<p>The end-point temperature prediction model of Ladle Furnace (LF) is crucial for LF temperature control and stable continuous casting operations. Currently, data-driven prediction models have been widely adopted due to their high accuracy. However, because of the limitations of their modeling mechanisms, methods to improve accuracy can easily lead to overfitting and lower stability of the models. Additionally, the generalization ability of data models in response to process changes is also being challenged. This paper uses a combination of k-means clustering analysis and improved dropout regularization to optimize conventional neural network models, ensuring both the accuracy and stability of the prediction models. A mechanistic model was then established and an error compensation method was used to integrate this mechanistic model with the optimized data model, enhancing its generalization ability. The results show that the hybrid model based on optimized data-driven model and mechanistic model achieved hit rates of 81.20%, 87.31%, and 91.35% within error ranges of ± 5&#xa0;°C, ± 7&#xa0;°C, and ± 10&#xa0;°C, respectively. The hybrid model exhibits high computational accuracy close to that of pure data models. Through validation with actual production data, it demonstrates better performance in stability and generalization compared to pure data models, which is beneficial for achieving “narrow window control” of molten steel temperatures. At the same time, calculations show that for every 1&#xa0;°C reduction in temperature fluctuation, electricity consumption can be reduced by 0.4335 kW·h per ton of steel, or CO₂ emissions by 0.3144 kg per ton of steel. This research contributes to the greening and decarbonization of the steelmaking process.</p> Graphical Abstract <p></p>