Assessing the Performance of Various Stochastic Optimization Methods on Chemical Kinetic Modeling of Combustion
Assessing the Performance of Various Stochastic Optimization Methods on Chemical Kinetic Modeling of Combustion
复制标题
DOI：
10.1021/acs.iecr.0c04009
复制
发表时间：
2020-10-28
影响因子：
4.2
通讯作者：
Pfaendtner, Jim
中科院分区：
文献类型：
--
作者：
Ashraf, Chowdhury;Pfaendtner, Jim
特定反应器配置的化学动力学模型的解通常由一组刚性常微分方程（ODE）组成，其中必须估计许多速率参数以适应实验观察。这提出了双重挑战：首先，准确有效地求解刚性 ODE 集；其次，通过优化目标函数来精确估计模型参数。近年来，用于参数估计的随机优化方法比经典优化方法更受欢迎，因为前者不需要合理的初始猜测并且具有逃避局部最小值的能力。在这项研究中，我们系统地检查了 10 种不同的随机优化算法，并评估了它们的性能，以估计先前开发的丙烷氧化机制（通常称为圣地亚哥机制）的模型参数。为此，我们开发了一个开源 python 包 kinexns，以使用 CVode 求解器高效求解动力学模型，执行灵敏度分析以确定重要的模型参数，并使用不同的随机方法优化模型参数。我们考虑的不同算法是蒙特卡罗（MC）、拉丁超立方采样（LHS）、最大似然估计（MLE）、马尔可夫链蒙特卡罗（MCMC）、混洗复杂进化算法（SCE-UA）、模拟退火（SA）、鲁棒参数估计（ROPE）、人工蜂群（ABC）、适应度尺度混沌人工蜂群（FSCABC）和动态维度搜索算法（DDS）。结果表明，MLE 和 DDS 在所有评估的算法中提供了更可靠的参数近似。
The solution to chemical kinetic models for a particular reactor configuration is usually composed of a set of stiff ordinary differential equations (ODEs), where a number of rate parameters have to be estimated to fit with experimental observations. This presents a twofold challenge-first, to solve the stiff set of ODEs accurately and efficiently, and second, to estimate the model parameters precisely by optimizing the objective function. In recent years, stochastic optimization methods for parameter estimation have gained popularity over the classical optimization methods as the former do not require a reasonable initial guess and have the capability to escape local minima. In this study, we systematically examined 10 different stochastic optimization algorithms and evaluated their performance to estimate the model parameters for the previously developed propane oxidation mechanism, popularly known as the San Diego mechanism. In doing this, we developed an open source python package kinexns to efficiently solve the kinetic model using CVode solver, perform sensitivity analysis to determine important model parameters, and optimize the model parameters by using the different stochastic methods. The different algorithms we considered are Monte Carlo (MC), Latin hypercube sampling (LHS), maximum likelihood estimation (MLE), Markov chain Monte Carlo (MCMC), shuffled complex evolution algorithm (SCE-UA), simulated annealing (SA), robust parameter estimation (ROPE), artificial bee colony (ABC), fitness scaled chaotic artificial bee colony (FSCABC), and dynamically dimensioned search algorithm (DDS). The results indicated that the MLE and DDS provide more reliable parameter approximation among all of the algorithms evaluated.