A Hybrid GLM Model for Predicting Citywide Spatio-Temporal Metro Passenger Flow

Yong Han, Tongxin Peng, Cheng Wang, Zhihao Zhang, Ge ChenISPRS Int. J. Geo Inf 2021 · DOI

TL;DR

This paper introduces a novel hybrid deep learning model (hybrid GLM) that combines Graph Attention Networks (GAT) and Long Short-Term Memory (LSTM) networks to predict citywide metro passenger flow. By effectively capturing both spatial dependencies within the metro network and temporal patterns, along with external factors, the hybrid GLM achieves significant improvements in prediction accuracy compared to existing state-of-the-art methods.

What Problem Does This Paper Address?

Accurately predicting metro passenger flow is crucial for urban management, transport scheduling, and enhancing the efficiency of intelligent transportation systems. While deep learning models have shown promise, predicting passenger flow remains challenging due to the complex spatial relationships within metro networks (which are often asymmetric) and dynamic temporal traffic patterns. Existing graph-based methods often struggle with asymmetric networks or fail to consider nodes beyond immediate neighbors. Furthermore, many models neglect the impact of external factors like weather and holidays, which can significantly influence passenger behavior. This paper aims to address these limitations by proposing a model that can handle asymmetric networks, incorporate broader spatial dependencies, and integrate external influences for more robust predictions.

Methods

The core of this paper is the proposed hybrid GLM (hybrid GAT and LSTM Model). This model is built upon several key components:

  • Graph Attention Networks (GAT): Unlike Graph Convolutional Networks (GCN) which require symmetric graph structures, GATs are employed to handle the asymmetric nature of real-world metro networks. GATs use an attention mechanism to assign different importance to neighboring nodes, allowing for a more nuanced capture of spatial dependencies.
  • Improved Adjacent Matrix: To account for the influence of nodes beyond immediate neighbors, the authors propose an improved adjacency matrix. This matrix assigns different weights to nodes based on their distance, allowing the model to consider "further neighbors."
  • Long Short-Term Memory (LSTM) Networks: LSTMs are used to model the temporal dependencies in passenger flow data. The model incorporates LSTMs to capture both short-term (close pattern) and longer-term (daily and weekly patterns) temporal dynamics.
  • External Factor Integration: The model explicitly incorporates external factors such as weather conditions (temperature, rain) and air quality (AQI, PM2.5, etc.) into a dedicated LSTM branch. This allows the model to account for how these external influences affect passenger flow.
  • Multi-Branch Architecture: The model is structured into five branches: three for spatial dependency (close, daily, weekly patterns) using GAT, one for temporal dependency using LSTM, and one for external factors using LSTM. These branches are then fused to generate the final prediction.
  • Feature Fusion: A weighted fusion mechanism is used to combine the outputs from the different branches, allowing the model to learn the relative importance of each component.

Datasets

The study utilizes Smart Card Data (SCD) collected from the metro system of Shanghai, China. This dataset covers approximately nine million card records per day for 289 metro stations over the period of April 1st to April 30th, 2015. The data includes transaction details like card number, date, time, metro line, business type, price, and discount information. External factors such as weather and air quality data were also collected.

Key Insights and Ideas

  • Asymmetric Networks are the Norm: The paper highlights that real-world metro networks are often asymmetric, making traditional GCNs unsuitable. GAT's ability to handle this is a significant advantage.
  • Beyond Immediate Neighbors: The improvement to the adjacency matrix, considering nodes at varying distances, is crucial for capturing the broader spatial influence in a city-wide network.
  • External Factors Matter: Explicitly incorporating weather and air quality data significantly contributes to prediction accuracy, demonstrating that human mobility is not solely driven by network structure and past flow.
  • Rush Hours are Key: The model shows a particular strength in predicting passenger flow during peak rush hours, which is highly valuable for operational management.
  • Temporal vs. Spatial Trade-offs: The results suggest that while temporal dependencies are important, capturing complex spatial dependencies (especially in asymmetric networks) can be more challenging but ultimately leads to better overall performance when combined effectively.

Key Results

The hybrid GLM model demonstrated superior performance compared to 10 baseline models, including statistical methods (LR), machine learning methods (KNN, RSVR), and various deep learning models (LSTM, CNN, ResNet, GAT, STGCN).

  • Overall Accuracy: The hybrid GLM achieved the smallest Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) across different time intervals (TIs).
  • Significant Improvements: Compared to the best baseline (STGCN), the hybrid GLM achieved a 6-10% reduction in RMSE for different TIs.
  • Rush Hour Performance: In rush hours on working days, the hybrid GLM showed a significant improvement, with an 8.62% reduction in RMSE for inflow and a 10.06% reduction for outflow compared to STGCN.
  • TI Robustness: The model maintained high accuracy across different time intervals (10, 15, 20, and 30 minutes), indicating robustness. For example, at a 30-minute TI, it achieved a 43.27% RMSE reduction compared to CNN.

Limitations and Future Work

The authors acknowledge several limitations:

  • Non-Rush Hour Prediction: The prediction errors for the hybrid GLM were slightly larger than STGCN during non-rush hours on working days. This is attributed to lower passenger flow, potentially leading to less discernible patterns in graph structures.
  • Limited Temporal Scope: The validation dataset covered only one month, which might not capture seasonal variations or long-term temporal external factors.
  • Single-Step Prediction: The study employed single-step ahead prediction, which can lead to error accumulation.

Future work directions include:

  • Exploring External Features: Investigating a wider range of external factors and their correlations for improved model performance.
  • Multi-Step Prediction: Implementing multi-step ahead prediction to mitigate error accumulation.
  • Longer Datasets and Other Flows: Applying the model to longer historical datasets and exploring its applicability to other types of flows (e.g., bike flow, crowd flow, traffic flow).
  • Advanced GAT Capabilities: Further exploring GAT's capabilities in modeling dynamic graphs, multi-graphs, and bidirectional traffic line characteristics.

Summary updated

BibTeX
@article{han2021hybrid,
  author = {Yong Han and Tongxin Peng and Cheng Wang and Zhihao Zhang and Ge Chen},
  title = {{A Hybrid GLM Model for Predicting Citywide Spatio-Temporal Metro Passenger Flow}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2021},
  doi = {10.3390/ijgi10040222},
}
Figure