Traffic Flow Prediction Based On Spatiotemporal Multilevel Graph Convolution Neural Network

Traffic flow prediction is an important part of intelligent transportation system,and accurate traffic flow prediction plays an important role in smart city,traffic control,road planning and other related decisions.Due to the spatio-temporal dependence,uncertainty and complexity of traffic flow,the accuracy of existing single-step longtime prediction models is relatively high but still has room for improvement,while the prediction accuracy of multi-step prediction models is not satisfactory,and the training cost of existing related models,especially those based on deep learning,is generally high.To this end,this paper constructs spatio-temporal traffic flow prediction models with different prediction steps based on graph neural networks,and its main research contents are as follows:(1)A lightweight spatio-temporal graph convolutional neural network model based on incremental edge discarding is proposed to address the shortcomings of singlestep models in long-time traffic flow prediction and their potential accuracy improvement.First,an incremental edge discarding strategy is used to build a graph convolution module to extract high-order spatial information of road network;then,a single "sandwich" spatio-temporal convolution module is built by combining temporal causal convolution to extract spatio-temporal features of traffic flow while keeping the overall structure lightweight;finally,a dynamic initial learning rate is introduced in the model training to further improve the optimizer Finally,the dynamic initial learning rate is introduced in the model training to further improve the performance of the optimizer and ensure the overall superiority of the proposed model.The proposed model is lightweight,simple and effective.(2)A feature fusion multi-module graph self-attentive network model is proposed to address the limitation of the existing multi-step traffic flow prediction model with high training cost and relatively low accuracy.First,a feature fusion layer is constructed to smooth the input features and improve the robustness of the model against missing data and noise;then,the main body of the model adopts a multi-module parallel structure with the same structure but different layers of graph convolution to form a multi-level graph convolution;meanwhile,a dynamic self-attentive network is built in the sub-module to capture the dynamic correlation between graph nodes with finer information granularity;finally,the adaptive combination of multiple modules to effectively integrate the output.The proposed model captures the spatio-temporal information of the road network in a multi-level and dynamic manner,which is accurate and efficient.(3)The proposed model and recent related baseline models are subjected to detailed comparative experiments and ablation analysis on several publicly available benchmark data.The experimental results show that the components involved in the model are well designed,the model is effective and feasible as a whole,and the prediction accuracy of the proposed model is higher and the training time is shorter compared with the baseline model.In summary,this paper designs separate prediction models based on graphical neural networks for single-step and multi-step traffic flow prediction problems,taking into account the global characteristics of the road network while capturing local spatial features,and the final model incorporates a lightweight timing module to accurately capture the spatial and temporal characteristics of the road network,while considering the structural complexity,training speed,and effectiveness of each component of the model.