A method for inversing natural gas composition from flue gas information based on neural networks

With the increasing globalization of the natural gas market, regional variations and fluctuations in natural gas composition have become increasingly prominent. Such compositional variations alter the physical and combustion properties of natural gas, which can induce combustion instability in gas-fired equipment such as boilers and gas turbines. However, the industry still lacks effective, high-precision online monitoring technologies for natural gas composition, which are urgently needed to realize real-time feedback and targeted combustion control of gas-fired facilities. In this study, we designed and constructed a natural gas combustion experimental system, and conducted 237 groups of combustion experiments with systematically varied gas mixing ratios based on the typical compositional ranges of commercial natural gas and liquefied natural gas (LNG). A database was established covering four key alkane components (CH₄, C₂H₆, C₃H₈, C₄H₁₀) and corresponding flue gas monitoring parameters (CO, CO₂, NO_x, etc.). Spearman rank correlation analysis was performed to eliminate redundant input variables and optimize the neural network architecture. A data-driven model for natural gas composition inversion from flue gas measurements was developed using a back-propagation (BP) neural network, and its reliability and feature contribution mechanism were verified via SHapley Additive exPlanations (SHAP) interpretability analysis. The proposed method enables real-time online inversion of natural gas composition using routine flue gas monitoring data.