The ironmaking industry is nowadays facing the great challenge of process optimization and transformation. Since the blast furnace is still the main facility for metallic iron production, a stable blast furnace operation aiming at lowering reducing agent consumption remains the main target in the daily business. Thus, a well-controlled blast furnace thermal state should base on a reliable thermal state prediction. This work aims to improve the performance of existing thermal state prediction models by increasing the frequency of hot metal temperature measurement. A multi-wavelength pyrometer was installed at each taphole of one blast furnace to continuously measure the hot metal temperature at the taphole. Three contrast tests were carried out using continuous immersion thermocouple lances in the skimmer to validate the pyrometer measurements. The results of the final contrast test showed an excellent correlation between the trustable pyrometer measurements and the continuous thermocouple measurements. The trustable measurements were applied to the existing hot metal temperature prediction model based on mass and energy balance calculations. An offline calculation with the dynamic gradient showed an improvement in prediction accuracy. To enhance the accuracy of the hot metal temperature prediction, two GRU-based (Gated Recurrent Unit) deep learning neural networks were established and trained on the same input dataset, with the exception of the hot metal temperatures. Both models demonstrated similar MAEs (Mean Absolute Error), although their performances differed slightly. The potential for further improvement of hot metal temperature prediction models is discussed later in this work.