More Efficient Bayesian-based Optimization and Uncertainty Assessment of Hydrologic Model Parameters

摘要

An important consideration in assessing the performance of model calibration software is that of run time. Minimizing the number of hydrologic model runs required during the calibration process is nearly always important, but particularly when the objective function landscape contains multiple local minima or hydrologic model run times are high. Minimizing the number of required model runs was one of the primary factors driving the research and development activities encapsulated in this report, such that the resulting optimization and uncertainty tool(s) are more compatible with the computationally expensive physics-based models that are becoming more commonly used within the practice community. SCEM-FA is a modified version of the Markov Chain Monte Carlo sampler SCEM-UA. It is more efficient than the native SCEM-UA algorithm, through employment of function approximation, while effectively inferring the posterior parameter distribution of model parameters and also the most likely parameters within this high probability density region. Based on a summary of thirty random trials, SCEM-FA was able to infer, effectively, the same posterior probability distribution for thirteen SAC-SMA hydrologic model parameters as that of SCEM-UA with an average twenty-one percent savings in total forward model calls.