China's urban rail transit is developing rapidly. Short-term passenger flow prediction is of great significance for operational safety, network optimization, and then smart city building. While the urban rail passenger flow is cyclical and random in the aspect of temporal characteristics, passenger flows in certain time slots are similar and passenger flows at adjacent stations are spatially correlated. Considering the above spatiotemporal characteristics, this research proposes a deep learning model, named k-means-ConvLSTM(k-ConvLSTM) for urban rail short-term passenger flow prediction based on the ConvLSTM and the adaptive k-means clustering algorithm. Experiments are designed to optimize the key parameters of the model. Also, in order to examine the performance of the proposed model, abundant experiments are conducted based on the real passenger flow data of the Shenzhen Metro IC card. The results show that proposed k-ConvLSTM model performs better than deep learning models that only consider spatiotemporal characteristics---parallel architecture comprising the convolutional neural networks(CNN) and the long short-term memory network(LSTM), the ConvLSTM, and deep learning models that only consider temporal characteristics---the LSTM and the bi-directional long short-term memory network(Bi-LSTM), and shallow learning models--- the back propagation neural network(BPNN) and the support vector regression model(SVR), in terms of root mean square errors, mean absolute errors and mean absolute percentage errors.