The nonlinearity and complexity of a boiler-turbine unit make its precise modeling a challenging task. In this paper, a hybrid of the two-phase feature selection and deep belief network (DBN) approach is proposed to predict the key variables of the boiler-turbine unit (e.g., steam pressure, steam flow, and active power). Actual operational data were collected from a local thermal power plant and were standardized to eliminate the effect of units with different variables. Moreover, data processing for normalization was carried out by Box-Cox transformation. Subsequently, a two-phase feature selection strategy was implemented. Using this strategy, variable selection based on Pearson correlation analysis was first carried out. Then, the PCA reconstructed the new features of the prediction model. Finally, a multilayer DBN with a back-propagation-based fine-tuning algorithm was developed to model the nonlinear relationship between the reconstructed input variables and key output variables. To validate the effectiveness of the proposed approach, numerical experiments were conducted based on practical data. The experimental results confirmed the outstanding performance of the proposed modeling approach in comparison to other classical approaches.