Artificial-intelligence-Assisted Analysis of Flares and Fugitive Gases

Abstract

Fugitive gas is the gas emission to the atmosphere that results from gas and oil activities, in which, flare is often used as a safety measure to intentionally burn the related gases so thus to avoid potential explosion. In addition to their use, flare burns hundreds of million cubic meters of natural gases per year, which causes a significant energy loss, and releases a massive amount of greenhouse gases, such as CO2, which also heavily contribute to global warming. Thus, for the purpose of better monitoring and controlling these emissions, the development of advanced, smart, and convenient gas analysis tools emerges as a pressing technical necessity.

Line of sight spectroscopy and Optical Gas Imaging (OGI) are two often used tools in noninvasive gas analysis. Limited by their underlying principles, conventional physics-based solutions cannot support the use of line-of-sight spectroscopy in turbulent flows, which is responsible for non-uniform temperature and concentration profiles along the light path. Meanwhile, the information contained inside OGI is difficult to be extracted, understood, and utilized for qualitative and quantitative analysis, let alone, be utilized for the purpose of process control. The emergence of modern Artificial Intelligence (AI) methods has shed light in such challenges, and provides capabilities to develop new tools. For this possibility to realize, a series of exciting scientific problems lying at the intersection of spectroscopy/OGI-based gas analysis and AI have to be addressed.

This dissertation proposes novel theoretical solutions and demonstrates their performance within the context of flares and fugitive gases analysis. Considering the reliability issue of utilizing supervised learning algorithms in estimating temperature and concentration information from spectroscopic data, an intelligent optimization paradigm is developed, in order to provide both efficient and effective estimations. Two algorithms, SVPEN and EEE, are respectively proposed, in which, SVPEN exploits the original pretrained model architecture, while EEE realizes a better optimization efficiency and convergence. In addition, the feasibility of quantifying average temperature from laser absorption spectroscopy measured from nonuniform profiles is proved, through a machine-learning-based data analysis method. Moreover, an approach to reconstruct temperature profiles from emission spectroscopy acquired from nonuniform profiles is proposed, which is shown to outperform state-of-the-art methods by a considerable margin. As for OGI-based analysis, few-shot learning is introduced for the first time into the qualitative analysis of combustion, which includes flares, which results in a massive reduction of the needed data amount. Subsequently, a number of advanced optimization algorithms, such as RRAM and contrastive learning, are proposed to improve the performance of quantitative analysis of combustion. An important outcome is that the novel contrastive learning-based framework can realize almost unbiased quantification.

The work reported in this dissertation substantially improves the capability and quality of AI-assisted analysis of flares and fugitive gases, and offers a solid theoretical foundation for further development within advanced, smart, and easy-to-use emission analysis tools.

Date of Award	Apr 2023
Original language	American English
Supervisor	Dimitrios Kyritsis (Supervisor)