<p>Pre-trained neural network models for machine tool spindle thermal displacement prediction have been widely applied to improve machining accuracy. However, when adapting to new spindle operating conditions, these models are prone to catastrophic forgetting during knowledge updates. To address this issue, this study proposes a replay-based continual learning framework for spindle thermal displacement modeling. A generator network is employed to produce virtual samples representing previously learned tasks, which are combined with data from new tasks to form an updated training dataset, enabling continuous optimization while retaining prior knowledge. To enhance virtual data quality and distribution fidelity, a Thermal Distribution-Regularized Generative Adversarial Network (TDR-GAN) is developed. The method integrates an authenticity discriminator to ensure sample realism and a distribution regularization mechanism that explicitly constrains the statistical structure of generated data. By directly penalizing distribution discrepancies between real and generated samples, the proposed method promotes global diversity and structural consistency during replay. Experimental results indicate that the proposed approach improves the quality of generated virtual data compared with conventional GAN-based methods. In continual learning scenarios involving spindle thermal expansion tasks, the prediction model achieves an average RMSE of 1.38&#xa0;μm and demonstrates favorable and stable performance compared with representative continual learning strategies. These results suggest that the proposed framework effectively balances knowledge retention and adaptation, providing a practical solution for intelligent spindle thermal displacement prediction.</p>