In the wind tunnel, the Mach number in the test section is an important parameter that cannot be calculated accurately. Several mechanism models have been used to estimate it based on the aerodynamics laws in the past decades. However, the accuracy of estimation cannot satisfy measurements. An alternative approach is to design data-driven models. Unfortunately, the high-dimensional input features and the large-scale data sampled from measurements make it difficult to predict the Mach number. To solve the two issues, based on the multivariate fuzzy Taylor theorem, the feature subsets ensemble (FSE) method is proposed in this paper. An FSE is developed on the basis of the set of direct, exhaustive, independent subdivisions of the feature space. Learning on substantially lower dimension feature subsets, the FSE is characterized by low complexity. The FSE models are examined by data of measurements from the wind tunnel of China Aerodynamics Research and Development Center. Experiments show that the FSE speeds up the testing time that would otherwise be infeasible for the individual BP, Bagging and Random Forest. The FSE models meet the requirements of forecasting speed, accuracy and generalization of the Mach number prediction.