Reducing the computational cost of automatic calibration through model preemption

Abstract

[1] Computational budget is frequently a limiting factor in both uncertainty-based (e.g., through generalized likelihood uncertainty estimation (GLUE)) and optimization-based (e.g., through least squares minimization) calibration of computationally intensive environmental simulation models. This study introduces and formalizes the concept of simulation model preemption during automatic calibration. The proposed "model preemption" method terminates a simulation model early to save computational budget if it is recognized through intermediate simulation model results that a given solution (model parameter set) is so poor that it will not benefit the search strategy. The methodology proposed here is referred to as deterministic model preemption because it leads to exactly the same calibration result as when deterministic preemption is not applied. As such, deterministic preemption-enabled calibration algorithms which make no approximations to the mathematical simulation model are a simple alternative to the increasingly common and more complex approach of metamodeling for computationally constrained model calibration. Despite its simplicity, the deterministic model preemption concept is a promising concept that has yet to be formalized in the environmental simulation model automatic calibration literature. The model preemption concept can be applied to a subset of uncertainty-based and optimization-based automatic calibration strategies using a variety of different objective functions. Results across multiple calibration case studies demonstrate actual preemption computational savings ranging from 14% to 49%, 34% to 59%, and 52% to 96% for the dynamically dimensioned search, particle swarm optimization, and GLUE automatic calibration methods, respectively.

Calibrating Environment Canada's MESH Modelling System over the Great Lakes Basin

“…Percentage Bias: This number also seems to work well according to some previous studies with MESH (e.g., Dornes et al, 2008;Razavi et al, 2010). The model calibration was conducted using DDS as implemented in the Optimization Software Toolkit for Research Involving Computational Heuristics (OSTRICH; Matott, 2005), which is a model-independent calibration and optimization tool consisting of a number of popular optimization algorithms including DDS.…”

An 825-year long varve record from Lillooet Lake, British Columbia, and its potential as a flood proxy

“…For example, Sander et al (2002) documented a 2000-year-long record of varves from an estuary in central Sweden and concluded that varve thickness significantly correlates with maximum annual daily discharge. Czymzik et al (2010) studied a sequence of varves in Lake Ammersee in southern Germany and compared it to the record of Ammer River floods. Other researchers have examined the seasonality and frequency of flooding in the southern European Alps (Wirth et al, 2013) and the Austrian Pre-Alps (Swierczynski et al, 2012).…”

A Monte-Carlo-Based Method for the Optimal Placement and Operation Scheduling of Sewer Mining Units in Urban Wastewater Networks

“…Indeed, the optimization algorithm may need to call the simulation model hundreds or thousands of times so as to accomplish a good solution, depending on each problem's parameters and restrictions. In order to address this issue, several optimization approaches have been developed aiming to minimize the computational burden, such as parallel computing and surrogate modeling techniques [46][47][48].…”