To reduce the impact of thermal errors on CNC machine tools, improve the machining accuracy of workpieces, and solve the problem of poor thermal error prediction accuracy under different working conditions. The thermal error measurement experiment of the CNC machine tool feed system is conducted under working conditions of a feed speed of 10 m/ min and an ambient temperature of 20°. The Pelican optimization algorithm is used to optimize the neural network, determine the optimal weight and threshold of the BP neural network, and the thermal error of the feed system prediction model of POA-BP is established. The experiment is compared and analyzed with traditional BP neural network, GA-BP neural network and the SCN random configuration network. The results show that the average relative error of traditional BP neural network prediction is 12. 23%, the average relative error of GA-BP neural network is 11. 5%, the average relative error of SCN prediction model is 12. 71%, and the average relative error of POA-BP prediction model is 9. 93%, which improves the accuracy. Conclusion: The neural network improved by the proposed Pelican optimization algorithm has strong effectiveness and accuracy in thermal error prediction, which can improve the accuracy of feed motion and provide theoretical guidance for the realization of thermal error compensation.