Original Article

Prediction of manganese element recovery rate in the ladle section of a converter based on a hybrid integrated deep neural network

¹ College of Automation and Electrical Engineering, Inner Mongolia University of Science and Technology, Key Laboratory of Integrated Automation for Process Industry in Higher Education Institutions of Inner Mongolia Autonomous Region, Baotou 014010, Nei Mongol, PR China
² Infrastructure Department of Inner Mongolia University of Science and Technology, Baotou 014010, Nei Mongol, PR China

^* e-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.

Received: 22 December 2025
Accepted: 10 March 2026

Abstract

Post-converter ladle alloying represents a critical stage in the steelmaking process. Obtaining precise post-converter recovery rates for alloying elements enables more effective guidance for alloy charge additions, thereby enhancing product quality while reducing costs. Using actual production data from a steel mill’s converter smelting of Q355B steel grade as the research subject, Spearman correlation analysis and metallurgical mechanism analysis ultimately identified seven process parameters as input variables. By integrating the local feature extraction capability of convolutional neural networks (CNN) with the deep multidimensional global perception capability of bidirectional long short-term memory networks (BiLSTM), the model fully captures global patterns and local dependencies within complex data. A multi-strategy improved Iterative Whale Migration Algorithm (IWMA) was employed for global optimization of challenging hyperparameters. Furthermore, the AdaBoost algorithm from ensemble learning was introduced to weight-train multiple IWMA-CNN-BiLSTM models, enhancing generalization capability. This culminated in the proposal of a novel IWMA-CNN-BiLSTM-AdaBoost manganese element recovery rate prediction model. The effectiveness of each model component in enhancing overall performance was validated through ablation studies. Comparative experiments demonstrated the superiority of the proposed model over three baseline deep learning models, with its mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) all outperforming the alternatives. Specifically, the proposed model achieved a maximum improvement of 55.52% in MAE, 53.96% in RMSE, and 0.5511 in R². Within the 3% and 5% error margins, the hit rates reached 82.92% and 96.25%, respectively, indicating the models’ robust predictive capability.