<p>Aluminum smelting relies on precise furnace temperature measurement to ensure product quality and yield, yet accurate real-time measurement faces technical barriers: thermocouples are susceptible to damage in high-temperature environments, resulting in increased measurement costs, while protective sheaths of thermocouples introduce thermal inertia, causing delayed temperature response in molten metal monitoring. Furnace temperature prediction models are usually used to achieve accurate prediction of furnace temperature to solve these issues. However, traditional methods often struggle to manage the dynamic spatial-temporal relationships among variables, leading to suboptimal predictive outcomes. Considering the aforementioned challenges, a proposal of a furnace temperature prediction model is made. The proposed model utilizes a graph convolutional network that combines dynamic graph fusion structures to capture evolving spatial correlations, supplemented by a sliding fully connected LSTM architecture with gating mechanisms, aiming to extract temporal dependencies through time series data analysis. Experimental results obtained from real-world aluminum smelting furnace data demonstrate that the proposed model achieves superior prediction accuracy compared to LSTM and Transformer baselines. Ablation experiments further validate the critical contributions of each architectural component. These results validate its ability to manage significant numerical variations and dynamic spatial-temporal relationships, establishing it as a robust solution for real-time temperature prediction in aluminum smelting furnaces.</p>