Dynamic Graph Convolutional Traffic Flow Prediction Based On Spatiotemporal Features

As an important part of Intelligent Transportation System(ITS),traffic flow prediction plays an important role and value in improving traffic efficiency.Traffic flow data is a complex spatiotemporal data,and the traffic status is also influenced by external factors such as holidays and weather conditions.Therefore,when predicting traffic flow,it is necessary to consider both the spatiotemporal and nonlinear characteristics of traffic flow data.This article focuses on the prediction of traffic flow on urban highways,fully exploring the hidden features of traffic data such as periodicity,spatiotemporal correlation,and spatial heterogeneity,further improving the predictive performance of the model.The main research contents of this paper are as follows:(1)This paper proposes a traffic flow prediction model based on spatiotemporal dynamic graph(STF-DGCN)to address the issue of insufficient extraction of dynamic spatial features of road network structure in existing traffic flow prediction models,which overlook the periodicity and nonlinearity of traffic flow data.Considering the periodic characteristics of traffic flow data,the model divides historical traffic flow data into hours,days,and weeks,and obtains dynamic road network structure information in advance through spatiotemporal embedding modules.Considering the spatiotemporal characteristics of traffic flow data,the model utilizes time graph attention networks and dynamic graph convolutional networks to extract dynamic spatiotemporal features of traffic flow data.Considering the impact of external factors such as holidays and weather conditions on traffic flow status,the model obtains external feature vectors through the external factors module,integrates them with dynamic spatiotemporal features,and outputs prediction results by the decoder.Through experiments on two datasets,Pe MS04 and Pe MS08,the results showed that the STF-DGCN model decreased MAE by 3.93% and4.15% compared to the benchmark model,respectively;Decreased by 3.21% and 6.03%respectively on RMSE;The MAPE decreased by 5.72% and 4.70% respectively.(2)This paper proposes a traffic flow prediction model based on spatiotemporal sparse dynamic graph(STF-SDAM-DGCN)to address the issue of existing traffic flow prediction models not fully mining the deep dynamic spatiotemporal features of traffic flow data,and not modeling the correlation of sensor nodes in real scenarios reasonably.The model builds a sparse directed Adjacency matrix according to the road network structure,which ensures the sparsity of the matrix and more reasonably models the correlation between nodes,the historical traffic flow data after cycle division is fused with the spatiotemporal embedded information,and the dynamic temporal features are extracted by using the attention network of time map,and the dynamic spatial features are extracted by using the dynamic map convolution network based on the sparse directed Adjacency matrix,the weighted fusion layer is used to model the relationship between historical and future time steps,assigning different weights to each historical time step to reduce errors in the data propagation process,the final prediction result is output by the decoder.Through experiments on two datasets,Pe MS04 and Pe MS08,the results showed that the STFSDAM-DGCN model decreased MAE by 6.46% and 6.41% compared to the benchmark model,respectively;Decreased by 4.79% and 8.01% respectively on RMSE;The MAPE decreased by 9.70% and 5.78% respectively.(3)In order to prove the validity of the model,comparison experiment,ablation experiment and multi-step prediction experiment were conducted on two public data sets Pe MS04 and Pe MS08.Experimental results show that the performance of the two prediction models proposed in this paper is better than that of other benchmark models,and the STF-SDAM-DGCN model has better prediction accuracy than the STF-DGCN model,which finally proves that the model can fully explore the hidden spatiotemporal features of traffic flow data and has good prediction performance.