California Traffic Flow Prediction Based On Multi-Scale Modeling Idea And Improved LSTM

The spatiotemporal series data is the most common multivariate time series data,which is usually represented by multiple interrelated time series in the spatial dimension.The mining and prediction of this kind of data is of great significance to human activities.Taking traffic flow data as an example,this type of data usually contains the traffic data collected by the monitoring point of each section of the traffic road network for a certain period of time.Up to now,the spatiotemporal graph neural network has become a classic method to solve such problems.However,the spatiotemporal graph neural network does not have a unified highly efficient prediction model.For data with different characteristics,the same model will still produce different effects.In this paper,a spatiotemporal series model is designed according to the multi-periodic characteristics of California traffic flow data and the high correlation of data between monitoring stations.Referring to the idea of temporal graphs,this paper first converts the flow data into temporal graphs,each of which represents the state of each monitoring point in a specific time scale,and the mutual changes between the positions of points on the temporal graphs represent the changes of their mutual influence over time.Therefore,this paper refer to the LSTM model and the idea of spatial graph convolution neural network and make targeted improvements to the LSTM to adapt to the three-dimensional tensor time series data,so that the LSTM can capture the dependence of the time dimension while capturing the change of the relationship between the monitoring points.At the same time,considering the multiple periodic characteristics of traffic data,based on the periodic scale estimated from time series,this paper establishes models at different scales according to the above ideas,and finally fuses the features obtained by modeling at different scales through the scale fusing layer,and forecasts the future data.By comparing the prediction effects of different models on California highway flow data,it is found that the MSGATLSTM model proposed in this paper is superior to other models.At the same time,the ablation experiment of the MSGLSTMs series models shows that the multi-scale modeling idea and improved LSTM are beneficial for improving MSGLSTMs framework prediction performance.