Research On Short-Term Traffic Flow Prediction Algorithm For Road Networks Based On Graph Neural Networks

With the accelerated process of urbanization,intelligent transportation systems have greatly facilitated people’s daily travel.Traffic flow prediction,as a fundamental and challenging task in the field of intelligent transportation,is essential for urban development as it involves understanding the dynamics of traffic flow and predicting the future state of traffic systems.Due to the complex spatial and temporal dependencies in traffic flow,the key challenge in efficiently and accurately predicting traffic flow lies in modeling the intricate spatial and temporal correlations within the traffic flow time series data.To address the insufficiency of existing traffic flow prediction models in exploring spatiotemporal features and modeling spatiotemporal relationships comprehensively,this thesis proposes two graph convolutional networkbased traffic flow prediction models that utilize multi-information fusion:1.To overcome the limitation of existing research that focuses solely on attributebased feature extraction without incorporating external factors,this thesis presents a traffic flow prediction method based on multi-information fusion and graph attention networks.This approach fills the gap in previous studies by incorporating external information into the prediction process.The nonlinear variation characteristics of traffic flow are not only related to temporal and spatial features but also influenced by various external factors,such as attributes of building clusters around sensor nodes and weather conditions.Previous research has seldom integrated external factors into prediction models.In this study,external factors are categorized as static and dynamic factors and integrated into separate modules for training.Spatial features are captured using graph attention networks,while time features are captured using a gated recurrent unit-based encoder-decoder.Experimental results demonstrate the effectiveness of multiinformation fusion and graph attention models in improving prediction accuracy.2.To address the challenge of capturing the dynamic spatial features of traffic flow,which are difficult to learn using predefined graphs in existing traffic flow prediction research,this thesis proposes a traffic flow prediction method based on multiinformation fusion and dynamic graph networks.This approach overcomes the limitation of conventional methods in capturing the dynamic changes in spatial flow patterns.The spatial features of traffic flow are often predetermined by the physical structures of roads,but the actual flow characteristics between traffic nodes not only span multiple nodes but also change over time.To tackle this issue,the thesis designs a method that synchronously performs convolutions on dynamic and static graphs.A spatial encoding module is used to embed nodes,and the resulting spatial feature vectors are input to static graph convolution and dynamic graph generation processes.The two types of graphs are combined using bidirectional graph convolutions to output node spatial features.Additionally,a multi-head attention mechanism is incorporated into the gated recurrent unit to enhance the modeling of long-term dependencies.Experimental results demonstrate the outstanding performance of the proposed model in traffic flow prediction tasks.