Building prediction model to predict and compensate thermal error is a common method to solve the problem of thermal error of machine tools. In this method, the prediction accuracy and robustness of the model are easily affected by the environmental temperature, so a robust thermal error modeling algorithm based on partial least square method is proposed. Firstly, the correlation coefficient method is used to screen the temperature sensitive points, and the partial least squares regression prediction model of thermal error is established. Then, based on the multi batch thermal error experimental data under the annual ambient temperature, the optimal number of temperature sensitive points is analyzed. Finally, the partial least squares regression model of thermal error is established and compared with the ordinary multiple linear regression model. The results show that the average prediction accuracy of the proposed algorithm is 5. 7 μm, and the robustness of the model is 0. 56 μm. Compared with the ordinary multiple linear regression algorithm, the prediction accuracy and robustness are improved by 13. 8% and 49. 5% respectively. It shows that the thermal error robust modeling algorithm proposed in this paper can maintain high prediction accuracy and high robustness when the ambient temperature changes greatly.