Development of a parallel computing enabled optimisation tool for hydrological model calibration

A river model is a semi-distributed hydrological model and it includes many processes such as flow routing, irrigation diversion, overbank flow, ground water interaction for simulating flows a river system for water resources planning and management. A number of calibration parameters are introduced in such models to represent various processes using simplified mathematical equations. Traditionally, a river model is calibrated using a reach-by-reach calibration approach starting from the top of the system cascading down to the end of the system. While the reach-by-reach approach is suitable for obtaining optimum model performance at a single river reach with high quality observed data, it does have the limitation of error propagation from upstream to downstream reaches if poor quality data are used in the calibration. A system-wide calibration approach has recently been developed for river system modelling in large river basins. Comparing with traditional reach-by-reach calibration, this new method optimises parameters of all river reaches within a region simultaneously using a weighted global objective function. The results of its application of this new approach in the Murray-Darling basin, Australia have shown its potential to overcome over-fitting and improve fitness of each individual gauge. However, due to the system-wide optimization of multiple reach parameters in a region, the search space and computational time required for system calibration increase exponentially with the increase of number of parameters. This limits the number of parameters that can be optimised and thus, the size of the region. To potentially overcome this limitation, a parallel computing enabled shuffled complex evolution (SCE) optimisation tool has been developed. A series of comparison studies have been conducted to evaluate the performance of this approach over normal SCE. These are: 1) comparison of computation time and performance for the same number of parameters; 2) comparison of performance with the same computation time and the same number of parameters and 3) comparison of the maximum number of parameters that can be optimised and performance within the same computation time. The results show that the run time with the new approach is about 25% of those with the normal SCE and its efficiency increases with increased number of calibration parameters.