Research On Traffic Flow Prediction Based On Graph Convolution Deep Learning Model

Over the past few decades,the urbanization and motorization in Chinese cities has been accelerating,which has brought about a series of problems such as traffic congestion and pollution and posed severe challenges to urban management.On the other hand,with the continuous improvement of living standards,people’s expectations for travel are also rising.The imbalance between traffic supply and demand is increasingly acute.In the 1990 s,the United States took the lead in putting forward the concept of intelligent transportation system(ITS),which represents an effective integrated traffic management system through advanced information collection and processing technology,improved the service level and efficiency of existing transportation facilities,and has been widely promoted and applied in various regions of the world since then.Accurate and real-time traffic flow prediction is the premise and basis of intelligent traffic system.Due to the complex formation of traffic flow,which is influenced by weather conditions,road conditions,holidays,travel choices and other factors,traffic flow data present complex and highly nonlinear characteristics,which makes the prediction task quite hard.The traditional prediction method is only based on time dependence,ignoring the influence of road space attribute.The latest research,such as spatial-temporal graph convolution network,integrates both of them,but the periodicity of traffic flow is not considered in the model structure,leading to the phenomenon of lagging prediction.Taking time dependence of traffic flow,spatial dependence and periodic characteristics into consideration,this paper puts forward a new deep learning architecture.It uses an graph convolution layer to extract spatial characteristics,followed by recurrent neural network to extract short-term dependence.We also use a parallel skip-recurrent neural network to model the periodic feature.The final prediction is computed then by concentrating outputs of these two layers.In this paper,two real traffic flow data sets in California,USA and Wujiang district,Suzhou,China are explored and used to verify different model’s performances.First,descriptive and exploratory analyses are performed on the two datasets to illustrate the temporal and spatial dependence of the traffic flow data.Secondly,four methods,HA,ARIMA,LSTM and STGCN,were used as the benchmark method to make predictions for horizon 3,horizon 6 and horizon 12 on two data sets,and the prediction effect was compared with that of the method proposed by this paper.Experimental results show that our model based on graph convolutional recurrent neural network has the optimal prediction performance.Since it takes periodicity into consideration,this is more obvious in long term prediction.On two data sets,STGCN achieve the best metrics in horizon 3,while our model is suboptimal.Whereas for horizon 6 and 12,this paper models outperform others.With horizon 12,our model’s mean absolute percentage error,MAPE,reduced by more than 10% than LSTM and STGCN,showing that this method can effectively improve the long-horizon prediction performance of short-time traffic flow.