基于秩自适应贝叶斯张量分解模型的交通流量数据修复方法

A traffic flow data imputation method based on a rank-adaptive Bayesian tensor decomposition model

针对传输线路故障、通信故障等原因造成智能交通系统在某时刻或时段无法识别到车辆,导致数据缺失的问题,提出一种基于秩自适应贝叶斯张量分解模型的交通流量数据修复方法.首先,考虑交通数据的时空相关性,基于张量模型构建数据结构.其次,使用贝叶斯模型在张量分解的参数和超参数上设置灵活的先验和超先验分布,构建秩自适应算法解决张量分解模型的秩选择问题.最后,采用长沙市车辆牌照识别(License Plate Recognition,LPR)系统记录的2019年7月1日至2019年7月28日793个交叉路口的交通流量数据,检验在不同的张量数据结构、丢失方式、丢失率以及张量分解秩的情况下该模型的数据修复精度.研究结果表明:秩自适应算法能够捕捉张量分解最佳秩的大小,避免预设秩过大导致的过拟合现象;与传统的CP分解(CANDECOMP/PARAFAC decomposition)和均值法相比,本文所提算法的平均绝对百分比误差在丢失率达到30%的情况下降低了20%,有效提升了交通流量数据修复的准确性.研究成果可为交通流量预测、交通出行时空特征分析中的数据修复提供参考.

Transmission line failures and communication breakdowns in intelligent transportation systems can result in data loss due to the inability to detect vehicles at specific times or intervals.To address this issue,this paper proposes a traffic flow data imputation method based on a rank-adaptive Bayesian tensor decomposition model.First,considering the spatiotemporal correlations in traffic data,the proposed method constructs data structures using a tensor model.Then,a Bayesian approach is employed to set flexible priors and hyperpriors on the parameters and hyperparameters of the tensor decomposition,and a rank-adaptive algorithm is developed to address the rank selection problem.Finally,the method is validated using traffic flow data recorded by the License Plate Recognition(LPR) system from 793 intersections in Changsha between July 1 and July 28,2019.This study examines the model's imputation accuracy under various tensor data structures,data loss patterns,loss rates,and tensor decomposition ranks.The results indicate that the rank-adaptive algorithm effectively captures the optimal rank for tensor decomposition,avoiding overfitting due to an excessively high preset rank.Compared to traditional CANDECOMP/PARAFAC decomposition and mean imputation methods,the proposed algorithm reduces the mean absolute percentage error by 20% even with a 30% data loss rate,significantly enhancing the accuracy of traffic flow data imputation.These findings provide valuable insights for data imputation in traffic flow prediction and spatiotemporal traffic pattern analysis.