Study On Short-term Traffic Flow Prediction Based On Deep Learning

In order to alleviate the problem of urban traffic congestion,intelligent transportation systems have received more and more attention.The basic support for real-time decision making in intelligent transportation systems – the study of short-term traffic flow prediction becomes extraordinarily important.In the early days of the lack of available data,scholars could only rely on expert experience modeling,and their prediction ability was very limited.With the extensive deployment of traffic data sensors,a large amount of historical data has been accumulated,which has promoted the development of data-driven models.In recent years,benefiting from the rapid development of dedicated computing devices,designing data-driven models based on deep learning has been a research hotspot in the field of traffic prediction.However,a large number of studies have stopped at the short-term dependence of capture time.There is no good solution on how to consider the long-term dependence of time,and the research on the spatial correlation of road network is not yet mature.In order to fully consider the spatiotemporal characteristics,this paper designs a new type of urban highway network spatiotemporal prediction model,which has better prediction performance than the previous deep learning model.In consideration of short-term dependence,an inverted tooth grid recurrent unit is designed,and a short-term dependency model is established by combining the general framework of sequence mapping.By re-examining the state flow of the gated loop unit,it is found that the flow state information in a similar "inverted tooth shape" not only reduces the number of state variables,but also enhances the short-term dependency capture capability.In addition,the general framework design of sequence mapping is adopted,taking into account the short-term dependence between successive moments in the future.Through real data validation,the model predicts better results than the commonly used deep learning model.In the long-term dependence,the input tensor is reconstructed by shearing,and then fed into the convolutional neural network for feature up-scaling,and the short-term dependency model is connected at the top of the model,which obtains better prediction performance than the previous model.In the aspect of capturing traffic flow periodic characteristics,a periodic module combining periodic trend item extraction and inverted tooth grid recurrent module was developed,and the prediction performance was improved by incorporating the existing model.In order to combine the local spatial correlation of the road network,a new type of road network spatiotemporal prediction model is developed by introducing the diffusion convolution operator in the field of graph convolutional neural network and considering the spatiotemporal fusion method.The diffusion convolution operator considers the directionality of the road network,which is in line with the actual situation of the road network.It is also proved by experimental analysis that the performance of the spatiotemporal prediction model using diffusion convolution is better than that of the spectral convolution.At the end of the paper,a new type of urban highway network spatiotemporal prediction model is designed by considering the long-term dependence of time,periodic characteristics and spatial correlation,and the best prediction performance is obtained.