Research On Short-term Traffic Flow Prediction Of Directed Road Network Based On Deep Learning

With the expansion of the city scale and the deepening of the urbanization process,the population of many super cities around the world continues to increase,which leads to an increase in car ownership,and road infrastructure construction cannot meet the growing traffic demand,resulting in urban congestion Frequent phenomena.Traffic flow prediction is an important application and research hotspot of computers in the field of traffic.It is of great significance to road traffic planning,real-time dynamic traffic signal optimization,real-time dynamic route planning,and traffic management decisions.Therefore,the problem of short-term prediction of traffic flow can provide a great help to solve the problem of urban road congestion.Although the problem of short-term prediction of urban traffic congestion has always been a focus of attention in the field of transportation,urban congestion is affected by multiple factors,and the previous methods of congestion prediction have certain limitations.In the past,when solving the problem of short-term prediction of traffic congestion,two models were mainly used.One is a model based on statistical theory,such as AR / MA / ARMA / ARIMA and Kalman filtering,and the other is based on machine learning and depth.Learning-based models based on knowledge discovery,these two types of methods have poor ability to capture the spatial characteristics of traffic flow.The emergence of graph convolutional neural network has improved the accuracy of road network spatial structure prediction,accurate to road segment level prediction,and provided feasibility for road network spatio-temporal prediction.However,T-GCN,ST-GCN and other models based on graph convolution require that the structure of the road network is undirected when performing traffic flow prediction,and cannot reflect the upstream and downstream characteristics of the road segment.Based on the above problems,this paper takes the road segment as the basic unit,and designs a multi-scale segmentation method to segment the road according to the characteristics of the road network congestion and the spatial structure of the road network,and then expresses the traffic road network as a directed graph,Use the deep learning model to predict the short-term traffic flow of the directed road network.First,the paper preprocesses the congestion data and analyzes the spatio-temporal distribution characteristics of traffic congestion.In terms of temporal distribution characteristics,congestion mainly presents a bimodal pattern.In terms of spatial distribution characteristics,congestion hotspots were used as experimental areas according to statistics.Then,this paper makes a short-term prediction of congestion in the congestion hotspot experimental area.Propose a multi-scale road segmentation method to divide roads,use the topological relationship of roads to graphically express the segmented roads,and then use a deep learning model that can fully capture spatiotemporal features for short-term traffic congestion prediction.In terms of spatial features,this paper uses a Diffusion Convolution model and Graph Attention model that can capture the spatial features of directed graphs.The temporal features are learned using GRU and LSTM models.The experimental results show that the DCRNN model has the best prediction effect in small-scale road network congestion prediction.The prediction effect of the GAT-LSTM model on the directional road network is close to it.The two prediction models for the directional road network are for HA and ARIMA.The traditional model has great advantages,and the prediction accuracy is improved compared with the traditional GCN method,which verifies the effectiveness of the model in predicting the traffic flow of the directed road network.