Abstract: In order to solve the problems of standard deviation and non-standard deviation in the process of automatic rail transit train parking, an effective solution is proposed based on the analysis of running log big data and the LSTM (long and short term memory) network algorithm. Firstly, the big data analysis is conducted for a large amount of historical information on train parking accuracy in the running log. In which the data is divided into stages with 1 d as a statistical cycle, then,the data is preprocessed and multi-type fitted. After comparison, the time series of the best fitting parameters is obtained. On this basis, a deep learning model is built by LSTM network algorithm to predict the distribution of automatic train parking accuracy. Finally, based on the historical data of train parking accuracy of a subway line in Chengdu, the LSTM prediction model is trained and verified. Results show that the prediction model can meet the statistical requirements for the similarity greater than 0.9. Thus, the effectiveness and accuracy of the model are verified.