Real-time accurate prediction of short-term railway passenger demand can provide basis for real-time adjustment of passenger service structure. Railway passenger flow data has characteristics such as time-varying, nonlinear and stochastic volatility. Traditional forecasting models cannot predict the short-term passenger traffic volume accurately. This study proposed a hybrid deep learning model based on Wavelet Packet Analysis and Long Short-Term Memory (WPA-LSTM). Firstly, the original passenger volume time series is decomposed into several low-frequency and high-frequency sequences with different scales by wavelet packet. Then, the LSTM model training and prediction are carried out respectively for each sub-sequence. Finally, the prediction values of each sub-sequence are superimposed as the output of WPD-LSTM model. The model was validated by the daily passenger flow data of 367 days of one high-speed railway. The model was compared with the seasonal model and the empirical model. The experimental results show that WPD-LSTM model can effectively improve the accuracy of railway passenger traffic forecasting.