Research On Traffic Flow Prediction Based On Multi-scale Spatial Temporal Graph Convolution Networks

As an important part of the intelligent transportation system,traffic flow prediction is of great significance for maintaining the healthy operation of the transportation network.In recent years,with the rapid development of deep learning,spatial temporal graph convolutional neural networks have become the main means of traffic data prediction.Aiming at the problem that the existing algorithms express the temporal and spatial characteristics of traffic flow in a single dimension and measure the dynamic periodic relationship with fixed weights,two traffic flow prediction algorithms were proposed in this paper based on spatial temporal graph convolutional neural networks.The details are as follows:In order to solve the problem that the existing algorithms express the spatiotemporal features of traffic flow data in single dimension,this paper proposes a Multi-Scale Spatial Temporal Graph Convolution Network(MSSTGCN).Using dilated convolution,first-order approximate graph convolution and temporal attention mechanism to extract the temporal and spatial characteristics of traffic flow data.Considering that the output of the graph convolutional network constructed by "series" only expresses spatiotemporal features in a single scale,the introduction of Res2 Net with a "parallel" structure uses multi-scale spatiotemporal feature fusion to perform multiple scales of spatiotemporal features.expression.At the same time,in order to consider the periodicity of traffic flow,three periodic network components with the same structure are constructed to model the traffic flow sequence of the near-term,daily and weekly periods in the near forecast period,and the final forecast is obtained by weighted fusion result.For the purpose of solving the problem that existing algorithms use fixed weights to measure the relationship between different periodic sequences of traffic flow data,this paper proposes an Adaptive Periodic Spatial Temporal Graph Convolution Network(APSTGCN)model based on MSSTGCN.It uses the adaptive periodic feature fusion network(APFN)to improve the parameter-based periodic feature fusion strategy in MSSTGCN,and uses the attention mechanism to mine the dynamic relationship between different periodic sequences.According to this relationship,adaptive weights are assigned to the output of each component to obtain a prediction closer to the real traffic flow.Finally,the experimental results on the two public highway datasets Pe MSD4 and Pe MSD8 show that the MAE,RMSE,and MAPE of MSSTGCN on the two datasets are better than other benchmark models.And with the assistance of APFN,the accuracy of APSTGCN has been further improved,and has shown obvious advantages in long-term forecasting.In addition,the APFN proposed in this paper can also be flexibly applied to other models in order to more comprehensively mine the periodic patterns of traffic data.