1 Material Technology Research Institute, Hesteel Group, Shijiazhuang 050023, Hebei, China 2 Tangsteel Company, Hesteel Group, Tangshan 063611, Hebei, China 3 School of Artificial Intelligence, Beijing Technology and Business University, Beijing 100048, China 4 School of Metallurgical and Ecological Engineering, University of Science and Technology Beijing, Beijing 100083, China
TSC prediction and dynamic control of BOF steelmaking with state-of-the-art machine learning and deep learning methods
1 Material Technology Research Institute, Hesteel Group, Shijiazhuang 050023, Hebei, China 2 Tangsteel Company, Hesteel Group, Tangshan 063611, Hebei, China 3 School of Artificial Intelligence, Beijing Technology and Business University, Beijing 100048, China 4 School of Metallurgical and Ecological Engineering, University of Science and Technology Beijing, Beijing 100083, China
摘要 Mathematical (data-driven) models based on state-of-the-art (SOTA) machine learning and deep learning models and data collected from 12,786 heats were established to predict the values of temperature, sample, and carbon (TSC) test, including temperature of molten steel (TSC-Temp), carbon content (TSC-C) and phosphorus content (TSC-P), which made preparation for eliminating the TSC test. To maximize the prediction accuracy of the proposed approach, various models with different inputs were implemented and compared, and the best models were applied to the production process of a Hesteel Group steelmaking plant in China in the field. The number of tabular features (hot metal information, scrap, additives, blowing practices, and preset values) was expanded, and time series (off-gas profiles and blowing practice curves) that could reflect the entire steelmaking process were introduced as inputs. First, the latest machine learning models (LightGBM, CatBoost, TabNet, and NODE) were used to make predictions with tabular features, and the best coefficient of determination R2 values obtained for TSC-P, TSC-C and TSC-Temp predictions were 0.435 (LightGBM), 0.857 (Cat-Boost) and 0.678 (LightGBM), respectively, which were higher than those of classic models (backpropagation and support vector machine). Then, making predictions was performed by using SOTA time series regression models (SCINet, DLinear, Informer, and MLSTM-FCN) with original time series, SOTA image regression models (NesT, CaiT, ResNeXt, and GoogLeNet) with resized time series, and the proposed Concatenate-Model and Parallel-Model with both tabular features and time series. Through optimization and comparisons, it was finally determined that the Concatenate-Model with MLSTM-FCN, SCINet and Informer as feature extractors performed the best, and its R2 values for predicting TSC-P, TSCC and TSC-Temp reached 0.470, 0.858 and 0.710, respectively. Its field test accuracies for TSC-P, TSC-C and TSC-Temp were 0.459, 0.850 and 0.685, respectively. A related importance analysis was carried out, and dynamic control methods based on prediction values were proposed.
Abstract：Mathematical (data-driven) models based on state-of-the-art (SOTA) machine learning and deep learning models and data collected from 12,786 heats were established to predict the values of temperature, sample, and carbon (TSC) test, including temperature of molten steel (TSC-Temp), carbon content (TSC-C) and phosphorus content (TSC-P), which made preparation for eliminating the TSC test. To maximize the prediction accuracy of the proposed approach, various models with different inputs were implemented and compared, and the best models were applied to the production process of a Hesteel Group steelmaking plant in China in the field. The number of tabular features (hot metal information, scrap, additives, blowing practices, and preset values) was expanded, and time series (off-gas profiles and blowing practice curves) that could reflect the entire steelmaking process were introduced as inputs. First, the latest machine learning models (LightGBM, CatBoost, TabNet, and NODE) were used to make predictions with tabular features, and the best coefficient of determination R2 values obtained for TSC-P, TSC-C and TSC-Temp predictions were 0.435 (LightGBM), 0.857 (Cat-Boost) and 0.678 (LightGBM), respectively, which were higher than those of classic models (backpropagation and support vector machine). Then, making predictions was performed by using SOTA time series regression models (SCINet, DLinear, Informer, and MLSTM-FCN) with original time series, SOTA image regression models (NesT, CaiT, ResNeXt, and GoogLeNet) with resized time series, and the proposed Concatenate-Model and Parallel-Model with both tabular features and time series. Through optimization and comparisons, it was finally determined that the Concatenate-Model with MLSTM-FCN, SCINet and Informer as feature extractors performed the best, and its R2 values for predicting TSC-P, TSCC and TSC-Temp reached 0.470, 0.858 and 0.710, respectively. Its field test accuracies for TSC-P, TSC-C and TSC-Temp were 0.459, 0.850 and 0.685, respectively. A related importance analysis was carried out, and dynamic control methods based on prediction values were proposed.
Tian-yi Xie,Cai-dong Zhang,Quan-lin Zhou, et al. TSC prediction and dynamic control of BOF steelmaking with state-of-the-art machine learning and deep learning methods[J]. Journal of Iron and Steel Research International, 2024, 31(1): 174-194.
2024, Vol. 31 
