<p>Thermal errors in long-stroke ball screw feed drive systems (LSBSFDS) significantly affect the motion precision. The thermal error models often suffer from limited accuracy and adaptability due to inaccurate thermal boundary conditions. This paper enhances thermal error modeling for LSBSFDS by establishing a thermal network to simulate the heat transfer mechanism and developing a time-varying state-space representation of the thermal network. A novel method is proposed to identify thermal boundary conditions that emphasize predictive consistency and adherence to physical laws. This method improved model performance with a 1.23 µm reduction in the maximum deviation on the test dataset. By incorporating heat dissipation during the measurement process, the maximum deviation is reduced by an additional 4.16 µm on average, and 97.69% overall average prediction accuracy is obtained. Finally, a statistical method is introduced to assess uncertain disturbances and detect potential model overfitting. This work provides a robust framework for high-precision thermal error prediction.</p>