[Paper Review] AST-GIN: Attribute-Augmented Spatial-Temporal Graph Informer Network for Electric Vehicle Charging Station Availability Forecasting

Ruikang Luo, Yaofeng Song+3 more|arXiv (Cornell University)|Sep 7, 2022

Transportation and Mobility Innovations4 citations

TL;DR

This paper proposes AST-GIN, a novel deep learning model that integrates graph convolutional networks and Informer architectures with attribute-augmented encoding to forecast electric vehicle (EV) charging station availability by modeling both spatial-temporal dependencies and dynamic/static external factors. Evaluated on real-world data from Dundee City, AST-GIN achieves state-of-the-art performance, outperforming baselines by up to 11.54% in accuracy on 90-minute forecasts and 8.97% on 120-minute forecasts.

ABSTRACT

Electric Vehicle (EV) charging demand and charging station availability forecasting is one of the challenges in the intelligent transportation system. With the accurate EV station situation prediction, suitable charging behaviors could be scheduled in advance to relieve range anxiety. Many existing deep learning methods are proposed to address this issue, however, due to the complex road network structure and comprehensive external factors, such as point of interests (POIs) and weather effects, many commonly used algorithms could just extract the historical usage information without considering comprehensive influence of external factors. To enhance the prediction accuracy and interpretability, the Attribute-Augmented Spatial-Temporal Graph Informer (AST-GIN) structure is proposed in this study by combining the Graph Convolutional Network (GCN) layer and the Informer layer to extract both external and internal spatial-temporal dependence of relevant transportation data. And the external factors are modeled as dynamic attributes by the attribute-augmented encoder for training. AST-GIN model is tested on the data collected in Dundee City and experimental results show the effectiveness of our model considering external factors influence over various horizon settings compared with other baselines.

Motivation & Objective

• Address the challenge of accurately forecasting EV charging station availability amid complex spatial-temporal dynamics and external influences.
• Overcome limitations of existing models that fail to integrate comprehensive external factors such as weather, POIs, and time-of-day patterns.
• Develop a unified deep learning framework capable of modeling both static (e.g., POI types) and dynamic (e.g., weather, time) attributes alongside spatial and temporal dependencies.
• Improve prediction robustness and interpretability by explicitly encoding external attributes into the model’s attention and message-passing mechanisms.
• Validate the model’s effectiveness across multiple forecasting horizons using real-world data from Dundee City.

Proposed method

• Propose an Attribute-Augmented Spatial-Temporal Graph Informer Network (AST-GIN) that combines GCN for spatial dependency modeling and Informer for long-range temporal attention.
• Introduce an Attribute Augmentation Unit (A2Unit) to dynamically encode external factors—such as weather, time-of-day, and POI types—into node-level embeddings.
• Use graph convolutional layers to capture topological relationships between charging stations, modeling them as nodes in a spatial graph.
• Leverage the Informer’s multi-head self-attention mechanism with a factorized attention design to model long-range temporal dependencies efficiently.
• Train the model end-to-end using historical usage data, external attributes, and spatial graph structure to predict future availability.
• Apply normalization and residual connections to stabilize training and improve generalization across different forecasting horizons.

Experimental results

Research questions

• RQ1 Can a unified deep learning model effectively capture both spatial and temporal dependencies in EV charging station usage data?
• RQ2 To what extent do dynamic and static external factors—such as weather and POI types—improve forecasting accuracy?
• RQ3 How does the integration of attribute-augmented encoding enhance model performance compared to models that ignore external factors?
• RQ4 Does AST-GIN maintain superior performance across short-, medium-, and long-term forecasting horizons?
• RQ5 How robust is AST-GIN to data noise and perturbations in real-world sensor data?

Key findings

• AST-GIN achieves the highest accuracy across all forecasting horizons, with a 0.8388 accuracy at the 90-minute horizon, outperforming the best baseline (STTN) by 11.54%.
• At the 120-minute horizon, AST-GIN achieves 0.7507 accuracy, which is 8.97% higher than the best baseline model (STTN).
• The combination of static and dynamic external factors yields the best performance, especially in short-term forecasting (e.g., 30–90 minutes), where accuracy improvements are most significant.
• In long-term forecasting (120 minutes), the performance gap between combined and individual factor types diminishes, indicating that temporal patterns dominate over external attributes.
• Perturbation analysis confirms AST-GIN’s robustness, as performance remains stable under Gaussian noise injection, indicating strong noise immunity.
• Visualization results show that AST-GIN consistently tracks ground truth values across different horizons, demonstrating stable and accurate temporal forecasting capability.