This paper develops a novel data-driven modeling strategy and predictive controller for boiler-turbine unit using subspace identification and multimodel method. To deal with the nonlinear behavior of boiler-turbine unit, the system is divided into a number of local regions following the analysis of the nonlinearity distribution along the operation range, and then the corresponding measurement data are organized to identify the local models through the subspace method. By transforming local models into the same basis, the resulting multimodel system (MMS) is shown to represent the boiler-turbine unit very closely, and thus, used in designing a multimodel-based model predictive control (MMPC). As an alternative approach, a data-driven direct predictive controller (DDPC) is developed by utilizing the intermediate subspace matrices as local predictors. Online update of the predictor is also implemented on the multimodel structure to make the controller responsive to plant behavior variations. Simulation results demonstrate the feasibility and effectiveness of the proposed approach.