Research On Spatiotemporal Traffic Flow Forecasting Based On Graph Convolutional Network

Transportation plays a vital role in socio-economic development and is an essential cornerstone for the regular operation of all aspects of our socio-economic life,as it can rationally connect production,exchange,distribution,and consumption in society.The acceleration of urban modernization and the development trend of national diversity travel have brought enormous pressure to traffic management,and intelligent transportation systems have emerged as the times require.With artificial intelligence and network communication technology,the intelligent transportation system coordinates and assists scientific decision-making,improves the perception ability of urban road traffic,and optimizes the traffic management structure.Traffic flow prediction is an important part of the intelligent transportation system.It can provide a reliable scientific basis for the acute perception of traffic congestion.Building a multi-dimensional,high-time-effective traffic flow prediction model has becomed an important role in promoting the construction of intelligent transportation systems.Traffic flow forecasting is not a simple linear regression forecasting problem.The data contains complex spatiotemporal correlation features.Traditional forecasting models based on statistical theory and machine learning are hard to fit urban traffic road network data with topological structure,and the forecasting effect is barely satisfactory.Existing deep learning type-based prediction methods cannot fully extract the spatiotemporal dependencies,and there are also defects to varying degrees in terms of realtime performance.To address the above issues,this paper would research the problem of traffic flow prediction based on the theoretical knowledge of deep learning.The main contributions of this paper are as follows:(1)This paper proposes a Spatial-Temporal Attention Graph Convolutional Network for Traffic Forecasting(STAGCN).The model uses a novel spatial graph learning module to explore the spatial topological connections in the original data from multiple perspectives,including global spatial correlation and local dynamic change properties.In addition,aiming at the gradient diffusion problem in long-sequence timing modeling,the model designs a gated timing-aware attention mechanism that focuses on the local trend change of the sequence.This mechanism uses a multi-headed attention mechanism and the local perceptual field property of the convolutional kernel to extract the local contextual semantic information.(2)To address the problems of error accumulation and the large number of redundant parameters arising from modeling the multidimensional spatial correlations,this paper proposes a Dynamic Adversarial Graph Learning with Semantic Knowledge for Traffic Forecasting(DAGS).The model combines adaptive node encoding and dynamic temporal encoding to construct a dynamic adaptive spatial graph to simultaneously capture global spatial correlation and dynamic features,reducing model complexity.In addition,the model also introduces a predefined graph structure with prior semantic knowledge and designs a spatiotemporal graph fusion module to measure different types of spatial graph structures,further improving the extensive applicability of the graph convolution model.Finally,the model uses a generative confrontation training framework to alleviate the problem of error accumulation from a global perspective and uses a discriminant network to fit the data distribution of the authentic samples,which improves the convergence speed of model training.(3)We evaluate the performance of the above models on the traffic datasets,and the prediction accuracy is significantly improved compared with each benchmark model,and the model also achieves good performance in fitting the data changes.In this paper,the validity of each module is tested by conducting ablation experiments with different modules of the model,and the excellent predictive performance of the proposed model is validated through a series of comparative experiments.