A Comprehensive Spatio-Temporal Model for Subway Passenger Flow Prediction

Zhihao Zhang, Yong Han, Tongxin Peng, Zhenxin Li, Ge ChenISPRS Int. J. Geo Inf 2022 · DOI

TL;DR

This paper introduces GCTN, a novel hybrid deep learning model that significantly improves subway passenger flow prediction by comprehensively modeling both spatial and temporal dependencies. By integrating graph convolutional networks (GCNs) for spatial relationships and a combination of LSTM and Transformer networks for temporal patterns, GCTN achieves superior accuracy and stability, especially for long-term predictions and during peak demand periods.

What Problem Does This Paper Address?

Accurate prediction of subway passenger flow is critical for efficient operation, scheduling, and the development of intelligent transportation systems (ITS). However, existing methods often struggle with the complex spatio-temporal dynamics of subway networks. Many models either focus solely on temporal patterns, neglecting spatial relationships between stations, or fail to capture both local and global temporal dependencies effectively. Furthermore, predicting long-term passenger flow, which is crucial for proactive planning, remains a significant challenge due to error accumulation in traditional autoregressive methods. This paper aims to bridge these gaps by proposing a unified model that addresses these limitations.

Methods

The core of the proposed model is the GCTN (Graph Convolutional and Comprehensive Temporal Neural Network). It's built around a "spatiotemporal block" that integrates several key components:

  • Graph Convolutional Network (GCN): Used to capture the spatial features of the subway network. The authors propose an improved adjacency matrix that incorporates station import/export quantities, aiming to better represent the spatial influence of station characteristics.
  • Convolutional Neural Network (CNN): Applied for initial feature diffusion and to extract intermediate features within the Transformer component.
  • Bidirectional Long Short-Term Memory (Bi-LSTM): Employed to capture local temporal dependencies, considering both past and future time steps.
  • Transformer Network: Utilized to capture global temporal dependencies through its self-attention mechanism. The authors also incorporate a 1x1 convolutional layer within the Transformer to further enhance feature extraction and aggregation.

These components are combined to model both local and global temporal dependencies, and these temporal features are then fused with spatial features. The model is designed for multi-step prediction, directly modeling long-term temporal dependencies to avoid error propagation.

Datasets

The study utilizes Smart Card Data (SCD) from the Shanghai Subway System, China. The dataset spans from April 7th to April 30th, 2015, covering approximately 9 million card records daily across 289 subway stations. The data was processed into time intervals of 10, 20, and 30 minutes for passenger flow counting. The data was split into training, validation, and testing sets with a ratio of 6:2:2.

Key Insights and Ideas

  • Enhanced Spatial Representation: The improved adjacency matrix, incorporating station import/export quantities, significantly boosts the model's ability to capture spatial influences, leading to better prediction accuracy.
  • Comprehensive Temporal Modeling: Combining Bi-LSTM for local and Transformer for global temporal dependencies allows GCTN to capture a richer set of temporal patterns than single-model approaches.
  • Direct Long-Term Prediction: By directly modeling long-term temporal dependencies rather than relying on step-by-step autoregression, GCTN mitigates error accumulation and provides more stable long-term forecasts.
  • Importance of Spatio-Temporal Fusion: The effectiveness of fusing spatial features (from GCN) with comprehensive temporal features (from Bi-LSTM and Transformer) is crucial for accurate overall prediction.
  • Impact of Station Surroundings: The study highlights that station import/export quantities and nearby Points of Interest (POI) are positively correlated with prediction error (RMSE), suggesting these factors are important for spatial modeling.

Key Results

The GCTN model demonstrated superior performance across various time intervals (10, 20, 30, 60 minutes) compared to several baseline models, including statistical methods (LR), machine learning methods (KNN, RSVR, LSVR), and other deep learning models (Bi-LSTM, Transformer, CNN, STGCN, STTN).

  • Overall Accuracy: GCTN achieved the lowest RMSE, MAE, and WMAPE in most scenarios, particularly for longer time intervals. For example, in 20-minute intervals, GCTN achieved an RMSE of 50.79/55.78/59.38 for 20/40/60 min predictions, outperforming other models.
  • Peak Prediction: The model showed a strong ability to predict peak passenger flow values, which is vital for operational decision-making.
  • Stability and Robustness: GCTN exhibited better prediction stability and robustness, especially during periods of rapid passenger flow changes.
  • Temporal Pattern Fusion: Combining close, daily, and weekly temporal patterns (1HD1W) yielded the best prediction performance, indicating the value of incorporating diverse temporal information.
  • Ablation Studies: Removing components like the LSTM or Transformer, or using an un-improved adjacency matrix, significantly degraded performance, underscoring the importance of each element in GCTN.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Data Period: The validation dataset's duration was insufficient to study the influence of long-term temporal factors like seasons.
  • External Factors Ignored: The model does not currently incorporate external factors such as weather, which can influence passenger flow.
  • Static Spatial Characteristics: The study focuses on static spatial characteristics; dynamic spatial features are not considered.

Future work suggested includes:

  • Exploring the influence of external features and applying GCTN to longer datasets.
  • Investigating dynamic spatiotemporal features within the deep learning model.
  • Studying the differences between various attention mechanisms for global temporal features.

Summary updated

BibTeX
@article{zhang2022comprehensive,
  author = {Zhihao Zhang and Yong Han and Tongxin Peng and Zhenxin Li and Ge Chen},
  title = {{A Comprehensive Spatio-Temporal Model for Subway Passenger Flow Prediction}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2022},
  doi = {10.3390/ijgi11060341},
}
Figure