It is difficult to establish the accurate mapping relationship between thermal error and temperature of machine tools. Aimed at the problems of adaptability and robustness of the thermal error model based on the traditional shallow network, a method of thermal error prediction and compensation based on parallel deep learning network was proposed. A deep learning prediction model based on three sub depth belief networks in parallel was established. Each sub depth belief network had the same network structure and different weight parameters. And the restricted Boltzmann machine of the input layer was shared to each sub depth belief network. A construction method of the parallel depth network structure based on prediction error was designed to determine the number of neurons in each RBM hidden layer. A parallel depth network training method based on initial weight sharing was proposed. One of the depth belief networks of the model was pretrained based on the unsupervised learning method with logarithmic divergence. Other depth belief networks shared the initial weight. And the backpropagation algorithm was used to further adjust the optimal weights of each sub depth belief network. The experimental results showed that the root mean square error of thermal error model based on parallel deep learning network was 2.2μm. This method improved the adaptability and robustness of thermal error compensation greatly while improving the accuracy of prediction.