Hydrologic model parameter optimisation

Deterministic rainfall-runoff models require parameter calibration with the aim of matching the modeled streamflow record to an observed record as closely as possible. Much attention has been focused on automated methods of model calibration using meta-heuristics, such as genetic algorithms. In this study, two meta-heuristic algorithms were compared: the shuffled complex evolution (SCE) algorithm and a real coded genetic algorithm (GA). The GA was modified with a fitness scaler based on a Boltzmann distribution, which was used to adjust the level of elitism of the selection operator. This was found to improve the performance of the GA. The algorithms were compared using a range of configurations on two test problems - the two parameter Rosenbrock function and the six parameter Hartman function. A relatively small number of objective function evaluations were allowed so as to test the ability of the functions to converge under adverse conditions (that is, limited computational runtime). These functions were found to be best solved with different strategies. The Hartman function was found to respond better to a more elitist strategy, and the Rosenbrock function was solved with a more egalitarian strategy with higher population diversity. On both functions, the best configuration of the SCE algorithm was found to perform favorably compared to the best configurations of the GA.

The best performing configurations of these algorithms were applied to the calibration of a rainfall runoff model of a catchment in Tasmania. The model was a two-tap Australian Water Balance Model (AWBM) consisting of eight parameters, with 10 years of data and a six month warm up period. 20 trials were performed on each algorithm, and in each trial the data were randomly split into train and test sets of nine and one year’s data respectively. The more egalitarian strategies (which maintain higher population diversity, as taken from the algorithm configurations found to best suit the Rosenbrock function) were found to be favorable for both the SCE and the GA. When comparing results on the test data, this difference was marginal. The difference in performance was more apparent on the results of the SCE algorithm.

A recommendation of this paper is to develop a catalogue of the performance of several algorithms under different configurations on a range of test objective functions. This could be used to assist in the configuration of meta-heuristics on a range of applied problems, such as the calibration of additional rainfall runoff models.