Research On Urban Traffic Flow Prediction Based On Deep Spatio-temporal Attention Neural Network

In recent years,the rapid development of the economy has led to the transformation of public lifestyle.Meanwhile,the social traffic network has become more and more complex,and people’s awareness of travel efficiency and public safety has been continuously improved.Therefore,an accurate transportation traffic forecasting is not only of great significance to the government in traffic management and safety precautions,but also plays a pivotal role in the travel experience and personal security of the public.For example,accurate prediction of taxi demand can improve the efficiency of vehicle scheduling and reduce traffic congestion.Accurate prediction of shared bicycles can reduce operating costs,and accurate crowd flow prediction can prevent stampede events in advance.Traffic flow prediction is to predict future traffic values by building a deep learning model given a series of historical observation data.The main challenges in improving the prediction accuracy are as follows.First,traffic flow data has the features of proximity,periodicity and trend,and how to more accurately capture the potential similarity between the predicted moments and them is a big difficulty.Second,traffic flow,as spatio-temporal sequence data,has mutual influences in time and space,and how to capture the spatio-temporal features more accurately is another difficulty.Aiming at two difficulties,this paper proposes two traffic flow prediction models based on spatio-temporal feature extraction module and attention module.Spatio-temporal feature extraction module is used to capture spatio-temporal features,and attention module is used to obtain the potential similarity between predicted data and historical data(proximity data,periodic data,and trend data).The first model scheme proposed in this paper is the traffic flow prediction network(TDAConv LSTM)based on time-dependent attention module(TDAM)and spatio-temporal characteristic extraction module(STCEM).Firstly,the spatio-temporal characteristic extraction module is composed of convolutional neural network(CNN)and convolutional LSTM network(Conv LSTM),which is used to model complex spatio-temporal features.Secondly,the time-dependent attention module is used to pay attention to the potential law of proximity data and prediction data.Finally,the output of the time-dependent attention module and the output of the last time of the spatio-temporal characteristic extraction module are fused into the deconvolution neural network(DCNN)to obtain the final prediction data.The second model proposed in this paper is the traffic flow prediction network(MAPred RNN)based on multi-attention mechanism(MAM)and new spatio-temporal characteristic extraction module(NSTCEM).Firstly,the new spatio-temporal characteristic extraction module is composed of convolutional neural network(CNN)and predictive recurrent neural network(Pred RNN)to capture the spatio-temporal features of traffic data.Secondly,the multi-attention mechanism includes proximity attention module,periodic attention module and trend attention module,which are used to capture the potential laws of historical traffic data(proximity data,periodic data and trend data)and predicted data.Finally,the output of proximity attention module,the output of periodic attention module,the output of the trend attention module and the output of the new spatio-temporal feature module are aggregated and put into the deconvolution neural network(DCNN)to get the final prediction result.In addition,based on two real-world data sets,the two proposed deep learning methods are tested,and the results show that TDAConv LSTM and MAPred RNN achieve better prediction results.