Since the selection means of structures of back propagation (BP) neural network having some drawbacks such as, heavily based on human experience, low convergence rate, easy to fall into local optimization and slow speed of covergency, which caused low thermal error predictive accuracy to exist in numerical control (NC) machine tool based on the BP, therefor, a new predictive error modelling method based on improved particle swarm optimized BP neural network was proposed. By a updating strategy of particle position and speed based on the improved standard particle swarm alogrithm, the optimized threshold and weight of the BP neural network were found. On basis of it, the thermal error predictive model for NC machine tool was built. The simulation experimental results show that the proposed thermal error predictive method has a higher predictive accuracy, better generalization ability as compared with standard algorithms of BP neural network and support vector machine.