Citywide Crowd Flow Prediction Based On Spatio-Temporal Trajectory Data

The continuous urbanization process and the rapid growth of urban population have brought not only vitality but also burden to cities.By predicting the citywide crowd flow,we can allocate resources,manage traffic and maintain public safety in a more reasonable way,which is one of the important ways to reduce the pressure of cities and improve the quality of life and happiness of people.Therefore,citywide crowd flow prediction has gradually become a popular research topic in academia.However,since the problem is a nonlinear prediction problem,it not only has complex spatio-temporal dependencies,but also is influenced by various external factors,which makes the problem still does not have a perfect enough solution.Therefore,this paper will investigate the problem in detail and propose a new method for citywide crowd flow prediction based on spatio-temporal trajectory data.First,this paper proposes a single-step citywide crowd flow prediction model based on deep residual networks and long short-term memory networks.One of the biggest difficulties of the citywide crowd flow prediction problem is that the problem has a complex nonlinear spatio-temporal dependencies,making it difficult for general models to capture the changing characteristics of crowd flow.The proposed prediction model captures the spatial dependencies of near and distant locations through the depth residual module and the temporal dependencies through the long short-term memory module,and takes into account the influence of external factors on flow changes in the training process of the model,so that the crowd flow in citywide areas can be captured and predicted more accurately.Secondly,this paper further investigates the multi-step prediction problem of citywide crowd flow based on the proposed single-step prediction model,uses the optimized prediction model to make multi-step prediction of crowd flow,analyzes and summarizes the influencing factors and difficulties in the multi-step prediction problem through experimental results.Finally,based on the studies of multi-step prediction problem,this paper proposes a multi-step citywide crowd flow prediction model combined with sequence-to-sequence model by considering the citywide crowd flow multi-step prediction as a sequence-to-sequence learning problem.In this model,the cumulative transmission of errors in the multi-step prediction problem is mitigated by using scheduled sampling for decoding in training process,and attention mechanism is added to enhance the model’s learning of the time dependencies of the crowd flow changes,which enables the model to make more accurate predictions of citywide crowd multi-step flow.In this paper,the single-step prediction model is trained and tested using Taxi BJ and Bike NYC datasets,and the prediction results are compared with several benchmark models,the experimental results prove that the single-step prediction model proposed in this paper outperforms the comparison models.this paper uses optimized prediction model to perform multi-step prediction in different ways,and conduct an in-depth studies of the citywide crowd flow multi-step prediction problem through a large number of experiments,and propose a multi-step citywide crowd flow prediction model based on the results of the study,and it is demonstrated that the proposed multi-step prediction model converges faster than other comparative models during training,and has lower prediction error in multi-step prediction.