UWB positioning system was used in an indoor three-line autonomous driving rail transit system, where high-precision positioning of vehicles was a key technology to improve operational reliability and scheduling efficiency. A three-step UWB location algorithm including non-line-of-sight (NLOS) discrimination, ranging error compensation and neural network location error compensation was proposed using a joint convolutional variational auto-encoder and predictor (VAE-CNN), based on the analysis of UWB location accuracy. Firstly, the ranging error and channel impulse response (CIR) data between the tag and the base stations were collected and used to train the VAE-CNN model. The non-line-of-sight ranging values were eliminated according to the confidence threshold of the original and reconstructed CIR. Secondly, the original ranging values were compensated by the prediction errors of the predictor. The coordinates and the direction cosine of the coordinates with respect to the coordinates of each base station were calculated, which were used to train the neural network to fit the relationship between the localization error and the direction cosine. The NLOS discrimination capability of the VAE-CNN model was validated on a publicly available UWB ranging and CIR dataset, which includes both line-of-sight (LOS) and NLOS measurements. The effectiveness of NLOS discrimination and ranging error compensation based on the VAE-CNN model on improving positioning accuracy was also evaluated. The effect of positioning error compensation neural network on improving positioning accuracy was evaluated based on the simulated vehicle trajectories under different ranging variances. An UWB localization system was built to verify the practical performance of the three-step UWB localization algorithm in dynamic localization. The results show that in dynamic localization, in full line-of-sight environment, the algorithm achieved an average localization error of 28. 68 mm, a root-mean-square localization error of 16. 67 mm, and a maximum localization error of 76. 68 mm. In the presence of non-line-of-sight environment, the average localization error is 38. 73 mm, the root mean square localization error is 20. 61 mm, and the maximum localization error is 116. 47 mm. It can be seen that the three-step UWB location algorithm offers high accuracy, low cost, and excellent stability, meeting the positioning requirements of indoor rail transit systems. Keywords:three-line indoor rail transit; UWB location; convolutional variati