Short-time Passenger Flow Prediction Model of Urban Rail Transit Considering Multi-timescale Features

Current prediction models on short- time passenger flow of urban rail transits always ignore the period dependence of data in feature construction. To address this problem, a hybrid deep learning model (GRU-Transformer) considering multi-timescale temporal features is proposed. The model consists of two blocks in parallel, a GRU neural network with added attention mechanism (Attention-GRU) and an improved Transformer (Conv-Transformer). First, passenger flow data at three time scales, namely weekly periodic, daily periodic, and recent time segment, are modeled separately and combined as model inputs. Second, the Attention-GRU and Conv-Transformer blocks are built to mine the continuity and periodicity features respectively and the prediction values are output after feature fusion. Finally, the AFC passenger flow data of two stations of Shanghai Metro Line 2 were collected for the prediction of inbound and outbound passenger flow under the 5-minute time granularity. To analyze the influence of parameters, tuning experiments are carried out and the model is evaluated based on the optimal parameter combination. The results show that compared with five baseline models (BPNN, CNN, GRU, CNN-GRU, Transformer) and four GRU-Transformer ablation models, the GRU-Transformer model has the highest prediction accuracy and good practicability. © 2022 Science Press. All rights reserved.