💻 computer science

A Metro Passenger Flow Prediction Method Based on Adaptive Graph Convolution and Hybrid Attention

This paper proposes AGHA-Net, a novel deep learning framework that integrates adaptive graph convolution, spatiotemporal gated recurrent units, and a dual-branch hybrid attention mechanism to significantly improve the accuracy and stability of short-term urban rail transit passenger flow prediction by effectively capturing dynamic spatial dependencies and complex temporal patterns.

Original authors: lu zeng, qingshuo sha, tao gao, qinzhen fang

Published 2026-08-03

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: lu zeng, qingshuo sha, tao gao, qinzhen fang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to predict the weather, but instead of clouds and wind, you're tracking thousands of people rushing through a subway system. This is the world of urban rail transit prediction, a field where scientists try to guess how many passengers will enter a train station in the next few minutes. Why does this matter? Because if you know exactly when a crowd is coming, you can open more ticket gates, send extra trains, and keep everyone from getting squished or stuck. To do this, computers need to understand two tricky things at once: space (how stations are connected like a web) and time (how crowds ebb and flow like tides). Old computer programs were like rigid robots; they could only follow simple, straight-line rules and couldn't handle the messy, changing reality of a busy city. They struggled to see that a station might be quiet in the morning but chaotic at lunch, or that a delay at one stop could ripple out to affect a station miles away.

Enter a new team of researchers who built a smarter, more flexible digital brain called AGHA-Net. Think of this new method as a super-observant conductor who doesn't just look at a static map, but actually watches the crowd move in real-time. Instead of assuming the connections between subway stations are fixed, AGHA-Net learns to redraw the map on the fly, noticing which stations are suddenly "talking" to each other because of a rush hour surge. It combines this with a special attention system that acts like a human memory, remembering not just what happened five minutes ago, but also spotting patterns from hours or days before. The researchers tested this on real data from the Hangzhou Metro in China. They found that AGHA-Net is significantly better at predicting passenger numbers than older methods, making fewer mistakes whether it's guessing for 15 minutes or 45 minutes ahead. It's like upgrading from a crystal ball that only works on sunny days to a high-tech weather radar that sees through the storm.

The Problem with Old Maps

Imagine you are trying to navigate a city using a map that was drawn ten years ago and never updated. If a new bridge opens or a road closes, your map is useless. This is what happened with many old subway prediction tools. They relied on a predefined map (called an adjacency matrix) that said, "Station A is connected to Station B," and stuck with that forever. But in the real world, the "connection" between stations changes constantly. During morning rush hour, a station near an office park might be the most important hub, while at night, a station near a shopping mall takes over. Old models couldn't adapt to this; they were stuck with their static map, missing the dynamic dance of the crowd.

Furthermore, these old models were bad at remembering the past. They could look at the last few minutes of data but struggled to understand long-term trends, like the difference between a Tuesday commute and a Saturday night party. They often made mistakes that piled up the further into the future they tried to predict.

The New Solution: AGHA-Net

The authors propose AGHA-Net, a method that fixes these issues by giving the computer three superpowers:

1. The Shape-Shifting Map (Adaptive Graph Convolution)
Instead of using a fixed map, AGHA-Net uses a learnable node embedding. Imagine if every subway station had a secret ID card that changed its shape depending on the time of day. The computer looks at these ID cards and figures out which stations are "friends" right now. If the morning rush makes Station A and Station B act like they are right next to each other, the computer draws a new line between them. This is called Adaptive Graph Convolution (AdaGCN). It allows the model to capture the real, changing relationships between stations, rather than just the ones on a paper map.

2. The Memory Gears (Spatiotemporal Gated Unit)
Once the map is updated, the model needs to process the flow of people. The researchers built a special engine called ST-GRU (Spatiotemporal Gated Recurrent Unit). Think of this as a gear system that mixes two things at once: the where (spatial) and the when (temporal). In older models, the computer would look at the map, then look at the clock, then try to guess. AGHA-Net does both simultaneously. It updates its memory of the crowd's location and movement in the same heartbeat, allowing it to catch sudden surges or drops in traffic much faster.

3. The Double-Lens Attention (Hybrid Attention)
Finally, the model needs to look at the big picture. The researchers added a Hybrid Attention layer with two different lenses working in parallel:

Sequential Self-Attention: This is like a wide-angle lens that scans the entire timeline to find long-range patterns. It asks, "What happened last week at this time?" to help predict today.
Bilinear Temporal Attention: This is a special, efficient lens that compresses information. Instead of trying to look at every single second (which would be too slow), it squishes the data into a manageable shape to find hidden trends. It's like a detective who doesn't read every word of a diary but quickly spots the recurring themes.

What the Experiments Showed

The team tested AGHA-Net using real passenger data from 80 stations on three metro lines in Hangzhou, covering data from January 1 to January 25, 2019. They looked at different time intervals: 15 minutes, 30 minutes, and 45 minutes.

The results were clear: AGHA-Net was the best at the game.

For the 15-minute prediction, it achieved an error rate (MAE) of 21.65, compared to 24.01 for the next best model.
For 30 minutes, the error was 22.37, beating the competition.
Even for the tricky 45-minute prediction, it stayed accurate with an error of 22.96.

The researchers also checked how well it worked on different types of stations:

Commuter Stations (Station A): Places with heavy morning and evening peaks. AGHA-Net nailed the "tidal" patterns of people rushing to work.
Commercial Stations (Station B): Busy all day with high volumes. The model handled the constant chaos without getting confused.
Residential Stations (Station C): Quieter during the day with an evening spike. It accurately predicted the calm before the evening rush.

Why It Matters

To prove that every part of their new machine was necessary, the researchers ran "ablation studies"—basically, they took parts of the model out one by one to see what broke.

When they removed the Adaptive Graph Convolution (the shape-shifting map), the error jumped by over 35%. This proved that a fixed map just isn't good enough.
When they removed the Bilinear Temporal Attention (the efficient lens), the model started making more mistakes as they tried to predict further into the future. This showed that this part is crucial for stopping errors from piling up over time.

The paper concludes that by combining a flexible map, a memory that updates instantly, and a dual-lens attention system, AGHA-Net provides a much more reliable way to predict subway crowds. This isn't just a math trick; it means city planners can make better decisions, send trains when they are actually needed, and keep the subway running smoothly for everyone. While the model is currently very good at handling the data it was given, the authors suggest that future versions could also look at outside factors like weather or holidays to get even smarter.

Technical Summary: AGHA-Net for Metro Passenger Flow Prediction

Problem Statement

Accurate short-term prediction of inbound passenger flow in urban rail transit is critical for network planning, smart services, and refined operations. However, existing methods face three primary limitations:

Static Spatial Modeling: Most methods rely on predefined adjacency matrices with fixed weights, failing to capture the dynamically evolving, nonlinear spatial dependencies between stations (e.g., during peak hours or transfer coordination).
Limited Temporal Modeling: Temporal modeling often depends on single recurrent or convolutional structures with limited mechanisms for capturing long-range dependencies, multi-scale periodic patterns, and long-term trends, leading to error accumulation in long-term predictions.
Insufficient Spatiotemporal Fusion: Spatial and temporal modules are often simply connected in series, resulting in insufficient interaction and low information fusion efficiency, which limits model stability and generalization.

Methodology: AGHA-Net

The authors propose AGHA-Net (Adaptive Graph Convolution and Hybrid Attention Network), a framework designed to jointly model dynamic spatial dependencies, local spatiotemporal features, and global long-range temporal trends. The architecture consists of three cascaded components:

1. Spatiotemporal Gated Recursive Layer (ST-GRU)

This layer integrates spatial aggregation and temporal state updating into a single computational graph to capture local spatiotemporal features.

Adaptive Graph Convolution (AdaGCN): Instead of using a fixed adjacency matrix, AGHA-Net employs learnable node embeddings ( $E_{adj}$ ) to generate an adaptive adjacency matrix ( $\tilde{A}$ ) via inner products and softmax normalization. This allows the model to learn time-varying spatial correlations directly from the data while retaining static topological priors.
Spatiotemporal Gated Unit: The adaptive graph convolution is embedded directly into the Gated Recurrent Unit (GRU) gating mechanism (update and reset gates). This enables the simultaneous propagation of spatial information and temporal state updates, effectively modeling coupled patterns like peak aggregation and transfer coordination.

2. Hybrid Attention Layer

To address the limitations of the ST-GRU in capturing long-range dependencies, a dual-branch parallel attention mechanism is introduced:

Sequential Self-Attention (SSAM): This branch incorporates temporal positional encoding and one-dimensional convolution into the query and key matrices. It captures global long-range temporal dependencies while preserving the integrity of the temporal structure.
Bilinear Temporal Attention (TABL): Designed to reduce computational overhead and explicitly model multi-scale periodic patterns, this module replaces full self-projection with dual-dimensional compression. It rearranges the input tensor to compress feature and node dimensions via convolution, then constructs a temporal attention matrix using bilinear interaction. This efficiently models internal correlations within long sequences and long-term trends.

3. Feature Processing Layer

This layer addresses potential spatial representation inconsistencies and attention weight biases.

Spatial Alignment: It uses a predefined adjacency matrix as a spatial inductive bias to inject physical topology priors, guiding the model to prioritize real interactions between adjacent stations and suppress spurious cross-region correlations.
Prediction Mapping: A one-dimensional convolution maps the refined features to the prediction space, outputting the passenger flow for the next time step.

Key Contributions

The paper highlights four main contributions:

Adaptive Graph Convolution: A strategy that separates static topological priors from data-driven associations, solving the reliance on predefined adjacency matrices and enabling the adaptive learning of nonlinear spatial dependencies.
ST-GRU Unit: An innovative spatiotemporal gated unit that embeds graph convolution into the GRU structure, breaking the separation between spatial and temporal modules to enhance real-time perception of coupled patterns.
Hybrid Attention Mechanism: A dual-branch design combining sequential self-attention for global dependencies and bilinear temporal attention for efficient long-term trend modeling via dual-dimensional compression.
Empirical Validation: Comprehensive experiments on real-world Hangzhou Metro data demonstrating the model's superiority over baselines and its ability to handle diverse station types (commuter, commercial, residential).

Experimental Results

The model was evaluated on a dataset of inbound passenger flow from 80 stations across three Hangzhou Metro lines (January 1–25, 2019). Experiments were conducted at 15-minute, 30-minute, and 45-minute prediction granularities.

Performance Comparison: AGHA-Net significantly outperformed 12 baseline models, including traditional statistical methods (HA, ARIMA), classical deep learning (LSTM, GRU, GCN), and advanced spatiotemporal graph networks (ASTGCN, T-GCN, PDFormer, etc.).
- At the 15-minute granularity, AGHA-Net achieved an MAE of 21.65, RMSE of 34.68, and MAPE of 21.83%, surpassing the next best model (ST-ConvLSTMNet) by a notable margin.
- The model demonstrated superior stability in long-term prediction, with the MAE growth from 15 to 45 minutes (6.1%) being far lower than that of ASTGCN (27.2%), indicating effective suppression of error accumulation.
Station Adaptability: The model maintained high accuracy across diverse station types:
- Commuter Stations: Successfully captured morning/evening double-peak patterns.
- Commercial Stations: Robustly tracked high-volume, complex fluctuations.
- Residential Stations: Accurately modeled single-evening peak patterns and low-volume periods.
Ablation Studies: Removing any core component (AdaGCN, ST-GRU, SSAM, or TABL) resulted in performance degradation.
- Removing AdaGCN caused the most significant drop (MAE increase of 6.82), confirming the necessity of adaptive spatial modeling.
- Removing TABL led to accelerated error accumulation in longer prediction steps, validating its role in long-term trend modeling.
Module Effectiveness: A specific comparison between the proposed Bilinear Temporal Attention (TABL) and a traditional temporal attention module (STAGCN) showed that TABL reduced MAE by 2.39 and RMSE by 4.33 at the 30-minute mark, proving its effectiveness in suppressing long-term error accumulation.

Significance and Claims

The paper claims that AGHA-Net provides a reliable decision support tool for intelligent rail transit management, passenger flow control, and capacity optimization. By unifying dynamic spatial dependency modeling, local spatiotemporal feature extraction, and long-term trend capture within a single framework, the method addresses the fragmentation often found in existing spatiotemporal models.

The authors emphasize that the method's ability to adaptively learn spatial correlations without predefined matrices, combined with its efficient handling of long-range temporal dependencies, offers a significant improvement in prediction accuracy and stability. The study concludes that the proposed approach is well-suited for the complex, nonlinear, and dynamic nature of urban rail transit passenger flow.

Note: The authors acknowledge that future work will incorporate multimodal data (weather, holidays, special events) to further enhance robustness and explore causal interpretability and cross-domain transfer.