Research On Crowd Flow Prediction Based On Spatio-Temporal Attention Network

As an integral part of intelligent transportation system,crowd flow prediction can effectively help government departments integrate and utilize existing transportation resources,in terms of urban traffic management and public transport safety,they play a very important role.However,due to the complexity of crowd flow data,it is not an easy task to accurately predict the crowd flow in a certain period of time in the future.Specifically,crowd flow is mainly affected by three aspects: time dependence,space dependence,and external factors such as weather,holidays,and accidents.How to more accurately explore the internal relationship of each factor and how to integrate multiple factors for predicting crowd flows is already becoming an essential research issue in order to improve the accuracy of forecasts.In attempt to overcome this problem,researchers have come up with many reasonable solutions,and the attention mechanism has become an important research point due to its good achievements in other fields of deep learning.Since its introduction,researchers have focused a significant amount of interest on the attention mechanism,mainly to enable the model to pay attention to the important parts of the data and selectively ignore some of the unimportant information.The introduction of attention mechanism can better exploit the time dependence and spatial dependence of crowd flow data,and thus improve the accuracy of crowd flow prediction.Therefore,based on the above multiple considerations,this paper proposes the following two deep learning frameworks.(1)In this paper,a Conv LSTM-based spatio-temporal attention network for traffic flow prediction is proposed.This model is based on Conv LSTM,and a new temporal attention module and spatial attention module are constructed for better capturing the temporal dependence and spatial dependence in traffic flow data.For external factors,this paper introduces a preliminary feature extraction module and an information fusion module.The preliminary feature extraction module can extract contextual features from the original traffic flow data and fuse them with external factors,while the information fusion module dynamically learns the weights of the prediction results of different time channels and fuses the prediction results of different channels by external factors.Experimental results demonstrated the CLSTAN model outperformed than base-line methods.(2)In this paper,a spatio-temporal augmented learning model based on AGCRN for crowd traffic prediction is proposed.Unlike the AGCRN model that completely adopts an adaptive approach to learn and update the model,the model is designed with two enhanced learning modules,a spatial enhanced learning module and a temporal enhanced learning module,which are used to help the whole updating iteration process of the AGCRN model.And in order to avoid the introduction of the auxiliary learning module to bring a large number of learnable parameters,the details are designed in an adaptive way,so that the number of parameters does not increase much.The experiment results showed the model achieves the best experimental results for the two publicly available datasets,Pe MSD4 and Pe MSD8.