The optimization selection of thermal sensitive points is a key issue in the thermal error modeling process. The accuracy and robustness of thermal error model is directly affected by the selected thermal sensitive points. A thermal sensitive point optimization method for machine tool spindle system combining gray relational analysis (GRA) and principal component analysis (PCA) was proposed. The GRA was used to screen the temperature measurement points that had great influence on thermal error, and the temperature values of several measurement points at different positions of the machine tool spindle and the thermal drift of the spindle at corresponding temperatures were taken as the analysis data.Through calculating the gray relational degree between the temperature variable and the thermal drift, the gray comprehensive relational matrix was obtained, and the temperature variable was selected after determining the correlation between the two variables. According to PCA, the highly correlated temperature data were simplified to fewer independent principal components. They were taken as the inputs of the subsequent thermal error model to realize the thermal sensitive points optimization. The method was applied to the CNC machine tool spindle system, and the optimized thermal sensitive point data were used as the input variables of the spindle thermal error model. The results show that taking the optimized thermal sensitive points as the inputs of BP thermal error model,the mean residual of the predicted thermal error and the actual thermal error is lower than that obtained by using GRA and PCA. The mean residuals are 0.83,5.18 and 4.57μm respectively.z-axis thermal deformation prediction accuracy〖JP〗 of the machine tool is significantly improved, which is beneficial to improve the machining accuracy.