PASSENGER FLOW PREDICTION IN METRO SYSTEMS USING LSTM

Abstract

Improved metro service and efficiency could result from more accurate
predictions of the Origin-Destination (OD) passenger riders. While OD prediction in tube
circuits has received less attention, subsequent research has concentrated on outgoing versus
incoming flow forecasts at individual stations. There are three potential origins of problems with
OD fluxes: First, there are fragmented and sparse data sets; second, there are complicated
geographical linkages and large temporal variability; and third, there are external variables.
Our proposed Flexible Function Fusion the network (AFFN) can do the following: a) accurately
represent recurrent passenger traffic patterns contingent on the auto-learned influence
compared to external factors; and b) subsequently merge geographic relationships generated by
several independently constructed knowledge graphs and undetected connections among
stations. We multi-task AFFN to tackle the sparsity and insufficient detail of OD matrices by
predicting each station's intake and outflows as a side project to increase the accuracy of OD
predictions. Two real-world metro journey datasets obtained in Xi'an and Nanjing, China, were
subjected to extensive testing. The evaluation findings show that our AFFN and multitasking
AFFN perform better than the most advanced baseline approaches and AFFN variations in
many accuracy measures. This proves that AFFN and its components are valuable for OD
prediction.