An accurate thermal error prediction model is crucial to the improvement of machining precision for numerical control machine. Aimed at the idea of modeling data generated during process operation usually presented with the characteristics of multiphase, multi-variables and three dimensional （3-D）, based on standard processing of the data, the partial least square method was employed to derive the predictive relationship between timeslice matrix and thermal error in high dimensional space, and the data space was reduced. K-means cluster algorithm was used to divide the models into different group under low dimensional properties space, next, the whole properties of manufacture process for a part were analyzed and knowledge discovered, therefore the thermal error predication model was established. The simulation experiment results show that as comparing with modeling method of BP thermal error, the proposed method has obviously improved predictive and generalization ability, which provides a novel way of idea for studying of thermal error prediction of NC machine, at the same time, it is an effective and practical solution.