A method for reverse-inferring fuel gas composition from flue gas information based on residual network and physical constraints

The gas quality of different gas sources fluctuates, and direct combustion will adversely affect the operation of gas equipment. This paper proposes a method for reverse-inferring fuel gas composition from flue gas information based on residual network and convex optimization (ResNet-CVX), achieving data and physical driving. The method takes flue gas information as input to achieve real-time prediction of methane, ethane, propane, and butane content in fuel gas. The data in this paper are from the energy efficiency test reports of boiler products. Data-driven experiments results show that the maximum absolute prediction errors of the model for CH4, C2H6, C3H8, and C4H10 are 1.75 vol%, 1.25 vol%, 0.7 vol%, and 0.15 vol%, respectively. The model output satisfies physical constraints such as component content and carbon conservation constraints, and no error propagation or interval variation due to inter-component constraints is observed. The relative error of the calorific value calculated from the output fuel gas components is controlled within 1.65 %. Compared with the ResNet, it exhibits a lower RMSE and a higher R2. This indicates that ResNet-CVX has stronger predictive capabilities, superior generalization performance and physical authenticity. In summary, this method provides technical support for energy-saving modifications and stable operation of gas equipment.