A novel modeling strategy for the prediction on the concentration of H₂ and CH₄ in raw coke oven gas

The composition of raw coke oven gas, mainly consisting of H₂ and CH₄, has a significant impact on its utilization in the industry. To better utilize this product, a novel modeling strategy is presented in this work to predict the concentration of H₂ and CH₄. Initially, both thermodynamic- and kinetic-based models are established to simulate the coal coking process in Aspen Plus. Following this, a sample pseudo-data set is generated from the kinetic-based model to better describe the coal coking process, and this is followed by data calibration using actual operating data from the industry. Based on these calibrated simulation data and the collected actual operating data, a machine learning model is developed to predict the concentration of H₂ and CH₄. In this work, four well-known and high-performance algorithms (i.e., artificial neural network, Random forest, XGBoost, and LightGBM) are used and compared in the model development. LightGBM provides the best modeling performance, with the coefficient of determination (R²) on H₂ and CH₄ being 0.99952 and 0.99964, respectively. Furthermore, the Shapley Additive exPlanations (SHAP) technique is employed to identify the ranking of key parameters that have a major impact on the concentrations of H₂ and CH₄.