Aiming at the problem that the common thermal error models have low prediction accuracy when the ambient temperature changes greatly, a piecewise weighted modeling method of thermal error based on model library of ambient temperature was proposed. The UPM120 CNC milling machine was used as the experimental object. Modeling and prediction accuracy analysis of the piecewise weighted model with ambient temperature between 15℃ and 35℃ were accomplished by using seven batches of data in different quarters. The experimental results showed that when the ambient temperature was varied within 5℃, the prediction accuracy of multiple linear regression model was better than that of BP neural network model, distributed lag model, grey theory model and support vector machine model, so multiple linear regression model was used as the basic model of the piecewise weighted model library. When the ambient temperature had small change, the piecewise weighted model based on multiple linear regression had a prediction accuracy of 1.39μm, which was slightly lower than that of the multiple linear regression model, but higher than that of the other four common thermal error models. When the ambient temperature had great change, the prediction accuracy was 1.51μm, which was much higher than the accuracy of multiple linear regression model of single ambient temperature sample, multiple linear regression model of multienvironment temperature samples and support vector machine model with strong generalization ability. The piecewise weighted model had high prediction accuracy under both large or small changes in ambient temperature.