In order to realize stable production in the steel industry, it is important to control molten steel temperature in a continuous casting process. The present work aims to develop a gray-box model that predicts the molten steel temperature in the tundish (TD temp). In the proposed approach, two parameters in the first-principle model, i.e., overall heat transfer coefficients of ladle and tundish, are optimized for each past batch separately, then the relationship between the two parameters and measured process variables is modeled through random forests (RF). In this inner gray-box model, the statistical models update the physical parameters according to the operating condition. To enhance the accuracy of TD temp estimation, another RF model is developed which compensates errors of the inner gray-box. The proposed approach was validated through its application to real operation data at a steel work.