基于注意力机制的城市轨道交通网络级多步短时客流时空综合预测模型

Attention-based Multi-step Short-term Passenger Flow Spatial-temporal Integrated Prediction Model in URT Systems

准确、可靠的短时客流预测可为城市轨道交通提供运营决策支持。本研究以基于Transformer机制的LSTM网络、深度注意力模块和CNN网络为基础,提出了城市轨道交通网络级多步短时客流预测模型(STIPM)。该预测模型由3个分支组成,分支一以时间序列进站客流为输入,提出了基于Transformer机制的LSTM网络提取该数据中的时间相关性;分支二以基于时间步的OD数据为输入,提出了深度注意力模块挖掘数据中大量的时间、空间相关性,利用基于时间步的OD数据能够更好地展现站间联系紧密程度和全局信息,从而完成了拓扑网络信息提取;分支三的输入为POI数据,使用CNN网络获取其时空相关性,并作为时间与空间特征之间的纽带。为了保证在预测精度足够高的条件下,获得更长的预测时间和更详细的预测信息,本文采用“神经网络多输出”策略,完成了多步预测任务。本文在2个大规模城市轨道交通真实数据集中对该模型进行测试,并将预测结果与10个基准模型和4个消融实验模型进行对比,在RMSE、MAE与WMAPE评估指标中,STIPM模型均得到最高的预测精度,结果表明该模型具有一定的优越性与鲁棒性。

Accurate and reliable short-term passenger flow prediction can support operations and decisionmaking of the URT system from multiple perspectives.In this paper,we propose a URT multi-step short-term passenger flow prediction model at the network level based on a Transformer-based LSTM network,Depth-wise Attention Block,and CNN network,named as Spatial-Temporal Integrated Prediction Model(STIPM).The STIPM comprises three branches.The first branch takes time-series inflow data as input,and a Transformerbased LSTM network is selected to extract the temporal correlations.The second one takes timestep-based OD data as input,and many spatial and temporal features are captured using Depth-wise Attention Blocks.Meanwhile,timestep-based OD data can better include inter-station relations and global information.The third branch takes Point of Interest data(POI)as input and CNN network is utilized for spatiotemporal features extraction,which can also become the bridge between spatial and temporal features.Moreover,the“Multi-inputmulti-output Strategy”for multi-step prediction is used to obtain a longer prediction period and more detailed information under a relatively high forecasting accuracy.The STIPM is applied to two large-scale real-world datasets from the URT system,and the obtained prediction results are compared with ten baselines and four variants from itself,in which STIPM model achieves highest prediction accuracy indicated by RMSE,MAE,and WMAPE evaluations,which demonstrates the superiority and robustness of the STIPM.