Short-term Traffic Forecasting Of Expressway Based On Spatial-temporal Graph Convolution

Accurate traffic flow forecasting is of the great importance for management and decision-making for intelligent traffic system.In fact,owing to the highly nonlinear and complex model shared by the traffic data,it is no easy task to predict the forthcoming traffic conditions.Methods like classical Mathematical Statistics Method,traditional machine learning algorithms and basic deep learning algorithms all have certain limitations in the face of a tremendous amount of traffic data with multiples of characteristics.In recent years,researchers have developed the spatial-temporal graph convolution models to capture the spatial-temporal characteristic of traffic data.By using the predefined adjacency matrix as the description of graph structure,graph convolution neural networks can mine the potential spatial correlations between nodes based on adjacencies and find well application in terms of predicting tasks.However,much of the graph adjacency matrix based on graph convolution models tends to be single and fixed,which makes it hard to accurately quantify the correlation between nodes and then comprehensively to describe the complex road network topology structure in the real world.At the same time,in order to effectively improve the accuracy of the traffic flow forecasting model,researchers have widely introduced the attention mechanism into the modeling of spatial-temporal graph convolution,which leads to an increase of the complexity of the model computing.In response to those issues,this paper has put forward an adaptive graph learning algorithm based on spatial-temporal embedding to improve the express capability of traffic roadmaps in the real world.Based on the above-mentioned analysis,this paper has proposed a spatial-temporal graph convolution traffic flow forecasting model added by the attention mechanism to make sure the model accuracy without increasing the complexity of time computing.The main research contents are as follows:Firstly,this paper has proposed the spatial-temporal characteristics of highway’s traffic flow.Based on the highway’s public data collections,this paper has analyzed the normal characteristics of traffic flow,then using the method of Pearson correlation coefficient to make further analysis of spatial and temporal characteristics,preliminarily mastering the spatial-temporal characteristics of highway’s traffic flow,which is conducive to follow-on forecasting models and purposedly setting up models for the spatial-temporal characteristics.Secondly,this paper has put forward the adaptive graph learning algorithm based on spatiotemporal embedding,which composes of multiple traffic flow forecasting modules and parametric graph learning modules.By developing parametric graph learning modules,possible correlations between nodes can be captured adaptively in the process of training.By optimizing the parameters of two modules in the form of alternating training and graph learning by way of redesigned loss function to ensure the sparsity of the generated affinity matrix.Finally,the optimal affinity matrix is generated by updating the graph structure in weighting method.The proposed AdapADJ algorithm can be applied to most traffic forecasting based on graph convolution-based networks.Experimental results carried out on the two datasets show that the AdapADJ model can not only further improve the accuracy of traffic forecasting,but also effectively mine the hidden correlation between nodes.Finally,this paper has put forward the spatial-temporal graph convolution traffic flow forecasting model added by the attention mechanism.The model consists of three separate parts that simulate hourly,daily,and weekly time-scale characteristics of traffic flow,with each component effectively capturing the dynamic spatial-temporal correlation in traffic data through the spatial-temporal attention module.What is to underscore is that the attention mechanism has replaced the multi-head attention mechanism with a locally sensitive hash,thereby reducing the time-computing cost to capture deep-level characteristics through spatial-temporal graph convolution module.In the final analysis,forecasting results are obtained after weighted fusion of the three dimensions.Experimental results on four data sets show that the proposed STRGCN model is superior to the baseline models in terms of efficiency and accuracy.