Short-term Traffic Flow Prediction Based On Deep Learning

Road traffic congestion has gradually become an important factor that restricts the modern economy and the healthy operation of society.Real-time and efficient short-term traffic prediction is a key technology for implementing advanced traffic control and guidance in intelligent transportation systems.It can achieve the goal of maximizing the rational use of urban road network capacity and reducing the probability of traffic accidents.It is a powerful measure to ease traffic congestion and improve overall social and economic benefits.Therefore,it is of profound practical and theoretical significance to thoroughly study short-term traffic flow prediction methods.Short-term traffic flow is usually highly non-linear,random and periodic.Deep learning can combine the low-level features to discover implicitly distributed feature representations in the data,resulting in more abstract high-level representation attributes or features.A deep neural network with multiple hidden layers has excellent feature learning capabilities that can characterize the internal characteristics of the data.Therefore,the use of deep learning methods and models to solve short-term traffic flow prediction problems has natural advantages and strengths.The main content and research work of this paper can be summarized as follows:(1)Review and summarize the past short-term traffic flow prediction research results,summarize and compare various common traffic flow prediction methods,and discuss the origin,difficulties and significance of the short-term traffic flow prediction;(2)Describe in detail the relevant concepts of traffic flow prediction,and describe the entire process from the collection and preprocessing of traffic data to the nature of the traffic flow prediction problems,the construction,training and evaluation of the prediction models;(3)For short-term traffic flow prediction,The effects of historical speed,occupancy information and historical traffic flow data of the adjacent vehicle detection station(VDS)on short-term traffic flow prediction performance of the targetVDS were studied.Different types of features have respectively established prediction models based on LSTM recurrent neural network.(4)In order to make full use of the spatial correlation between adjacent VDS and the continuity of their own historical traffic flow sequence to improve the accuracy of short-term traffic flow prediction on the target VDS,serveal Conv-LSTM models based hybrid convolutional recurrent neural network were established.