Multi‐Step Temperature Prediction for a <scp>TGAL</scp> Regenerative Aluminum Smelting Furnace

ABSTRACT This study addresses the industrial challenge that the temperature inside regenerative aluminum smelting furnaces cannot be directly or accurately measured. To overcome this issue, a TGAL hybrid model combining a Temporal Convolutional Network (TCN), Graph Convolutional Network (GCN), Multi‐Head Attention mechanism, and Long Short‐Term Memory (LSTM) network is proposed for multi‐step accurate prediction of furnace temperature. The method first applies wavelet denoising to suppress noise in the industrial data collected by the SCADA system and employs the Pearson correlation coefficient to select highly correlated features, thereby improving the quality of the input data. The proposed TGAL model exploits the synergy of TCN in capturing long‐term temporal dependencies, GCN in uncovering spatial correlations among variables, the attention mechanism in dynamically weighting features, and LSTM in temporal dynamic modeling. Validation on 44,640 one‐minute data samples from actual production shows that, compared with traditional models, the proposed model achieves maximum improvements of 7.44% in RMSE, 24.85% in MAE, and 25.27% in MAPE for 2‐step prediction, respectively. For 10‐step prediction, the improvement rates remain at least 4.23% in RMSE, 6.91% in MAE, and 6.31% in MAPE. Moreover, Diebold–Mariano statistical tests confirm that the TGAL model's predictive accuracy is significantly superior to that of the comparison models. Nevertheless, the model performance under extreme operating conditions remains limited by data noise and nonlinear dynamics, and the physical mechanisms of the smelting process have yet to be incorporated. To address these limitations, future work will focus on dynamic coupling modeling and the embedding of physical information to further enhance the model's generalization capability and physical consistency.