一种利用ConvGRU解决交通流量预测问题的方法

Novel Traffic Flow Prediction Method Based on ConvGRU

准确的交通流量预测可协助交通管理部门工作,减少交通压力.但现有方法对交通流时间特征与空间特征学习不充分,将二者割裂讨论,忽视了时间与空间的动态相关性.针对以上问题,本文提出了基于ConvGRU的多时间尺度时空卷积交通流预测方法(MTSTC),设计浅层时空卷积模块对数据中的时空相关性进行初步提取;提出以ConvGRU为核心特征提取器的深层时空卷积模块,对数据的时空特征进行更深层次挖掘;并从3种时间尺度范围的数据中提炼交通流的周期性特征;结合注意力机制设计了时空注意力模块辅助模型训练,提升模型收敛速率.在公开数据集PEMS04和PEMS08上进行实验验证,结果表明采用MAE和RMSE评价指标时,本文方法的准确率相较基线方法在两个数据集上提升了3.23%~5.64%.

Traffic flow prediction is an essential part of intelligent transportation system.Accurate traffic flow prediction can not only assist traffic management to reduce traffic pressure but also avoid congestion and traffic accidents.The current methods do not learn the temporal and spatial features of traffic flow adequately,and learn them in isolation,ignoring the intrinsic relationship between them.To address the above issues,this paper proposed a multi-scale spatio-temporal convolutional traffic flow prediction method based on ConvGRU(MTSTC).A shallow spatio-temporal convolution module is conceived to perform preliminary extraction of spatio-temporal correlations in the data;a deep spatio-temporal convolution module with ConvGRU as the core feature extractor is constructed to capture the spatio-temporal features of the data at a deeper level;the periodicity of traffic flow is refined by distilling data from three time scale ranges(recent,current day,and current week);the spatio-temporal attention module is designed in conjunction with the attention mechanism to assist model training and improve the convergence rate.Experiments are conducted on two public datasets,PEMS04 and PEMS08,and the results demonstrate that the MAE and RMSE of MTSTC are reduced by 3.23%~5.64%on the two data sets respectively,compared to the baseline method.