Blast furnace thermal state prediction method based on mechanism-feature enhancement and LSTM-TCN

To address the problems of lagging thermal state assessment, delayed laboratory assay feedback, and insufficient feature representation under complex operating conditions in blast furnace thermal regulation, a furnace thermal state prediction method based on mechanism-enhanced features and a hybrid LSTM-TCN model is proposed. Using nearly one year of production data from No.6 blast furnace at TISCO, abnormal samples were removed through a combined Z-score and boxplot approach, followed by multi-source heterogeneous data preprocessing, time alignment, and sample construction. A burden mechanism model was established to derive the slag-iron heat index, thermal loss in the high-temperature zone, and multi-scale temporal features. Meanwhile, the maximum relevance minimum redundancy method was introduced to optimize high-dimensional features through dimensionality reduction and to identify key variables strongly correlated with hot metal temperature and silicon content, thereby enhancing the model’s ability to characterize the nonlinear, time-varying, and multivariable coupling relationships in the blast furnace process. A hybrid LSTM-TCN network architecture was then constructed to achieve collaborative modeling of temporal dependencies and local sequential features. Experimental results show that, for hot metal temperature prediction, the proposed model achieved an E_MA of 4.74 ℃, an E_RMS of 6.01 ℃, and a prediction accuracy of 92.5%; for silicon content prediction, the E_MA and E_RMS were 0.016 7% and 0.022 8%, respectively, with an accuracy of 93.5%. Compared with conventional single models and comparative models without mechanism-based features, the proposed method exhibited superior prediction accuracy, stability, and generalization capability. Industrial application results further demonstrate that the model can effectively support early identification of blast furnace thermal states and operational regulation, and thus has significant engineering value for improving the stability of hot metal quality, assisting operators in precise intervention, and promoting stable blast furnace operation.