Hybrid model for predicting and optimizing energy consumption in converter processes based on feature selection and interpretability

To address the issues that energy consumption in the converter process is affected by multiple coupled factors, traditional prediction models suffer from feature redundancy and insufficient interpretability, and existing energy consumption optimization methods rely heavily on empirical decision-making, this paper proposes an energy consumption prediction and optimization method integrating Shapley Additive Explanations (SHAP), Support Vector Regression (SVR), and Grey Wolf Optimization (GWO). First, sample features including hot metal composition, temperature, and mass of scrap steel are extracted from the converter steelmaking process. Outliers are removed using box plots, followed by standardized preprocessing. Feature importance is then ranked via the maximum mutual information coefficient (MIC), Pearson correlation coefficient (PCC), and SHAP. Forward selection is employed to identify the optimal 7-feature subset, while grid search combined with K-fold cross-validation is used to optimize the hyperparameters of the SVR model. Experimental results show high predictive accuracy, with the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) on the test set reaching 3.12 kg/t, 2.51 kg/t, and 5.98%, respectively. Further global and local interpretability analyses are conducted using SHAP, leading to the development of a GWO-based energy optimization model with mass of scrap steel, hot metal temperature, and mass of lime as key regulatory factors. Optimization results indicate that the optimized energy consumption (in standard coal equivalent) is reduced by 2.63 kg/t. This model effectively reduces process energy consumption and provides decision support for energy-saving regulation in converter steelmaking, featuring both high accuracy and strong interpretability.