Abstract:

A metro transfer passenger flow prediction model was proposed based on the seasonal decomposition of time series by loess（STL）and Gated Recurrent Unit（GRU），in order to enrich the research on metro internal transfer passenger flow prediction and to better formulate the metro operation plan.The prediction process was divided into three stages by the model.In the first stage，the raw automatic fare collection（AFC）data are preprocessed，where the travel path of passengers is identified using the graph-based depth-first search algorithm and the transfer passenger flow time series are constructed.In the second stage，the transfer passenger flow time series are decomposed into the trend component，seasonal component and remainder component by the STL；while the outliers of remainder component are eliminated and filled using the 3σ principle.In the third stage，the GRU model is built and the related training and prediction are processed through the deep learning library Keras.The model performance was validated with the passenger flow data of Xizhimen Station of Beijing metro.The result shows that，compared to the following 3 models which are long short-term memory neural network（LSTM），GRU and STL-LSTM model，the STL-GRU prediction model can improve the prediction accuracy of transfer passenger flow on weekdays（excluding Friday），Friday and weekends，and the mean absolute percentage errors of the prediction results can be reduced by at least 2.3%，1.36%，and 6.42%，respectively.

Key words: urban traffic, transfer passenger flow prediction, GRU, metro, STL, deep learning, prediction accuracy