Temperature-sensitive points significantly affect the performance of the thermal error model of machine tools. To solve the problem that many measurement points need to be arranged and key temperature difference information may be lost in the existing temperature-sensitive point selection methods, this article proposes a temperature-sensitive point selection method based on active construction of temperature difference variables. The temperature difference variables are obtained from constructing a limited number of the original temperature measurement points, and added as an extension of the original temperature variables. The temperature-sensitive points are selected, which are based on fuzzy clustering and correlation coefficient analysis, and used for the thermal error modeling. This method can make up for the lack of potential key temperature information in existing methods, which has higher accuracy and stability. Compared with the traditional temperature-sensitive point selection methods, experimental results show that the proposed method can reduce the average root mean square error from 11. 1、 10. 3 μm to 3. 6 μm, which shows the remarkable effect.