用于预测煤燃烧CFD模拟的机器学习方法：从详细动力学到HDMR降
题目：用于预测煤燃烧CFD模拟的机器学习方法：从详细动力学到HDMR降阶模型-Fuel-Debiagi2020
摘要：由于固体燃料燃烧系统中存在复杂的多尺度湍流-化学-颗粒（ turbulence-chemistry-particle， TCP）相互作用，即使提高了大涡模拟（LES）方法的精度，开发预测模型仍然是一个艰巨的挑战。基于LES的煤炭燃烧模型主要以下三方面的工作：a）煤颗粒动力学的模型，b）亚网格TCP相互作用的模型，c）固体燃料反应动力学的模型。第三类是这项工作的重点，因为最近的一些研究表明，用于描述固-气动力学转化过程的反应动力学模型的准确性在可预测性方面至关重要。因此，在模拟煤燃烧时，需要实现详细的固体燃料反应动力学。然而，在大型电厂的锅炉中直接耦合详细的反应动力学模型还不现实。为了克服这一挑战，本文开发了一种基于机器学习（ML）的降阶模型，以可接受的计算成本准确地表示固相转化过程。ML模型的训练集是：在一个气体辅助煤炭燃烧室中提取的多种工况下，利用基于详细反应动力学模拟单个碳颗粒燃烧所建立的数据库，之后在测试集及气体辅助煤炭燃烧室LES模拟中提取的颗粒轨迹上进行测试。基于ML的模型可以以更低的计算成本准确预测煤炭颗粒燃烧的不同阶段。结果表明，基于ML的方法的使用有望在LES背景下实现详细的固体燃料动力学仿真。
Notes:
- HDMR: high-dimensional input–output relationship.
- 训练集构造方法：在Fig. 1中所示的燃烧室中提取了5*1000个颗粒的数据，得到颗粒加热速率、所在网格氧含量及最终温度的最大和最小值，从中进行插值，得到单颗粒仿真的边界条件，然后利用详细反应动力学对单个颗粒进行模型，得到训练集
- 在氧含量较低的情况下，ML模型的误差达到30%
Fig. Sketch of a generic multilayer perceptron of the HDMR method. Eq. (1) is determined by the glass box with a variance reduction method.
Fig.2 Instantaneous gas-phase temperature field on the section plane; individual particles colored by their temperature and sized proportionally to their radius. Only one in three particles is shown
Title: Machine learning for predictive coal combustion CFD simulations—From detailed kinetics to HDMR Reduced-Order models
Abstract: Because of the complex multiscale turbulence-chemistry-particle (TCP) interactions in solid fuel-supplied combustion systems, developing predictive models remains a formidable challenge even with the improved accuracy of the large eddy simulation (LES) approach. There are three main types of LES-based coal combustion model: a) those for coal particle dynamics, b) those for subgrid TCP interactions, and c) those for solid fuel kinetics. The third type is the focus of this work, as several recent studies have shown that the accuracy of kinetic models used to describe the solid-gas phase kinetic conversion process is of primary importance when it comes to predictability. Therefore, the implementation of detailed solid fuel kinetics is desired when simulating coal combustion However, it is far from being feasible to directly couple detailed kinetics in large-scale LESs of power plants. To overcome the challenge, a reduced-order model based on Machine Learning (ML) is developed in this work to accurately represent the solid-gas phase conversion process at an acceptable computational cost. The ML-based model is trained with databases from the simulations of single-particle combustion with detailed kinetics over a wide range of operating conditions extracted from a novel gas-assisted coal combustion chamber, and then validated by the test databases and unsteady particle trajectories from the LES of the gas-assisted coal combustion chamber. The ML-based model can accurately predict different phases of coal particle combustion at a reduced computational cost. The results indicate that the use of ML-based approaches is promising for implementing detailed solid fuel kinetics in the context of LES.
相关文献
