Watershed calibration using multistart local optimization and evolutionary optimization with radial basis function approximation

Calibration of computationally expensive watershed models is more feasible with algorithms that require fewer simulations. This paper compares the performance of seven global optimization algorithms on a 14-parameter and an 8-parameter watershed calibration problem. The optimization algorithms include Shuffled Complex Evolution (SCE), Differential Evolution (DE), an evolutionary algorithm that uses Radial Basis Function (RBF) approximation (ESGRBF), and four types of local optimization methods coupled with the Multi-Level Single Linkage (MLSL) multistart procedure. The four local optimization algorithms are: Sequential Quadratic Programming, which is a derivative-based method; Unconstrained Optimization by Quadratic Approximation (UOBYQA), which is a derivative-free trust-region method; Pattern Search; and Implicit Filtering. The results indicate that ESGRBF is the most effective algorithm on the two calibration problems, followed by Implicit Filtering coupled with the MLSL multistart approach. Hence, this study provides some promising alternatives to the currently most widely used methods in watershed calibration, which did not perform as well.