With the rapid development of intelligent transportation, accurate passenger flow forecasting in urban rail transit can effectively enhance resource allocation efficiency and reduce operational costs. However, the spatiotemporal dependency and uncertainty present in passenger flow time series make it challenging to predict short-term pas-senger flow accurately. To address this issue, this study proposes a deep learning-based intelligent forecasting model for short-term passenger flow at multiple urban rail transit stations, combining traditional time series fore-casting models with deep learning techniques to enhance the accuracy of short-term passenger flow predictions for different types of stations. First, the wavelet thresholding method is used to denoise the subway data, elimi-nating misleading information and enhancing data quality. Secondly, the K-Means clustering method is employed to categorize the subway stations based on passenger flow characteristics, in order to clarify the demand patterns of each station. Next, the processed short-term passenger flow samples are input into the three individual models: LSTM, GRU, and ARIMA for prediction. Weights are assigned based on the prediction error performance of each model on the validation set. Stacked ensemble learning is then used to train a meta-learner, obtaining the weight values for the different models and completing the model establishment. In the experiment, noise tolerance testing is conducted to validate the model's stability, using subway data from Beijing for model verification. The experi-mental results indicate that the intelligent forecasting model exhibits low prediction errors under various envi-ronmental conditions and demonstrates strong resistance to noise interference. Additionally, compared to indi-vidual models and other existing models, the accuracy of passenger flow predictions is improved by 4.819% and 8.714%, respectively. Therefore, the deep learning-based intelligent forecasting model for short-term passenger flow at multiple urban rail transit stations maintains high accuracy and robustness when facing diverse and com-plex data, effectively addressing the challenges