Automatic calibration of conceptual rainfall-runoff models: Optimization algorithms, catchment conditions, and model structure

From 32 CRR-catchment cases (combinations from four conceptual rainfall-runoff models (CRR) and eight catchments) calibrated with either two or three optimization methods, (1) the shuffle complex evolution method (SCE-UA), (2) the multiple start Simplex (MSX), and (3) the local Simplex, it seems that all three methods produced parameter sets of comparable, local-optimum quality. Even with comparable performance among the models, some parameter values derived by the three optimization methods for the same CRR-catchment cases are surprisingly different from each other. In addition, parameter sets of SCE-UA or MSX, which often produce marginally better results than the local Simplex at the calibration stage, could end up with worse results at the validation stage. Apparently, given the inherent limitations of calibration data, model inadequacies, and identifiability problems, it is impossible to achieve global convergence in the parameter search. However, other than those for dry catchments such as Ihimbu or Bird Creek, the parameter sets obtained are generally feasible. Both SCE-UA and the local Simplex are viable optimization tools, while MSX is inefficient computationally. SCE-UA can complete the parameter search in one run, while the local Simplex often requires multirun operations to get good results.