Abstract: Thermal error is one of the primary factors affecting the machining accuracy of CNC machine tools. Its modeling involves not only temporal dependencies but also spatial correlations among multiple sensors. Traditional approaches often fail to fully capture such spatio-temporal interactions and lack robustness when transferred across different machines. To address these limitations, this study proposes a novel thermal error modeling framework based on a Graph Convolutional Network (GCN) and Transformer architecture. In the proposed method, the GCN is employed to extract spatial dependencies among temperature sensors, while the Transformer captures temporal dynamics, enabling an efficient fusion of spatial and temporal features. A multi-domain feature extraction module is further designed to integrate these features for prediction loss computation as well as cross-machine domain adaptation. Given that thermal characteristics of machine tools may vary due to assembly deviations and performance degradation, a domain adaptation strategy based on multi-kernel maximum mean discrepancy (MK-MMD） is introduced to align the feature distributions between source and target domains, thereby enhancing cross-machine generalization capability. Experimental evaluations demonstrate that the proposed method consistently outperforms existing thermal error models under multi-condition and cross-machine scenarios. Moreover, when labeled samples on the target machine are limited, transfer learning significantly improves prediction accuracy and robustness. These findings highlight the effectiveness of the proposed GCN-Transformer framework and domain adaptation strategy, offering a promising solution to the challenges of thermal error modeling and compensation in complex operating environments.