Abstract: To reduce CO₂ emission intensity in blast furnace-basic oxygen furnace（BF-BOF） route steel production, this study develops an ensemble machine learning methodology integrating feature selection mechanisms with intelligent optimization algorithms. A multidimensional feature importance evaluation system was established based on Spearman's rank correlation coefficient. Through comparative analysis of the predictive performance between the Levenberg-Marquardt Backpropagation（LM-BP） neural network（single hidden layer with 8 nodes） and the Particle Swarm Optimized Support Vector Machine（PSO-SVM） with RBF kernel parameters（σ=0.85, penalty factor C=0.1）, this study quantitatively evaluates the dynamic impacts of process parameter adjustments on carbon emissions. Results demonstrate that the PSO-SVM model achieves a determination coefficient of 0.994 3, significantly outperforming the BP neural network's 0.894 0, with its global optimization capability effectively addressing sample nonlinearity and dimensional sensitivity. Factor analysis based on the prediction model reveals: Comprehensive energy consumption per ton steel, sintering temperature, and coal injection ratio exhibit significant positive correlations with CO₂ emissions; Blast furnace volume, scrap steel ratio, and coke oven gas volume show negative correlations.