Abstract: During the operation of a spindle unit, heat generated by the motor and bearings induces structural deformation, which consequently affects machining accuracy. To address this issue, the spindle unit of a precision horizontal machining center is investigated and research on its thermal characteristics and active temperature control is conducted. Firstly, heat generation and dissipation models of the spindle unit are established, and finite element simulations are performed to reveal the temporal evolution of the temperature and thermal deformation fields. Secondly, a multi-source time-series dataset is constructed based on both simulation and experimental data. A novel deep learning model, namely BiLSTM-DCIB-KCBAM-GRU(BDKG), is proposed, which integrates a dilated convolution interaction block (DCIB) and a kan-augmented convolutional block attention module (KCBAM), enabling high-accuracy mapping from temperature measurements to thermal deformation. Ablation studies demonstrate that DCIB significantly improves deformation prediction accuracy. Finally, based on the predicted thermal deformation trends, a progressive optimization method for coolant inlet temperature is developed by combining a hybrid Gaussian dung beetle optimizer (HGDBO) with extreme gradient boosting (XGBoost), forming an active temperature control strategy for suppressing thermal deformation. Experimental results show that, compared with conventional constant-temperature cooling, the proposed method reduces the maximum axial thermal deformation of the spindle unit from approximately 18 μm to 2.5 μm, markedly enhancing its thermal stability.