The modeling methods for thermal errors of a ball screw was researched to improve the positioning precision of a sem-closed loop ball screw driving system. The dynamic characteristics of heat sources, temperature fields of the ball screw were analyzed and how to establish the thermal error model based on Elman neural networks and operating conditions was proposed. Firstly, the internal heat sources and temperature field distribution characteristics of the ball screw were determined according to the structure studied. Then, the dynamic nonlinear functional relationship between the thermal deformation of ball screw and its internal heat sources was researched based on the temperature distribution function. Finally, on the basis of Elman neural networks, the thermal error prediction model of ball screw was established in consideration of the effect of operating condition on the thermal error. Experimental results indicate that the estimated residual error of thermal deformation by the proposed model varies from -3.1 μm to 2.4 μm when the ball screw is worked at a complex environment. It concludes that the prediction precision and robustness by the proposed model based on Elman neural network and operating conditions are better than those by the BP and Elman neural networks (just considering temperature data), and it shows a powerful engineering application prospect.