Accurate measurement of molten steel temperature is critical for ensuring product quality and improving process efficiency in steelmaking. However, conventional infrared thermometry is highly susceptible to disturbances such as smoke, emissivity fluctuations, and complex operating conditions, which often lead to significant measurement errors. To address this issue, this study proposes a Transformer-based infrared temperature prediction method. The proposed approach integrates multi-source features, including raw infrared temperature, environmental parameters, equipment conditions, and process variables, and incorporates a comprehensive preprocessing pipeline covering outlier filtering, missing value imputation, feature engineering, and differentiated feature scaling to improve data quality and robustness. The model employs a multi-head self-attention mechanism to capture both short-term fluctuations and long-range temporal dependencies, while integrating interaction features and physics-informed constraints to enhance interpretability. Experimental results demonstrate that the model maintains stable prediction accuracy under smoke interference and rapid temperature fluctuations, with errors consistently confined within an acceptable range, thereby meeting the precision requirements of industrial steelmaking. These findings confirm the effectiveness and practical value of the proposed Transformer-based approach for robust infrared molten steel temperature prediction in real-world industrial applications.