A Network Model Of Spatio-temporal Attention Mechanism Based On Missing Value Complementation Of Traffic Data

Traffic congestion prediction is an essential part of smart cities,which can help urban traffic management to make more accurate road control measures in advance and thus reduce the direct or indirect losses caused by traffic congestion.However,this task becomes extremely challenging due to the following three problems:(1)data missing problem: important parameters such as traffic congestion data will be missing due to the sparse distribution of sensors in the traffic road network and possible distortion of sensor data;(2)sample imbalance problem: when the temporal and spatial granularity of the traffic congestion prediction task is fine to the minute level and road level,respectively,the the proportion of the sample size of traffic congestion will be small;(3)data mutation problem: when fine-grained traffic congestion prediction,the traffic conditions of adjacent road sections are sometimes completely different,making the existing methods based on spatio-temporal smoothing fail.In this paper,we propose a novel model Meta-ST-I based on attention mechanism,which can make more accurate prediction of minute-level traffic congestion events at road granularity in the absence of data.To solve the problem of missing data,this paper proposes the Meta-I module,which can use the meta-semantic information of roads to calculate the similarity among roads,thus making the model have the ability of data complementation;to solve the problem of sample imbalance,this paper adopts an undersampling strategy,which randomly samples the non-congested samples and slow samples,so that the data amounts of congested samples,non-congested samples and slow samples are close to each other,thus In order to solve the problem of data mutation,this paper proposes Meta-ST module,which not only uses the dynamic features of each road,such as traffic speed,traffic flow,traffic congestion,etc.,but also uses the static features of each road,such as road length,road width,speed limit level,etc.,when calculating the spatio-temporal similarity,so as to better This allows for better access to important information about adjacent road sections.The experimental results in a realistic dataset-Xi’an city congestion dataset confirm the effectiveness of the proposed model.In particular,the proposed Meta-ST-I significantly outperforms the existing SOTA model in the classification task F1 index for slow speed and congestion event prediction,which indicates that the proposed Meta-ST-I can not only handle the data missing problem and sample imbalance problem,but also well handles the data mutation problem in fine-grained traffic prediction.