Research On Traffic Flow Prediction Model Based On Self-attention Mechanism

In recent years,with the rapid economic development,the per capita ownership of automobiles in our country has begun to rise year by year,and the contradiction between the excessively fast automobile growth rate and the lower capacity of roads has become more and more serious.To increase the capacity of roads,the capacity of green areas and sidewalks has been reduced.This can only be a temporary relief and cannot solve the problem fundamentally.For this reason,many scientific researchers have begun to study more scientific ways to solve traffic flow problems.Traffic flow prediction is not simply affected by the number of vehicles,but also contains various complex factors,such as: time period,road conditions,and people flow.The reason for the failure is that most researchers ignore the complex factors of roads and the correlation between roads.The following problems exist in traffic flow prediction:(1)Traffic flow data had complex spatiotemporal correlation.Researchers considered time and space as a whole and ignored the characterization of a single node and a single moment.(2)In previous research,the high-order adjacency relationship between nodes was ignored by people.In fact,each node was not independent of others and has influence on each other.(3)The implicit characteristics of traffic flow data were ignored,such as periodic characteristics.Combined with the existing basic research work,this dissertation proposed a traffic flow prediction model based on self-attention mechanism-BSAGCRN.The specific researched work are as follows:Firstly,the theory of spatiotemporal decoupling was used to divide each time of each node into finer particles.The spatial dependence of each node was captured by the graph convolutional neural network and the self-attention mechanism.The time dependence of a single node at any moment was captured by the convolutional network and the self-attention mechanism.The result was depicted as a feature of a single node and a single moment.Secondly,multi-module fusion was used to mine the potential periodic relationships in the data.The time characteristics of traffic flow: recent hours,day cycle and week cycle were modeled.The data of the three modules were side connected for mutual supervised learning.Next,Through the iterative method and the residual connection,the original characteristics in the data were guaranteed and the temporal and spatial correlation of traffic flow data is acquired.Finally,GRU was used to mine the potential time relationship of the three modules.The results of the three modules are input into the GRU module by splicing,to learn the potential time series relationship in the three time series data,and to learn the future trend of traffic flow.The verification experiment was carried out on two traffic flow data sets-Pe MS04 and Pe MS08 in the Caltrans Performance Measurement System,with RMSE,MAE and MAPE as performance evaluation indicators.In the same experimental environment,we proved that BSAGCRN had better performance than the baseline model and the significance of each module through baseline model comparison experiments and ablation experiments.