As the scale of urban rail transit systems in many of our cities gradually expands, the accuracy of short-term passenger flow forecasting, which serves as a basis for train operation planning, has become increasingly important with the surge in passenger traffic. To improve the short-term passenger flow forecasting accuracy for subway stations, an optimization model, ConvLSTM is proposed. The model clusters similar stations and dates by K-Means clustering algorithm, and based on the clustering results, feature extraction and prediction are performed using one-dimensional convolutional neural network and Long Short-Term Memory (LSTM).Taking the AFC swipe data of Wuhan for three weeks in 2022 as an example, the 15min granularity inbound passenger flow is predicted using the ConvLSTM model, and the ARIMA, SVM, LSTM, and ConvLSTM models are evaluated by two evaluation metrics (MAPE, RMSE). The results show that the stations of the whole network can be categorized into: Hybrid station with a commercial bias, Residential Station, Hybrid station, Hub Station, Hybrid station with a residential bias; and the days can be categorized into: weekdays, Saturdays and Sundays. The ConvLSTM model has the highest prediction accuracy on hybrid stations, with Mean Absolute Percentage Error of 6.886%, 10.546%, and 11.793% on weekdays, Saturdays, and Sundays, respectively, as well as outperforming the control model on a sample of six types of stations. It shows that ConvLSTM has higher prediction accuracy in short-term prediction than single model.