Abstract: Objectives: Precise orbit prediction of low earth orbit (LEO) satellites is the foundation of LEO augmentation navigation. Due to the impact of non-conservative perturbation forces, conventional dynamic extrapolation method relies on high-accuracy dynamic models and suffer from rapid error accumulation. Methods: We combine machine learning algorithms and dynamic integration method to improve the accuracy of LEO orbit prediction. Three machine learning models are analyzed, including support vector machine (SVM), back propagation (BP) and long short term memory (LSTM) neural networks. To compare the performance of different machine learning algorithms in the high-accuracy orbit prediction, we process the orbit error time series of Sentinel-3A from January 1, 2019 to July 14, with the dynamic features selected by extreme gradient boosting (XGBoost). The optimal input configuration parameters for each machine learning models are further investigated to exploit the potential of the models in the orbit prediction. Results: The result shows that the performance of three machine learning models in the prediction varies from prediction time. When the prediction time is limited to 20 min, LSTM model presents the best performance with an accuracy improvement of 38.1%, which is evidently superior to SVM (-0.1%) and BP (-1.2%), and the corresponding orbit accuracy is 1.07 cm (3D). For the prediction exceeding 40 min, SVM and BP models outperform LSTM. When the prediction time reaches 180 min, the predicted orbit can still achieve the accuracy around 6-7 cm with the employment of SVM/BP models. Conclusion: Three models can effectively improve the orbit prediction accuracy of LEO satellites. LSTM model can achieve the best performance in ultra-short prediction (20 min), while SVM and BP models are more suited to the prediction exceeding 40 min. Compared with the traditional orbit prediction method, the improvement of machine learning models is about 40-50% overall.