did not locate the global optimum in any of the ten runs. It is proposed that the systematic verification of the adequacy of optimization should be an integral part of model calibration exercises. The form of verification should depend on the specific problem at hand.