Technical Summary: Context-Enriched Graph Neural Networks for EV Charging Station Energy Demand Forecasting

1. Problem Statement

The rapid proliferation of electric vehicles (EVs) has intensified the demand for charging infrastructure, placing significant stress on power grids and creating challenges for grid stability and sustainable urban mobility. Accurate energy demand forecasting is critical for Charging Station Operators (CSOs) to balance supply and demand, manage grid load, and plan infrastructure expansion.

However, existing forecasting methods face two primary limitations:

1. Neglect of Spatial Dependencies: Many approaches treat charging stations as isolated entities, failing to capture the spatial interdependencies where the demand at one station influences or is influenced by neighboring stations and the surrounding urban environment.
2. Insufficient Exploitation of Heterogeneous Data: Current models often underutilize contextual urban data, such as Points-of-Interest (POI) distributions and population density, which are crucial drivers of charging behavior.

Consequently, there is a need for a framework that jointly models temporal dynamics and spatial correlations while integrating diverse contextual features to improve prediction accuracy and generalizability across different geographic settings.

2. Methodology

The authors propose a Spatio-Temporal Graph Neural Network (STGNN) framework designed for day-ahead energy demand forecasting at the station level. The methodology integrates heterogeneous data sources into a static spatial graph to model complex energy patterns.

2.1. Graph Construction and Representation

• Nodes: Each node represents a specific Electric Vehicle Charging Station (EVCS).
• Edges: A static spatial graph is constructed where edges encode relationships between stations. These relationships are defined by:
  • Physical Proximity: Derived from road network connectivity (shortest path travel time/distance).
  • Functional Proximity: Reflecting correlations in demand between stations serving similar urban roles, even if not geographically adjacent.
• Node Features: Each node is enriched with a vector of heterogeneous features, including:
  • Historical Load Profiles: Lagged energy consumption (1, 3, 7 days) and rolling statistics (mean, standard deviation).
  • Contextual Urban Data: POI distributions (e.g., parking, residential, office, road types) extracted from OpenStreetMap and population density.
  • Temporal Indicators: Day of the week, month, holiday indicators, and peak usage flags.
  • Session Data: Lagged session counts and durations.

2.2. Forecasting Framework

The framework employs a sliding-window mechanism to process temporal sequences.

• Input: A sequence of historical observations ($w$ time steps) for all $N$ nodes, represented as $X_{t-w+1:t} \in \mathbb{R}^{w \times N \times F}$.
• Processing: The model utilizes message-passing mechanisms to propagate information between connected stations, allowing each station to iteratively exchange information with its neighbors. This captures how local demand fluctuations influence surrounding nodes.
• Output: The model predicts energy consumption for the next time step (day-ahead) for all stations simultaneously.

2.3. Model Architectures

The study systematically benchmarks four advanced STGNN architectures against temporal-only baselines (LSTM and GRU):

1. MTGNN (Multivariate Time Graph Neural Network): Integrates temporal convolutions with adaptive graph structure learning to handle multivariate time-series data.
2. ASTGCN (Attention-Based Spatio-Temporal Graph Convolutional Network): Applies spatial and temporal attention mechanisms to capture dynamic dependencies.
3. STConv (Spatio-Temporal Convolutional Network): Efficiently models local temporal patterns using graph convolutions.
4. A3TGCN (Attention Temporal Graph Convolutional Network): Combines gated temporal units with attention mechanisms for sequence learning.

2.4. Data Preprocessing

• Datasets: Two geographically diverse datasets were used:
  • Palo Alto, USA: An urban dataset (46 stations, 2018–2020).
  • Yonne, France: A semi-rural dataset (133 stations, 2021–2022).
• Noise Reduction: Exponential Moving Average (EMA) was applied to smooth noisy data and reduce the impact of transient anomalies while preserving underlying trends.
• Data Cleaning: Outliers were corrected, and data from the COVID-19 pandemic period (characterized by atypical usage) was excluded to ensure training on regular operational patterns.

3. Key Contributions

The paper outlines three primary contributions:

1. A Novel Forecasting Framework: The proposal of a station-level, day-ahead forecasting framework using advanced STGNN architectures (MTGNN, ASTGCN, STConv, A3TGCN) that explicitly models both temporal evolution and spatial interdependencies.
2. Context-Enriched Data Fusion: The integration of multi-source data, including spatial relationships, historical usage, and contextual urban features (POIs, population density), into node feature vectors. This approach captures demand fluctuations shaped by spatial connectivity, user behavior, and demographic factors.
3. Comprehensive Evaluation: A rigorous evaluation across two contrasting real-world datasets (urban vs. semi-rural) comparing STGNNs against established temporal baselines (LSTM, GRU) using multiple metrics (RMSE, MAE, MSE, $R^2$).

4. Experimental Results

Experiments were conducted on the Palo Alto and Yonne datasets with a 60/20/20 train/validation/test split.

4.1. Performance Metrics

• Palo Alto Dataset (Urban):
  • Best Model: MTGNN achieved the lowest test RMSE (8.32) and highest $R^2$ (0.86), outperforming LSTM (RMSE 10.65) and GRU (RMSE 11.31).
  • Efficiency: MTGNN demonstrated high computational efficiency with an execution time of 0.12 minutes.
• Yonne Dataset (Semi-Rural):
  • Best Model: ASTGCN and MTGNN tied for the highest accuracy, both achieving an $R^2$ of 0.86. ASTGCN achieved a slightly lower test RMSE (15.16) compared to MTGNN (15.54).
  • Baseline Comparison: Both STGNNs significantly outperformed LSTM (RMSE 18.24) and GRU (RMSE 20.36).
  • Execution Time: MTGNN (1.39 min) and ASTGCN (2.14 min) were significantly faster than the baselines (LSTM: 73.03 min; GRU: 109.39 min).

4.2. Analysis

• Spatial Dependency: The superior performance of graph-based models over LSTM and GRU confirms that capturing spatial correlations between stations is critical for accurate forecasting.
• Model Specifics:
  • MTGNN proved robust across both datasets, particularly effective in handling multivariate time-series data and offering the best balance of accuracy and speed in the urban setting.
  • ASTGCN excelled in the semi-rural setting, leveraging its attention mechanisms to model complex spatial-temporal dependencies.
  • STConv showed competitive accuracy but suffered from high computational overhead (e.g., ~58 minutes for Palo Alto).
  • A3TGCN offered the fastest execution time but struggled with generalization in high-variability scenarios, resulting in lower $R^2$ scores.

5. Significance and Claims

The paper claims that graph-based spatio-temporal modeling enables accurate, efficient, and generalizable EVCS forecasting across diverse urban settings. By integrating heterogeneous urban data (POIs, population) with historical load profiles, the proposed framework addresses the limitations of existing methods that neglect spatial dependencies.

The study emphasizes reproducibility and adaptability across heterogeneous urban contexts with reduced data requirements, making it practically applicable for infrastructure planning and grid coordination. The results demonstrate that STGNNs, particularly MTGNN and ASTGCN, provide a robust foundation for managing the energy demands of electrified transportation, offering a significant improvement over traditional temporal-only models.

Future Directions: The authors suggest future work should focus on:

• Integrating real-time traffic information.
• Refining optimization using reinforcement learning.
• Applying transfer learning to adapt models from data-rich regions to data-scarce contexts (e.g., emerging markets).
• Utilizing meta-learning for rapid model adaptation to new stations or evolving demand conditions.