Traffic Forecasting Based On Graph Neural Network

In recent years,spatiotemporal forecasting has been widely used in various fields.Traffic forecasting is an example of this kind of learning task.The so-called traffic forecasting is to predict the future traffic speed given the historical traffic speed and the basic road network.A good long-term traffic flow forecast will provide great convenience for traffic planning and save a lot of time for people’s travel.Accurate and timely prediction of multi-scale traffic conditions is also of great importance to road users and regulators.However,the prediction of traffic flow is very challenging.On the one hand,the road network in the real world has a complex spatial structure,and the road traffic flow is noneuclidean and oriented.On the other hand,the change of road traffic flow with time is not stable,with strong time dependence,such as morning and evening rush hour,holidays or unpredictable traffic accidents will have an impact on the road traffic speed.With the development of deep learning technology,the traffic forecasting method based on deep learning has attracted more and more attention.In recent years,the emergence of graph neural network also provides a new idea for spatial dependence modeling of road network.In this thesis,a traffic flow prediction model based on graph neural network is proposed by combining space dependent modeling with time dependent modeling.Firstly,a traffic flow prediction model based on spatial and temporal characteristics is proposed,which extracts the spatial characteristics of traffic flow in the road network with two different graph neural networks.Then a simple and powerful variant gated recursive unit(GRU)in the recurrent neural network(RNN)is used to model the time dependence of traffic flow in the road network.Finally,the sequence to sequence(Seq2Seq)model is used for long-term prediction to generate the final prediction results.Then,an attention mechanism is introduced on the basis of this model to optimize the spatial dependence modeling of road network.At the same time,an automatic encoder mechanism is introduced in the prediction to reduce the complexity of the model,and a planned sampling mechanism is introduced in the training to improve the prediction accuracy.Finally,an optimized traffic flow prediction model is proposed.In this thesis,a large number of experiments are carried out on three real world traffic data sets.By comparing the existing traditional machine learning methods and deep learning methods,three traffic flow prediction and evaluation indicators are used to verify the performance of our proposed model.The experimental results show that the model in this thesis has achieved the best prediction effect on all the evaluation indicators,and it is superior to other methods both in the prediction accuracy and in the complexity of the model,especially in the long-term prediction,the superiority of the model will be more prominent.