Abstract: Thermal error is one of the primary factors affecting the machining accuracy of precision CNC machine tools, characterized by nonlinearity, time-varying behavior, and strong dependence on operating conditions. To address the problems of overfitting and insufficient feature utilization under small-sample conditions, this study takes the μ2000-400H CNC machine tool as the research object and proposes a lightweight Thermal Information-Enhanced Temporal Attention Network (TIETA), which is deployed within an edge digital twin framework for real-time thermal error compensation. The model employs channel reweighting to highlight critical temperature features, integrates gated recurrent unit (GRU) to capture temporal dependencies, and incorporates an attention mechanism to strengthen the extraction of key temporal patterns. Experimental results across multiple operating conditions demonstrate that TIETA outperforms traditional methods and complex deep-learning models in terms of accuracy and robustness. In the edge closed-loop application, the model enables real-time correction of spindle-end displacement, significantly reducing residual errors. These findings confirm that the proposed method offers high modeling accuracy, strong generalization, and practical applicability for thermal error modeling and compensation in machine tools.