Research On Traffic Flow Prediction Algorithm Based On Multi-dimensional Attention Mechanism

Traffic flow prediction is one of the most important tasks in urban system,which is of great significance to the management of urban system and public safety.It is widely used in traffic control,shared resource scheduling platform and intelligent transportation system.In the traditional grid-based traffic flow prediction problem,researchers divide the city into many grids of equal size,in which each grid represents each region in the city,gives the historical traffic data of each region,and predicts the traffic flow in the next day or several days.The main challenges to improve the prediction accuracy are as follows: first,the traffic flow data has a significant daily cycle and weekly cycle.How to capture the potential similarity more accurately between the prediction time and them is a major difficulty;Second,traffic flow has complex spatio-temporal correlation,and spatio-temporal dependence is dynamic.How to effectively capture this change is another difficulty.Aiming at these two difficulties,this paper proposes a traffic flow prediction model based on the interactive attention module and the spatio-temporal fusion attention module.The interactive attention module is used to learn the similarity and influence degree among closeness,period and trend,and the spatio-temporal fusion attention module is used to effectively model the dynamic spatio-temporal correlation of traffic data.The first model proposed in this paper is a spatio-temporal neural network based on interactive attention for traffic flow prediction(Att Deep STN+).Firstly,an interactive attention mechanism is proposed to learn the correlation among closeness,period and trend,and the influence degree among the three components.This mechanism can better capture the complex correlation of traffic data through the attention mechanism.Secondly,after the interactive attention layer,it uses fusion to capture the complex correlation between different levels of features.Finally,the final multi feature fusion is used in the model,which reduces the trainable parameters of the whole network model,has a clear structure,and improves the accuracy of model prediction.The second model proposed in this paper is the deep neural network based on spatio-temporal fusion attention for traffic flow prediction(ST-DSTN).By designing spatiotemporal fusion attention module,the attention mechanism adaptively assigns different weights to spatio-temporal data,which can effectively model the dynamic spatio-temporal correlation of traffic data,and use three non-shared convolution layers to convert them into the same tensor size,And feed them back to the non-shared ST module respectively,and conduct n times of iterative update processing at the same time,in order to more effectively model the spatiotemporal dynamic correlation of traffic data.Next,the three processed time features and external factor features are fused,and the fused results are sent to the Res Plus module.This early fusion method allows different levels of features to interact,to effectively capture the correlation between different features.At the end of the model,each output is fused by multiscale,to further capture the complex correlation between different levels.In addition,this paper tested the proposed deep learning method based on real-world data sets,and the results confirmed that Att Deep STN+ and ST-DSTN achieved better prediction results.