Abstract: Numerical control machine tool in the process of processing will cause thermal deformation of spindle due to heat. In order to reduce the influence of thermal deformation on the precision of machining parts, a numerical control lathe was taken as the research object. A Gaussian process regression (PSO-GPR) thermal error modeling and prediction method based on particle swarm optimization was proposed. By collecting and measuring the experimental data under five different working conditions, the thermal error model of the axial axis of the main shaft was built, and compared with the unoptimized GPR modeling method. On this basis, the influence of training data enhancement on the generalization of the thermal error model was studied. The experimental results show that the maximum residual error of PSO-GPR model for predicting thermal deformation is 0.49 μm, and the root mean square error (RMSE) is 0.11 μm, which is better than the unoptimized GPR model. The PSO-GPR model with data enhancement reduces the maximum residual in thermal error prediction of verified data by 35% and 33.7% in operating conditions 4 and 5 respectively, indicating that the enhancement of training data can improve the generalization ability of thermal error model.