Research On Traffic Speed Prediction Method Based On Spatio-Temporal Graph Convolutional Network

Intelligent transportation systems have identified traffic speed prediction as a crucial research area that can mitigate traffic congestion and enhance public transportation safety.The intricate nature of traffic speed prediction arises from the integration of spatial correlation and temporal dependence.To address this problem,researchers have applied various deep learning methods to traffic speed prediction,mainly using graph neural networks and recurrent neural networks,which can better capture spatial correlation and temporal dependence.To improve the accuracy of traffic speed prediction,improvements based on the original research are proposed in this thesis,which consider using contrastive learning to enhance model accuracy and external factors’ influence on traffic speed,instead of solely relying on fully connected networks for single-step prediction.To address the challenges and shortcomings of existing models and achieve more accurate traffic speed prediction,traffic speed prediction model based on spatial-temporal graph convolutional network combined with contrastive learning(CSTTS)is designed in this thesis,and further improves and perfects the research on the CSTTS model,traffic speed prediction model based on sequential to sequential spatial-temporal graph convolution network(SSTTS)is designed.The main research content and innovative contributions are as follows.(1)To improve the accuracy of model training,a contrastive learning method is adopted to train the model,which reduces the error in model training.Specifically,the regularization dropout method is used,which introduces the same sample and through different dropouts of the same model,the output of two sub models is restricted to the distribution of output data,so that the distribution of two data generated by the same sample in the same batch is as close as possible,and the final experimental results are significantly improved.(2)In response to the problem of most current traffic speed prediction models only conducting single-step predictions,a sequence-to-sequence model is employed in this thesis,which is an encoder-decoder structure that can alleviate the rapid accumulation of model errors caused by long inputs and multi-step predictions.Moreover,on this basis,attention mechanism is used to enhance crucial information and better explore the complex spatial-temporal correlations in road networks,thus improving the accuracy of long and short-term predictions.(3)To address the impact of external factors on traffic speed prediction accuracy,external factors such as weather in model inputs are considered in this thesis,attribute-enhanced units are used to encode and integrate external factors into the model.Experiments on real-world datasets show that considering external information in traffic speed prediction tasks is highly effective.To establish the efficacy and dependability of the enhanced model presented in this thesis,adequate comparative experiments and ablation analyses were conducted,and each model was tested using two real-world traffic datasets.Based on the experimental outcomes,models surpassed relevant baseline models in terms of performance evaluation metrics.