Long-term Traffic State Prediction Of Urban Road Network Based On Deep Learning

In current life,traffic congestion is getting worse.The goal of the intelligent transportation system(ITS)is to reduce the traffic congestion.The traffic state prediction is the basis for the traffic control and optimization,which plays a very important role in ITS.Because the transportation optimization plan requires time to design and deploy,and the expansion of the city scale also makes people a single trip longer,so long-term multi-step prediction of traffic state is more important.Besides,traffic state is a variable affected by many factors,especially the characteristics of temporal and spatial dimensions.Current studies focus on short-term prediction,and less studies on long-term prediction with more than half an hour and consideration of other features.In this paper,the traffic state is represented by the traffic speed,and the long-term multi-step prediction of urban road traffic state is studied in consideration of the influence of spatial-temporal correlation and other features on the traffic state.The main work is as follows1)Data preprocessing and correlation analysis on traffic state data were completed.Firstly,due to the poor quality of the traffic state data,data preprocessing such as outlier detection,missing value filling,and data sampling was completed.Then,the temporal and spatial correlations of traffic state were analyzed in detail.The correlations between traffic state data and other auxiliary features were also analyzed such as hours,holidays,road types,weather,etc.2)A multi-step prediction model of urban road network traffic state was proposed considering the temporal and spatial correlation of traffic states.Firstly,this paper proposed a graph convolutional network.This method can capture the spatial characteristics of the traffic network and flexibly process the upstream and downstream sections separately.Then the long short-term memory networks(LSTM)and sequence to sequence(Seq2Seq)structure were further integrated to realize multi-step prediction of traffic state,which took into account the spatial-temporal correlation.The introduction of the attention mechanism can enhance the effectiveness of the model and obtain the importance of each step in the historical data.Finally,the effectiveness of the proposed model was illustrated by comparing the effects of different models on the dataset,and the results were analyzed in detail.3)Based on the temporal and spatial correlation-based trafic state prediction,a multi-step traffic state prediction model considering other auxiliary features was further proposed.Firstly,one-hot encoding was performed on the category-type auxiliary features to achieve the discretization of the category features.For the situation that the dimensions become larger after discretization of features,the dimensions were reduced based on the Embedding idea,and the data was combined with traffic state data as the input of the prediction model.The experiment's results showed that the proposed method could achieve good multi-step prediction results with a small increase in the amount of parameters.At last,this study analyzed and explained the weight matrix used in the auxiliary feature dimension reduction and the output embedding vector.