Balancing exploitation and exploration: A novel hybrid global-local optimization strategy for hydrological model calibration

Abstract

Optimization problems in hydrological modeling are frequently solved using local or global search strategies, which either maximize exploitation or exploration. Thus, the elevated performance of one strategy for one class of problems is often offset by poor performance for another class. To overcome this issue, we propose a hybrid strategy, G-CLPSO, that combines the global search characteristics of the Comprehensive Learning Particle Swarm Optimization (CLPSO) with the exploitation capability of the Marquardt-Levenberg (ML) method and implement it into the hydrological model, HYDRUS. Benchmarks involving optimizing non-separable unimodal and multimodal functions demonstrate that G-CLPSO outperforms CLPSO in terms of accuracy and convergence. Synthetic modeling scenarios involving the inverse estimation of soil hydraulic properties are used to compare the G-CLPSO against the original HYDRUS ML solver, the gradient-based algorithm PEST, and the stochastic SCE-UA strategy. Results demonstrate the superior performance of the G-CLPSO, suggesting a potential use in other environmental problems.