Research On Long-term And Short-term Traffic Flow Prediction Technology Based On Transformer Networ

As urbanization rapidly progresses,the transportation system plays an increasingly vital role in the functioning of cities.With the continuous growth of population and vehicles,the pressure on the transportation system continues to mount.The high-quality operation of the transportation system is of utmost importance for economic development.The rapid advancement of artificial intelligence technology has led to the increasing intelligence of transportation systems.In the realm of intelligent transportation system research,traffic flow prediction serves as a fundamental component.However,traffic prediction tasks encounter numerous difficulties and challenges due to the dynamic nature of traffic conditions and the complex spatio-temporal correlation of traffic data.Traditional traffic prediction methods struggle to effectively leverage the underlying interdependencies among traffic data.With the swift progress of deep learning,many deep learning-based models have been employed to tackle traffic prediction problems,yet prediction accuracy still falls short.This paper focuses on the research of long-term and short-term traffic flow prediction techniques based on the Transformer network,proposing two prediction models that yield superior traffic prediction outcomes.The primary research contents of this paper are as follows:Firstly,this paper introduces an advanced traffic flow prediction model called LongShort Term Network(LSTN),which is based on the Transformer network.The model effectively captures the long-term dependencies in traffic data by incorporating both recurrent and skip recurrent connection modules,along with the multi-head attention mechanism of the Transformer network.It processes the long-term and short-term traffic data separately,considering their temporal proximity.Experimental results on real-world datasets demonstrate the excellent performance of the model in both long-term and shortterm traffic flow prediction tasks.Compared to existing models,the proposed approach achieves higher prediction accuracy,particularly for long-term traffic flow prediction,highlighting its significant performance advantage.Furthermore,in order to better model the spatio-temporal dependencies in traffic data,this paper extends the LSTN model and introduces an enhanced traffic flow prediction model called TE-TCN,which is based on the Transformer Spatio-Temporal Convolutional Network.The TE-TCN model incorporates GCN(Graph Convolutional Network)to effectively capture the spatial dependencies by modeling the road network structure,addressing the limitations of the LSTN model in terms of spatial dependency modeling.Furthermore,the TE-TCN model utilizes parallel temporal convolutional networks and the Transformer network to capture short-term and long-term temporal dependencies in traffic data,respectively.Lastly,experimental evaluations are performed on datasets including Pe MS03 and etc.The results demonstrate the exceptional performance of the model in both long-term and short-term traffic prediction tasks.Additionally,by considering spatial dependencies,the proposed method further enhances the accuracy of predictions.