A Research Of Traffic Volume Prediction Method Based On Deep Learning

Traffic congestion is a major problem in today's society.Intelligent transportation system is an emerging major means for traffic congestion,and traffic volume prediction is an indispensable module in the system.Traffic volume prediction is to analyze and summarize the rules based on the existing traffic data so as to predict the traffic volume in the future for a period of time.There is a guiding significance for the urban traffic planning,enterprise transportation scheduling,and users to provide travel and other fields.Such as,motor vehicle traffic volume prediction can help government traffic administration realize the city's congestion areas and take appropriate action in advance.Due to the increasing scale of traffic data in recent years,traffic volume prediction has attracted more and more attention from artificial intelligence researchers,who have proposed many methods,but the existing methods still have limitations.The existing methods do not take into account the characteristics of the spatial dependence changing dynamically with time and the periodicity of the temporal dependence.This paper describes the architecture and details of the traffic volume prediction based on deep learning model.In the model,we use both the local Convolutional Neural Network and the Long Short-Term Memory network to model the local spatial dependence and short-term time dependence of traffic data.The flow control mechanism is introduced to model the spatial dependence of dynamic changes between regions.The attention mechanism of periodic shift was introduced to model the long-term periodic dependence and periodic time shift which had been neglected in previous studies,so as to achieve more accurate prediction.Next,we will summarize the research content of this paper:(1)The architecture and details of the traffic volume prediction model based on deep learning are described in detail.In the model,we use both the local convolutional neural network model and the long and short time memory network model to model the local spatial dependence and the short-term temporal dependence of traffic data.And we take the external characteristics into account to obtain more accurate prediction.Experimental evaluation results show that comparing with 11 other baseline models(HA,ARIMA,KF,RR,XGBoost,MLP,CNN,local CNN,LSTM,ST-Resnet,Conv LSTM),our model is more accurate.This validates the effectiveness of our model by integrating the spatio-temporal and external features of traffic data.(2)We introduce a flow gated mechanism to model the dynamic spatial dependence between regions and a periodic shifted of attention mechanism to model long-term periodic dependencies and periodic time shifts that have been ignored in previous studies to achieve more accurate forecasts.The Rooted Mean Square Error(RMSE)of the start traffic volume on the Beijing taxi trajectory dataset and the New York City rental bike trajectory dataset are 12.17 and 4.81,respectively.Our model reduces the prediction error of 5.9% and 6.2% based on the original local CNN + LSTM combination model.Through extensive experiments,this paper proves that our proposed deep learning model can accurately predict road traffic volume.And its excellent performance on two vehicle datasets proves its versatility.Therefore,this paper has both theoretical research value and practical application significance.