基于双模态信息的出租车需求预测
Taxi Demand Forecasting Based on Bimodal Information
陈春源 1陈鹏1
作者信息
- 1. 武汉理工大学交通与物流工程学院,武汉 430063;武汉理工大学交通信息与安全教育部工程研究中心,武汉 430063
- 折叠
摘要
现有需求预测方法对路段相关性挖掘能力不足,导致空间特征提取不充分.针对此问题,利用路段的邻接关系和需求时序形状相似关系两种模态信息分别构建路段静态拓扑图和序列形状相似图,表征更深层的路段空间关系.其次,基于图注意力网络(Graph Attention Network,GAT)与门控递归单元(Gate Recurrent Unit,GRU)构建时空多图注意力网络(Spatiotemporal Multi-graph Attention Network,ST-MG AT),以使用两个图挖掘数据的时空特征.实验表明,ST-MGAT的平均绝对误差、均方根误差和平均绝对百分误差分别为0.758 1、0.641 2'和6.976 3％,均优于对比模型.此外,该方法在短时预测方面表现更佳.
Abstract
Existing demand forecasting methods lack the capability to adequately explore the correlation of road seg-ments,leading to insufficient extraction of spatial features.To address this issue,two modalities,namely the adjacency re-lationships of road segments and the temporal shape similarity of demand time series,are employed to construct road stat-ic topology graph and sequence shape similarity graph,respectively,to represent the deeper spatial relationships of road segments.Additionally,a Spatiotemporal Multi-graph Attention Network(ST-MGAT)is built based on the Graph Atten-tion Network(GAT)and Gate Recurrent Unit(GRU).This network utilizes two graphs to uncover spatio-temporal char-acteristics in the data.Experiments show that the mean absolute error,root mean square error and mean absolute percent-age error of ST-MGAT are 0.758 1,0.641 2 and 6.976 3％respectively,which are better than the comparison model.In addition,the method has better results in short-term prediction.
关键词
出租车需求预测/城市交通/深度学习/图注意力网络/门控递归单元Key words
taxi demand forecasting/urban transport/deep learning/graph attention networks/gated recurrent units引用本文复制引用
出版年
2024
NETL
NSTL
万方数据