Data-driven methods are commonly used for thermal error modeling, but the open-loop serial structure without mechanistic support makes it challenging to ensure model robustness under new operating conditions, leading to unreliable prediction performance. This paper introduces a semi-closed-loop spindle thermal error modeling approach based on dataset reconstruction. The original modeling batches are sorted and screened according to the ambient temperature of both the prediction and modeling batches. The modeling dataset is then reconstructed, and a semi-closed-loop thermal error model is developed. This method was applied to predict thermal errors in a lathe spindle, achieving root mean square errors of 1. 7 μm, 1. 7 μm, and 0. 9 μm in three test sets with the reconstructed models. Compared to conventional models, accuracy improved by 29. 2% , 39. 3% , and 64. 0% , respectively. This approach introduces a feedback loop to the existing serial open-loop thermal error modeling, offering significant potential for enhancing the performance of thermal error models.