Turbine blades are key hot components of a gas turbine, which operates in high temperature and harsh environment. Accurate and efficient prediction of turbine blade temperature field is of great significance for gas turbine service safety and stability. In this study, a turbine blade fluid-thermal-solid coupling model was established for different start-up conditions of gas turbines, and numerical simulations were carried out to form a dataset of turbine blade temperature field. An efficient prediction method of turbine blade temperature field is proposed by combining neural network and model order reduction technique. The temperature field of the turbine blade can be fast predicted through on-site sensor data. The results show that compared with the fluid-thermal-solid high-fidelity numerical simulation, the proposed method is able to efficiently predict the temperature field of the turbine blade in milliseconds, and the relative error is less than 12%. This work provides a feasible approach for real-time monitoring of three-dimensional temperature field of turbine blades.