Abstract: The temperature of molten steel during Ladle Furnace treatment is a critical process parameter that directly influences both the quality of the refined steel and the operational stability of the subsequent continuous casting process. However, existing molten steel temperature prediction models often face limitations in adaptability and prediction accuracy due to an overreliance on metallurgical mechanisms or global data-driven algorithms in the Ladle Furnace refining process. To address this issue, this study proposes a modeling approach that integrates metallurgical mechanisms with production data. The theoretical energy increase of the system is first calculated according to the principle of energy conservation. Subsequently, the end-point temperature of the molten steel is determined based on this calculation. Then, the difference between the actual detection end-point molten steel temperature and the calculated temperature by the mechanism model is defined as the residual, which is used to reflect the unmodeled errors caused by mechanism simplification and operational fluctuations, and is used as the output item of the Just-In-Time Learning model. Finally, a weighted Euclidean distance metric is employed to identify the most similar historical samples. Utilizing these selected samples, a local regression model is then constructed, the residual prediction (from the data-driven component) is combined with the output of the mechanistic model to produce the final predicted end-point molten steel temperature of the fusion model. Compared to traditional mechanism-based and global data-driven models, the proposed fusion model demonstrates significantly improved prediction performance. Model validation on an independent dataset shows that the prediction accuracy within -5, 5 and -10, 10 reaches 87.2% and 98.8%, respectively, with R2 of 0.986.