The design of cleaner and more sustainable combustion technologies represents nowadays a key task. Reliable numerical models, able to cope with a large variety of configurations, combustion processes and fueling mixtures are needed, especially for future applications in combustion monitoring and control, for thermal and environmental performances, which are of critical importance. In this work, alternative, low-computational cost modelling tools for pollutants and thermal efficiency predictions, represented by Chemical Reactor Networks (CRN), are designed from Computational Fluid Dynamic (CFD) simulations assessing a novel methodology, by exploring new possibilities offered by Machine Learning (ML) algorithms. In particular, unsupervised learning approaches are employed, in order to extract the key features of the system flow-field, adopting advanced clustering algorithms, such as Local Principal Component Analysis (LPCA) and K-Means, thus providing an efficient and automatic identification of similar thermo-chemical state compartments in the computational domain. The identified zones are modelled in a post-processing phase as a network of interconnected chemical reactors, and detailed kinetic mechanisms are employed for low concentration pollutants predictions. The case study, a quasi-industrial, flameless-capable combustion furnace, fed with methane-hydrogen mixtures in different compositions at a nominal power of 15 kW, has been investigated numerically by performing 2D CFD simulations with reduced chemistry and subsequently CRN simulations has been carried out with detailed kinetics, adopting the aforementioned approach. Results are validated upon experimental data, in order to provide a novel methodology for CRN design applications, which can be suited for future GTs applications.
Automatic extraction of Chemical Reactor Networks from CFD data via advanced clustering algorithms
Alberto Cuoci;
2021-01-01
Abstract
The design of cleaner and more sustainable combustion technologies represents nowadays a key task. Reliable numerical models, able to cope with a large variety of configurations, combustion processes and fueling mixtures are needed, especially for future applications in combustion monitoring and control, for thermal and environmental performances, which are of critical importance. In this work, alternative, low-computational cost modelling tools for pollutants and thermal efficiency predictions, represented by Chemical Reactor Networks (CRN), are designed from Computational Fluid Dynamic (CFD) simulations assessing a novel methodology, by exploring new possibilities offered by Machine Learning (ML) algorithms. In particular, unsupervised learning approaches are employed, in order to extract the key features of the system flow-field, adopting advanced clustering algorithms, such as Local Principal Component Analysis (LPCA) and K-Means, thus providing an efficient and automatic identification of similar thermo-chemical state compartments in the computational domain. The identified zones are modelled in a post-processing phase as a network of interconnected chemical reactors, and detailed kinetic mechanisms are employed for low concentration pollutants predictions. The case study, a quasi-industrial, flameless-capable combustion furnace, fed with methane-hydrogen mixtures in different compositions at a nominal power of 15 kW, has been investigated numerically by performing 2D CFD simulations with reduced chemistry and subsequently CRN simulations has been carried out with detailed kinetics, adopting the aforementioned approach. Results are validated upon experimental data, in order to provide a novel methodology for CRN design applications, which can be suited for future GTs applications.File | Dimensione | Formato | |
---|---|---|---|
GPPS-TC-2021_paper_349.pdf
accesso aperto
Descrizione: articolo principale
:
Publisher’s version
Dimensione
1.19 MB
Formato
Adobe PDF
|
1.19 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.