Clustering-enhanced transformer for temperature prediction in regenerative aluminum melting furnaces

Aluminum, with its lightweight, high strength, corrosion resistance, and recyclability, is widely used in transportation, construction, energy, and electronics, making it a key foundational material for modern industry and sustainable development. Accurate temperature prediction of regenerative aluminum melting furnaces is of great importance for energy efficiency and stable operation in aluminum production. However, the nonlinear dynamics and strong multivariable coupling of the process pose significant challenges. This paper proposes a Transformer-based temperature prediction method enhanced with clustering embeddings. Time-series clustering is first applied to identify different operating conditions, and the resulting cluster embeddings are incorporated into the model to improve adaptability to multiple conditions. The attention mechanism is then leveraged to effectively capture long-term dependencies and interactions among variables. Experiments on real furnace temperature data demonstrate that the proposed method outperforms conventional Transformer and LSTM models in both prediction accuracy and robustness, with a significant reduction in mean squared error and a clear improvement in the coefficient of determination. These results verify the effectiveness of incorporating clustering information into deep learning frameworks and provide a feasible approach for modeling and optimizing complex industrial processes.