Introduction
Accurate forecasting of metro passenger flow is a critical component of intelligent urban transportation systems. Unlike road traffic, where vehicles can choose various routes, metro passengers are constrained by fixed tracks and rigid schedules. However, their behavior remains highly dynamic; the spatial correlations between stations shift significantly throughout the day—for instance, morning rush hours connect residential areas to business districts, while weekend afternoons link various commercial hubs.
Existing Graph Neural Network (GNN) approaches often fall short because they rely on static or slowly evolving graph structures. Furthermore, these models frequently struggle with environmental noise—such as weather or calendar data—that can obscure the underlying flow dynamics. To address these challenges, this paper introduces ReDyNet (Responsive Dynamic Graph Neural Network).
Methodology: The ReDyNet Framework
ReDyNet is designed to move beyond rigid spatial representations by creating time-sensitive "virtual maps" of city-wide station interactions. The architecture consists of four integrated modules designed to filter noise and adapt to real-time shifts in passenger behavior.
Core Components
- Contextual Spatial Embedding (CSE): This module integrates historical flow data with external factors like date and weather into a unified latent representation.
- Redundancy Context Disentanglement (RCD): Utilizing Information Bottleneck (IB) principles, RCD acts as a filter. It explicitly separates essential signals required for prediction from redundant environmental noise. By minimizing mutual information between raw inputs and learned representations while maximizing task-relevant information, the model avoids simply "memorizing" the calendar.
- Responsive Dynamic Learning (RDL): Instead of static weights, RDL employs a parameter pool. It dynamically selects and combines weights based on the current flow state, generating a "responsive" adjacency matrix for every time step. This allows the graph structure to adapt to changing passenger patterns.
- Task Adaption: A final gated mechanism modulates the output to handle non-linearities before producing the final forecast.
Experimental Results and Analysis
ReDyNet was evaluated on three large-scale, real-world datasets from Beijing, Shanghai, and Hangzhou. The model consistently outperformed state-of-the-art baselines, including PDFormer and STGCN, across all time horizons (15, 30, 45, and 60 minutes).
Performance and Ablation Study
The effectiveness of ReDyNet's individual components is validated through extensive ablation studies. As shown in the performance metrics, removing the RCD or RDL (Dynamic Learning) components leads to a noticeable increase in error (MAE and RMSE).
Key findings from the evaluation include:
- Robustness to Horizon Drift: While traditional models like DCRNN show significant performance degradation as the prediction horizon increases, ReDyNet maintains high accuracy even at the 60-minute mark.
- Disentanglement Efficacy: Visual analysis (via t-SNE) confirms that the RCD module successfully produces date-independent representations, focusing the model on actual flow dynamics rather than raw temporal signals.
- Capturing Delayed Correlations: The model successfully identifies "chain reactions," such as how an inflow surge at a residential station predictably leads to an outflow surge at a workplace station later in the hour.
Conclusion
ReDyNet represents a significant advancement in metro flow forecasting by combining information theory with dynamic graph learning. By filtering out redundant contextual noise and allowing the graph structure to "breathe" in response to real-time data, it achieves superior accuracy and robustness. Future work will focus on integrating more granular environmental data and accounting for unplanned system shocks like accidents or major public events.
Summary updated