Abstract: Subway short-term passenger flow forecasting is an essential component in urban subway operation,and it aims to forecast the passenger flow of subway stations in a short time in the future.Aiming at the problem that the existing methods fail to make full use of the passenger flow information of stations,a short-term subway passenger flow forecasting method based on temporal knowledge graph embedding combined with residual network and long short-term memory network is proposed,which is called TKG-ResLSTM.First,we use subway passenger flow data to construct a temporal knowledge graph of subway passenger flow,and apply the graphical embedding of temporal knowledge to obtain the dynamic patterns of subway stations passenger flow.Then,the extracted dynamic patterns of passenger flow are converted into dynamic similarity matrices and applied to theforecasting architecture of subway passenger flow based on deep learning to complete the subway passenger flow forecasting task.Finally,experimental evaluations are carried out at time granularities of 10 min,15 min,and 30 min using the Beijing subway and city A subway passenger flow datasets,respectively.Experimental results show that TKG-ResLSTM can effectively extract the dynamic patterns of subway stations passenger flow.Without using external information,TKG-ResLSTM reduces the root mean square error of forecasting by 0.41 compared with ResLSTM in the time granularity of 10 min of the Beijing subway dataset.

Key words: Deep learning, Temporal knowledge graph, Spatial-Temporal forecasting, Dynamic embedding, Citywide metro network