The disadvantages of the common current thermal error modeling methods for CNC machine tool feed drive systems were analyzed, such as the requirement of many temperature sensors to reach high accuracy and poor applicability of different moving states. A new robust modeling method based on the heat transfer theory is proposed, and the procedure of the thermal tests for a feed drive system is presented. Multiple regression method and robust modeling method based on the heat transfer theory were, respectively, used to establish a thermal error model, and a pointer automatic optimizer was used to optimize the parameters in the robust model. A compensation simulation was conducted under five different moving states using these two modeling methods, and the advantages of the robust modeling method were proved. Finally, the compensation effect of the robust modeling method was verified under a random moving state on a vertical machining center.