With the intelligent transformation of process manufacturing, accurate and comprehensive perception information is fundamental for application of artificial intelligence methods. In zinc smelting, the fluidized bed roaster is a key piece of large-scale equipment and plays a critical role in the manufacturing industry; its internal temperature field directly determines the quality of zinc calcine and other related products. However, due to its vast spatial dimensions, the limited observation methods, and the complex multiphase, multifield coupled reaction atmosphere inside it, accurately and timely perceiving its temperature field remains a significant challenge. To address these challenges, a spatial-temporal reduced-order model (STROM) is proposed, which can realize fast and accurate temperature field perception based on sparse observation data. Specifically, to address the difficulty in matching the initial physical field with the sparse observation data, an initial field construction based on data assimilation (IFC-DA) method is proposed to ensure that the initial conditions of the model can be matched with the actual operation state, which provides a basis for constructing a high-precision computational fluid dynamics (CFD) model. Then, to address the high simulation cost of high-precision CFD models under full working conditions, a high uniformity (HU)-orthogonal test design (OTD) method with the centered L₂ deviation is innovatively proposed to ensure high information coverage of the temperature field dataset under typical working conditions in terms of multiple factors and levels of the component, feed, and blast parameters. Finally, to address the difficulty in real-time and accurate temperature field prediction, considering the spatial correlation between the observed temperature and the temperature field, as well as the dynamic correlation of the observed temperature in the time dimension, a spatial-temporal predictive model (STPM) is established, which realizes rapid prediction of the temperature field through sparse observation data. To verify the accuracy and validity of the proposed method, CFD model validation and reduced-order model prediction experiments are designed, and the results show that the proposed method can realize high-precision and fast prediction of the roaster temperature field under different working conditions through sparse observation data. Compared with the CFD model, the prediction root-mean-square error (RMSE) of STROM is less than 0.038, and the computational efficiency is improved by 3.4184  × 10⁴ times. In particular, STROM also has a good prediction ability for unmodeled conditions, with a prediction RMSE of less than 0.1089.