CO reduction in sintering flue gas by CFD-ML for process parameters optimization

To solve the problem of high CO concentration in sintering flue gas, a combined Computational Fluid Dynamics and Machine Learning (CFD-ML) method is presented in this study for predicting CO concentration in sintering flue gas and optimizing the process parameters to reduce CO emission. Firstly, numerical simulations were conducted to investigate the primary factors influencing CO emissions during the sintering process. An increase in carbon content from 3.2% to 4.0% was accompanied by a corresponding increase in CO emissions from 3222.06 ppm to 9651.77 ppm. Conversely, an increase in coke particle size from 1.2 mm to 2.4 mm resulted in a decrease in CO emissions from 8278.55 ppm to 4765.12 ppm. An increase in bed thickness from 600 mm to 800 mm resulted in a 4.83% reduction in CO emissions. However, an increase to 1000 mm resulted in a 15% increase in CO emissions. The optimal SFGR temperature of 250°C resulted in CO levels of 6513.61 ppm, whereas higher temperatures led to a corresponding increase in CO emissions. Correlation analysis based on CFD calculations revealed a strong correlation between carbon content, coke particle size, SFGR temperature, bed thickness, and average CO concentration. Subsequently, nine machine learning algorithms were employed to construct a CO concentration prediction model. Following a comparative analysis, it was determined that the Gradient Boosted Regression Tree (GBRT) exhibited satisfactory performance on the test set. ( R 2 = 0.988, MAE = 0.019, MSE = 0.00056). SHAP analysis of the GBRT model indicated that the order of contributions to average CO concentration prediction was carbon content > coke particle > SFGR temperature > bed thickness. Finally, particle swarm optimization (PSO) integrated with the ML model determined the optimal parameters: 3.0% carbon content, coke particle size of 2.98 mm, bed thickness of 682 mm, and SFGR temperature of 286.77°C, projecting a minimum CO concentration of 3374.48 ppm. The accuracy of the prediction model was validated against CFD calculations, which yielded a 0.68% error. • A two-dimensional mathematical model was developed to simulate the transient variation of CO content in sintering flue gas. • Gradient boosting regression tree the sintering flue gas CO concentration prediction model with better performance compared to other machine learning algorithms. • CO concentration in sintering flue gas can be quickly predicted in seconds using a trained machine learning model. • The process parameters were optimized by a particle swarm algorithm combined with a machine learning model.