[Paper Review] Multi-Graph Convolutional-Recurrent Neural Network (MGC-RNN) for Short-Term Forecasting of Transit Passenger Flow

Yuxin He, Lishuai Li+2 more|arXiv (Cornell University)|Jul 28, 2021

Traffic Prediction and Management Techniques4 citations

TL;DR

This paper proposes MGC-RNN, a deep learning model that integrates multi-graph convolutional networks and an LSTM-based seq2seq architecture to forecast short-term transit passenger flows by capturing spatial dependencies, temporal dynamics, and heterogeneous inter-station correlations. Evaluated on Shenzhen Metro data, MGC-RNN outperforms benchmark models, demonstrating superior accuracy and the ability to model complex spatiotemporal patterns through multi-source data fusion.

ABSTRACT

Short-term forecasting of passenger flow is critical for transit management and crowd regulation. Spatial dependencies, temporal dependencies, inter-station correlations driven by other latent factors, and exogenous factors bring challenges to the short-term forecasts of passenger flow of urban rail transit networks. An innovative deep learning approach, Multi-Graph Convolutional-Recurrent Neural Network (MGC-RNN) is proposed to forecast passenger flow in urban rail transit systems to incorporate these complex factors. We propose to use multiple graphs to encode the spatial and other heterogenous inter-station correlations. The temporal dynamics of the inter-station correlations are also modeled via the proposed multi-graph convolutional-recurrent neural network structure. Inflow and outflow of all stations can be collectively predicted with multiple time steps ahead via a sequence to sequence(seq2seq) architecture. The proposed method is applied to the short-term forecasts of passenger flow in Shenzhen Metro, China. The experimental results show that MGC-RNN outperforms the benchmark algorithms in terms of forecasting accuracy. Besides, it is found that the inter-station driven by network distance, network structure, and recent flow patterns are significant factors for passenger flow forecasting. Moreover, the architecture of LSTM-encoder-decoder can capture the temporal dependencies well. In general, the proposed framework could provide multiple views of passenger flow dynamics for fine prediction and exhibit a possibility for multi-source heterogeneous data fusion in the spatiotemporal forecast tasks.

Motivation & Objective

• To address the challenge of short-term passenger flow forecasting in urban rail transit systems, which is complicated by spatial dependencies, temporal dynamics, and heterogeneous inter-station correlations.
• To incorporate exogenous factors such as network distance, structural layout, and recent flow patterns into the forecasting framework.
• To develop a unified deep learning model capable of predicting inflow and outflow at all stations simultaneously across multiple time steps.
• To enable multi-source heterogeneous data fusion for improved spatiotemporal forecasting performance.

Proposed method

• The model employs multiple graphs to encode different types of inter-station correlations, including those based on network distance, topological structure, and recent flow patterns.
• Multi-graph convolutional layers aggregate spatial features from these distinct graph structures to capture heterogeneous spatial dependencies.
• An encoder-decoder architecture with LSTM units models the temporal dynamics of passenger flow across time steps.
• The seq2seq framework enables end-to-end prediction of inflow and outflow for all stations simultaneously over multiple future time steps.
• The model is trained end-to-end using a sequence-to-sequence loss function to minimize prediction error on historical passenger flow sequences.
• Exogenous factors such as station location and flow history are embedded and integrated into the graph construction and feature learning process.

Experimental results

Research questions

• RQ1 How can multiple heterogeneous spatial relationships—such as network distance, topological structure, and recent flow patterns—be effectively modeled in passenger flow forecasting?
• RQ2 To what extent can an LSTM-based encoder-decoder architecture capture long-term temporal dependencies in passenger flow sequences?
• RQ3 Can the integration of multiple graphs improve forecasting accuracy compared to single-graph or non-graph-based models?
• RQ4 How do exogenous factors like station proximity and historical flow influence inter-station correlations and prediction performance?
• RQ5 What is the relative contribution of different types of spatial correlations (e.g., structural vs. flow-based) to forecasting accuracy?

Key findings

• MGC-RNN achieves higher forecasting accuracy than benchmark models, including GCN-LSTM, ST-ResNet, and Transformer-based models, on real-world Shenzhen Metro data.
• The model demonstrates that inter-station correlations driven by network distance, network structure, and recent flow patterns are significant predictors of future passenger flows.
• The LSTM-encoder-decoder architecture effectively captures long-term temporal dependencies in passenger flow sequences.
• The use of multiple graphs significantly enhances performance by modeling diverse spatial relationships beyond simple adjacency.
• The framework enables multi-view analysis of passenger flow dynamics, offering interpretability through distinct graph attention mechanisms.
• The model shows strong potential for multi-source heterogeneous data fusion in spatiotemporal forecasting tasks.