Urban Flow Prediction Based on ConvLSTM and Capturing Periodicity

Nowadays, due to the availability of large-scale traffic data and the advancement of deep learning technology, urban flow prediction based on various deep learning methods has become the focus of academic circles and industry. Different from other spatio-temporal forecasting tasks, urban flow prediction has its own unique characteristics, such as the strong periodicity of flow sequences and the non-strictness of periodic patterns. In this paper, we propose a method called PCST based on ConvLSTM and capturing periodicity, which employs a periodically shifted capture module and another attention-based fusion module to make full use of periodicity. We evaluate our PCST in real-world traffic dataset and the results demonstrate the effectiveness of the proposed methods.