Analysis And Application Of Traffic Travel Big Data Based On Deep Learning

Owing to the original urban planning is simple,so that it cannot be suitable with the sophisticated circumstance of city at present.Therefore,the rapid speed of city development brings many traffic problems,such as traffic congestion,the unbalanced distribution of carrying capacity.These questions are not just hindering the growth of city and also slowing down the efficiency of operating of society.Hence,in order to accelerate the running speed of city and prevent the traffic jam,people ought to predict travel demand in the future,which is the foundation of traffic resource dispatched.While,with the progress of sensor and internet,many vehicles produce enormous traffic spatiotemporal data.The actual traffic condition can be reflected by these data,and it provides the research basis for analyzing.Thus,how to utilize these large number of data to predict the travel demand and offer a reference which is used for travel demand dispatched becomes the hot spot of current research.In the traditional domain of travel demand prediction,the classical time series forecasting methods is popular,such as AR,VAR,ARMA,ARIMA and so on.These old avenues only make use of the temporal feature in data,which does not explore the internal condition of data.At the same time,since these models are not easily to achieve convergence,so they often need long training time.Thus,these models have no ability to dispose of complicated scenarios.Nowadays,deep learning infuses extraordinary motivation to many fields,and makes great achievements in speech recognition,computer vision and so on.Thereby deep learning is brought into the traffic forecasting domain to improve the accuracy of prediction.However,many deep learning models only can catch one characteristic between spatial feature and temporal feature.Consequently,there is big space for improving the forecasting.Therefore,in order to break up these obstructions,this thesis proposes a novel forecasting traffic framework DeepSTCL(Deep Spatiotemporal ConvLSTM)which based on ConvLSTM neural network.This framework has an eminent performance in forecasting,because it can both gain the temporal and spatial attributes of data.DeepSTCL consists of three branches which named closeness,period and trend.These branches have same components,but they do not share weights each other.When they attain the branch result,the linear fusion method will be used to converge three middle results to an ultimate result.For the sake of checking the efficiency and accuracy of framework,this thesis finishes the corresponding experiments based on order data set of Cheng du from DIDI incorporate.The experimental results show that DeepSTCL overwhelms the other baseline models in three different evaluating indicators.Simultaneously,this thesis also explores how different ways of data processing influence the accuracy and efficiency,and then offers suitable advices for others.Finally,owing the visualization of traffic data can make people know more about current traffic situation,and then give the reasonable opinions for traffic resources dispatching.So that this thesis devises a visualization system based on the specialization of spatiotemporal data.This system makes the distribution of travel demand clear and intuitive,which can lead the peak of travel demand be found very quickly,thus the traffic resources can be arranged more smartly.