© 2020, The Brazilian Society of Mechanical Sciences and Engineering. The aim of this study is to reduce the influence of machine tools spindle thermal error on machining accuracy in load machining state. The temperature and thermal error measuring system and experiments of the 3-axis vertical milling machine tool spindle in idling and load machining state were established and carried out. The differences in temperature and thermal error between the idling and load machining states were analyzed. Upon combining the fuzzy clustering and gray correlation algorithm, the temperature-sensitive points were optimized. The thermal error prediction models of machine tool spindle system in load machining state with the optimal specific cutting energy were established based on the adaptive chaotic particle swarm optimization algorithm, and the model prediction effects were evaluated. The results showed that the spindle system temperature and thermal error in load machining were higher than the idle state. Two temperature-sensitive points were selected that not only reduced the redundancy of temperature measuring points but also ensured the model prediction accuracy. The thermal error models prediction accuracy was above 90%, and the root mean square error and residual error were better than PSO and regression. The experimental results showed that the thermal error prediction models have a high prediction accuracy and engineering application value.