<p>Thermal errors arising from machine tool thermal deformation significantly affect the accuracy and stability of CNC machining. Real-time thermal error compensation, implemented via high-precision data-driven models, is a widely accepted method for ensuring consistent machining quality. However, because the model’s prediction accuracy relies entirely on the availability of a sufficient number of labeled samples, the practical application of thermal error compensation technology is limited by challenges including difficulties in acquiring modeling data and high modeling costs. To address thermal error modeling of CNC machine tools under small sample conditions and to enable multi-model expansion of the thermal error model, this study proposes a thermal error modeling method based on incremental transfer learning. Firstly, using support vector regression (SVR), an initial thermal error model for the source domain machine tool is established based on small sample error data; subsequently, an incremental learning approach combined with a sample screening mechanism founded on Karush–Kuhn–Tucker (KKT) conditions is employed to optimize the model’s boundaries, thereby enhancing accuracy and generalization. Secondly, a thermal error transfer model for various types of machine tools is constructed by integrating KKT screening with correlation alignment-based domain adaptation methods (CORAL). Finally, the proposed model (K-CA SVR) is compared with the traditional SVR model and a convolutional neural network (CNN) model, while real-time error compensation experiments are conducted on various types of machine tools. The results indicate that the actual thermal error compensation accuracy for the source domain machine tool exceeds 90%, whereas the compensation accuracy for the target domain machine tool exceeds 80%. These findings confirm the effectiveness of the proposed model.</p>