<p>Accurate prediction of thermal error in high-speed electric spindle is crucial for maintaining machining precision and enhancing the reliability of the spindle unit. However, existing thermal error prediction methods ignore the multi-physical state information during spindle operation and complex spatial–temporal interactions, which limits the prediction performance of the model. To address these deficiencies, this work proposes a Spatial–Temporal Dynamic Adaptive Graph Convolutional Network (STDAGCN) based on multi-source information fusion for predicting thermal error. First, the spatial–temporal memory characteristics of thermal error were characterized through thermal error mechanism analysis. Next, the layout of temperature sensors was determined through thermal characteristics simulation analysis, and experimental data were collected under various operating conditions, including fine-grained temperature, electrical control, time, and cooling information. Subsequently, the architecture of the proposed STDAGCN model is detailed, in which complementary static, dynamic, and adaptive graph structures are used to dynamically extract spatial–temporal features at each moment. Finally, through model comparison evaluations, it was found that multi-source information input can significantly enhance model prediction performance, yielding an overall improvement in prediction precision of approximately 30%, with cooling information providing the most substantial contribution. The prediction accuracy of the STDAGCN model proposed in this work can reach up to 93.28%, showing excellent prediction performance. This model is integrated into the digital twin system, providing technological support for achieving higher machining precision and efficiency.</p>