STGATE: An Explainable Spatio-Temporal Graph Attention Network for Multi-Task EV Charging Analytics

STGATE: An Explainable Spatio-Temporal Graph Attention Network for Multi-Task EV Charging Analytics

Sakshi Limaye, Shravani Nagawade, Jay Paspule

Ms. Divya Mary Biji
Assistant Professor, Department of Computer Science
Ajeenkya DY Patil University
Charholi bk, Pune, Maharashtra, India

Abstract- The rapid rise of electric vehicles (EVs) has created a growing need for smarter charging infrastructure. Most existing
methods treat demand forecasting, user behavior, and grid monitoring as separate problems, even though they are closely connected
in real-world usage. In this work, we propose STGATE (Spatio-Temporal Graph Attention Network with Transformer Encoder), a
unified framework that addresses these challenges together. The model represents charging stations as nodes in a graph and uses a
two-layer Graph Attention Network with dual attention heads to capture spatial relationships. At the same time, a two-layer
Transformer Encoder learns temporal patterns from historical charging data. These spatial and temporal features are combined
using a fusion layer and passed through a fully connected network with BatchNorm, ReLU, and Dropout to support multiple tasks,
including demand forecasting, behavior clustering, grid stress prediction, and anomaly detection. To improve interpretability, the
model includes SHAP-based feature analysis, node importance evaluation, and attention visualization. We evaluated the system on
1,320 charging sessions across 462 stations in five U.S. cities, achieving strong results, including an R2 of 0.9494. The system is
deployed using FastAPI with a React dashboard for real-time monitoring.
Keywords- Electric Vehicles, Spatio-Temporal Graph Attention Network, Demand Forecasting, Anomaly Detection, Smart Grid, Deep
Learning.

I. INTRODUCTION
The rapid growth of electric vehicles (EVs) worldwide is transforming urban energy systems at an unprecedented pace. Global EV adoption crossed 40 million vehicles in
2023 and is projected to reach over 145 million by 2030 [ l]. While this growth is encouraging for sustainability and reduction of carbon emissions, it also places increasing
pressure on existing charging infrastructure and energy distribution networks. As EV usage expands in urban areas, charging networks are beginning to face issues such as peak demand spikes, uneven station utilization, and highly variable user behavior patterns across different regions and time periods [2].
Most traditional approaches to managing EV charging systems treat problems like energy demand forecasting, user behavior analysis, and grid monitoring as separate and independent tasks. In practice, however, these aspects are deeply interconnected, and separating them often leads to environment [ 4]. When combined with temporal models, these approaches can also capture how charging patterns evolve dynamically over time under varying conditions [ 5]. Machine learning techniques have shown clear advantages over traditional statistical methods, particularly in shortterm forecasting, adaptive learning, and anomaly detection tasks [6].
However, most ex1stmg solutions still focus on individual objectives rather than providing a unified and scalable framework. In addition, many deep learning models lack interpretability, making them less useful for real-world decision-making by infrastructure planners and policymakers who require transparency and trust in predictions [7] [8]. In this work, we propose STGATE (Spatio-Temporal Graph Attention Network with Transformer Encoder), a unified multi-objective model that simultaneously
performs next-hour demand forecasting, user behavior spatial and temporal dependencies between charging stations, ultimately reducing system efficiency and reliability [3].

To address this limitation, recent research has explored graph-based deep learning models, which are particularly effective in modeling spatial relationships across distributed charging stations within a networked within a single integrated architecture. By modeling charging stations as nodes in a spatial graph and incorporating an explainability framework, the proposed approach not only improves predictive performance but also provides meaningful and actionable insights. T he system is further supported by a React-based dashboard for real-time monitoring, interactive visualization, and efficient management of smart grid operations.

highlights the growing importance ofmodel interpretability and feature importance analysis in real-world deployments. Other works [13][14] explored hybrid machine learning setups for predicting charging patterns, availability, and waiting times, showing promising and consistent results across multiple evaluation benchmarks. Additionally, Attention-based LSTM models [8] improve short-term demand forecasting with better clarity and flexibility.

detection in isolation [7][8]. There is a lack of a unified framework that integrates spatial relationships and temporal dynamics while also ensuring interpretability. To bridge this gap, this work introduces STGATE, a unified spatio-temporal architecture with an embedded explainability layer. The proposed model is evaluated using real-world data collected from 462 charging stations across five U.S. cities.

1. METHODOLOGY

   1. Dataset and Preprocessing

      This study is based on a dataset containing 1,320 EV charging sessions collected from 462 stations across five major U.S. cities-Houston, San Francisco, Los Angeles, Chicago, and New York-during January and February 2024. Each session record includes details such as station ID, location, energy delivered, session duration, charging rate, initial state of charge, SoC gap, temperature, and user

      category (Commuter, Casual Driver, or Long-Distance Traveler).

      From these records, station-level features were derived and organized into sequences using a sliding window of six time steps. Each sequence captures five key attributes: total

      energy, session count, average SoC gap, temperature, and

      charging rate. Before training, all features were normalized

      using StandardScaler. The dataset was split into training (70%), validation (15%), and testing (15%) subsets.

   2. Graph Construction

      Charging stations are modeled as nodes within a graph structure, where connections are defined based on geographic closeness and similarity in usage patterns. Each node is described using 13 aggregated features, including energy statistics, session counts, charging behavior, peak usage, user distribution, charger types, and city encoding. The graph is implemented using PyTorch Geometric, allowing efficient handling of node relationships through an edge index representation.

   3. System Architecture

      Fig. 1 shows the overall STGATE system. The proposed STGATE framework follows a dual-encoder design that combines spatial and temporal learning. A Graph Attention Network (GAT) captures interactions between stations, while a Transformer Encoder processes historical usage sequences. The outputs from both components are merged through a fusion layer and passed to multiple prediction heads responsible for demand estimation, behavior analysis, grid stress evaluation, and anomaly detection.

      The model takes two inputs: a graph-based representation of stations and time-series sequences describing past activity. After preprocessing, each input is passed through its respective encoder. The GAT focuses on spatial dependencies, whereas the Transformer learns temporal trends. Their outputs are combined into a shared representation, which is further processed for multi-task predictions. An explainability module is also integrated to provide insights into model behavior.

      Fig. 1. STGATE system architecture

   4. Model Architecture Details

      STGATE consists of four main components. The GAT encoder uses two layers with dual attention heads and a hidden size of 32, along with batch normalization, ELU activation, and dropout to improve generalization. The Transformer encoder also contains two layers with a model dimension of 32 and a feedforward size of 128, followed by adaptive average pooling to produce a compact representation.

      The outputs of both encoders are concatenated and passed through a fully connected layer to create a unified feature vector. This vector is then used by four task-specific heads: one for predicting next-hour energy demand, one for generating embeddings for user behavior clustering, one for estimating grid stress as a continuous score, and one for classifying anomalies.

   5. Training Configuration

   The model is implemented using PyTorch and PyTorch Geometric. Training is carried out for up to 100 epochs with a batch size of 8. Early stopping is applied to prevent overfitting, and the Adam optimizer is used with a learning rate of le-3 and weight decay of le-4. A learning rate scheduler reduces the rate when validation performance plateaus.

   The overall loss function combines multiple objectives: mean squared error for demand and grid predictions, and cross-entropy loss for anomaly detection. Behavior clustering is performed separately using K-Means on the learned embeddings, with the number of clusters selected based on the Silhouette Score.

   F Explainability Framework

   To make the system more interpretable, a three-level explainability mechanism is included. SHAP values are used to identify important input features for demand predictions. Node importance scores highlight which stations have the most influence on overall results. Additionally, attention weights from the GAT layers are visualized to show how stations interact with each other. These explanations are accessible through a dedicated API endpoint.

   G. Deployment

   The trained model is deployed as a FastAPI backend that provides endpoints for all major functionalities,

   1. User Behavior Clustering

      The behavior modeling component generates 16- dimensional embeddings that were grouped using K-Means clustering. The optimal number ofclusters was found to be two, resulting in a Silhouette Score of 0.2177. While the score is moderate, it still indicates meaningful structure in the data, which is expected given the gradual and overlapping nature of real-world user behavior. The clusters broadly separate users into high-intensity and low­ intensity charging profiles. High-intensity users typically include long-distance travelers and those relying on fast charging, whereas low-intensity users are mostly commuters or occasional drivers. These pattens can help guide infrastructure planning-for example, identifying where higher power capacity is needed or where demand­ response strategies can be applied.

      including prediction, clustering, anomaly detection, and explainability. A React-based dashboard connects to this

backend using WebSockets and updates data every few seconds, enabling real-time monitoring. The system is hosted online using a Cloudflare tunnel, allowing remote access without requiring a dedicated server.

1. RESULTS AND DISCUSSION

   A. Demand Forecasting Performance

   The proposed STGATE model delivers strong performance in next-hour energy demand prediction, achieving an R2 score of0.9494 along with an MAE of4.05 kWh, RMSE of 5.33 kWh, and MAPE of 17.14%. An R2 value close to 0.95 indicates that the model captures nearly

   Fig. 3. User Behavior Clusters (K=2) and Silhouette Score vs K

   1. Grid Stress Prediction

      The grid stress module produced an average stress score of 0.4925 across all stations, suggesting that the overall system operates under moderate load conditions. However, 20 stations were identified as high-stress locations, with scores exceeding 0.7. These stations are primarily associated with areas that have a higher density of fast charging usage. Detecting such hotspots early is useful for grid operators, as it enables proactive measures like redistributing load or implementing load management strategies to maintain system stability.

      all variability in demand, which highlights the

      effectiveness of combining spatial and temporal learning. Compared to conventional LSTM-based approaches, which typically report R2 values in the range of 0.80—0.92 for similar tasks [9], STGATE performs noticeably better. This improvement can be attributed to the inclusion of graph-based spatial modeling, which allows the system to

      learn dependencies between nearby charging stations­ something purely temporal models often miss.

      Fig. 2. Demand Forecast-Actual vs Predicted (R’=0.9494)

      Fig. 4. Grid Stress Score Distribution and Timeline

   2. Anomaly Detection

      For anomaly detection, the model achieved an AUC­ ROC of 0.6273, indicating a reasonable but not highly strong ability to differentiate between normal and abnormal charging sessions. Anomalies were defined based on sessions with unusually high SoC gaps, specifically those above the 90th percentile. The relatively modest performance is partly due to class imbalance, as anomalous events are rare in real-world datasets [8]. To improve this component, future work may incorporate techniques such as oversampling (e.g., SMOTE) or modified loss functions

      like focal loss to better handle imbalance and improve detection sensitivity.

      Fig. 5. ROC Curve (AUC=0.6273) and Co11fusio11 Matrix

   3. Summary of Evaluation Metrics

   The STGATE model is evaluated across multiple tasks, including demand forecasting, clustering, grid stress analysis, and anomaly detection, to assess its overall effectiveness. For demand forecasting, the model performs very well, achieving a high R2 score of 0.9494, which shows that the predicted values closely match actual demand. The error values, including MAE (4.05 kWh) and RMSE (5.33 kWh), are low, indicating accurate predictions, while the MAPE of 17.14% is acceptable given the natural variability in EV charging behavior. In terms of clustering, the Silhouette score of 0.2177 suggests moderate separation between groups, with k = 2 providing a reasonable division of user behavior patterns.

   For grid stress analysis, the average stress score of 0.4925 indicates a moderate load on the system, and 20 instances are identified as high-stress events, showing the model’s ability to highlight potential pressure points in the grid. In anomaly detection, the AUC-ROC score of 0.6273 reflects fair perfonnance in detecting unusual charging activities. While this can be improved further, the overall results demonstrate that STGATE works effectively across different tasks, offering a practical and balanced solution for managing and analyzing EV charging systems in real­ world scenarios.

   TABLE I

   STGATE EVALUATION METRICS

   Task	Metric	Value	Result
   Demand Forecast	R’	0.9494	Excellent
   Demand Forecast	MAE(kWh)	4.05	Low Error
   Demand Forecast	RMSE(kWh)	5.33	Low Error
   Demand Forecast	MAPE(¾)	17.14	Acceptable
   Clustering	Silhouette	0.2177	Moderate
   Clustering	k	2	Optimal
   Grid Stress	Mean Score	0.4925	Moderate
   Grid Stress	High-Stress	20	Flagged
   Anomaly Det. AUC-ROC 0.6273 Fair

2. CONCLUSION

This paper presented STGATE, a spatio-temporal deep learning framework designed to handle multiple aspects of EV charging infrastructure in a unified and efficient

manner. The model combines a Graph Attention Network for capturing spatial relationships between stations and a Transformer Encoder for learning complex temporal usage patterns over time. By integrating these components, STGATE is able to perform four key tasks simultaneously: next-hour demand forecasting, user behavior clustering, grid stress prediction, and anomaly detection. Unlike many existing approaches that treat these problems separately, this unified design simplifies the overall system architecture, reduces computational redundancy, improves coordination between tasks, and makes it more practical and scalable for real-world deployment scenarios.

The model was evaluated on 1,320 charging sessions collected from 462 stations across five U.S. cities, providing a diverse and realistic dataset. The results show strong performance, particularly in demand forecasting with an R2 of 0.9494, indicating high predictive accuracy and reliable generalization within the dataset. The clustering output revealed two broad user groups-high­ intensity and low-intensity users-which can be useful for planning infrastructure capacity, optimizing resource allocation, and designing demand-response strategies. The grid stress analysis also helped identify high-load stations and potential bottlenecks, enabling better operational planning, while anomaly detection showed moderate but acceptable perfonnance given the inherent noise and imbalance in the data.

An important aspect of STGATE is its built-in explainability, which enhances its practical usability and trustworthiness. By combining SHAP analysis, attention visualization, and node importance scoring, the model provides insights that are easier to interpret compared to typical black-box approaches. This makes it more useful for planners and operators who need to understand the reasoning behind predictions and make informed decisions in dynamic environments. The system is further supported by a FastAPI backend and a React-based dashboard, enabling real-time monitoring, interactive visualization, and seamless integration into existing smart grid systems and operational workflows.

That said, there are still some limitations that need to be addressed. The dataset is relatively smal and limited to a short time period, which may affect the model’s ability to generalize across different regions and long-tenn conditions. The anomaly detection performance is also impacted by class imbalance, and the use of a static graph limits the model’s ability to capture evolving relationships between stations over time. Future work will focus on expanding the dataset, improving anomaly detection using techniques like oversampling and advanced loss functions, and exploring dynamic graph-based approaches to better model real-world variability, seasonal trends, and system evolution.

REFERENCES

[l] M. Rashid, T. Elfouly, and N. Chen, “A Comprehensive Survey of Electric Vehicle Charging Demand Forecasting Techniques,” IEEE Open

J. Veh. Technol., vol. 5, pp. 1348-1373, 2024, doi: I0.1109/ONT.2024.3457499.

1. X. Lin et al., “Electric Vehicle Charging Load Modeling: A Survey, Trends, Challenges and Opportunities,” Oct. 30, 2025, arXiv:2511.03741. doi: 10.48550/arXiv.25 l1.03741.

2. M. l(jm, W. Y. Park, H. S. Moon, and Y.-S. Kim, “A Review of Electric Vehicle Charging Demand Forecasting Models by Prediction Horizon: A Multi-Criteria Decision Analysis Approach,” J. Electr. Eng. Technol., Jan. 2026, doi: 10.1007/s42835-025-02556-0.

3. Y. Yang, Y. Zhang, and X. Meng, “A Data-Driven Approach for Optimizing the EV Charging Stations Network,” IEEE Access, vol. 8, pp. l l 8572-118592, 2020, doi: l 0.1!09/ACCESS.2020.3004715.

4. A. Hussain, V. Eswarakrishnan, A. Aslam, and S. Tripura, “Charging stations demand forecasting using LSTM based hybrid transformer model,” Sci. Rep., vol. 15, no. 1, p. 36639, Oct. 2025, doi: l 0.1038/s4 l 598-025-2042 l-y.

5. R. N. Akshayaa, S. SA, and R. R, “Session and energy forecasting for electric vehicle charging station using custom weighted ensemble machine learning,” Results Eng., vol. 26, p. 105104, Jun. 2025, doi: l 0.1016/j.rineng.2025.105104.

6. S. Sanami, H. Mosalli, Y. Yang, H.-G. Yeh, and A. G. Aghdam, “Demand Forecasting for Electric Vehicle Charging Stations using Multivariate Time-Series Analysis,” Feb. 22, 2025, arXiv: arXiv:2502.16365. doi: I 0.48550/arXiv.2502.16365.

7. V. T. Kesavan, Md. J. Hassen, R. Gopi, and E. R. Joseph, “Anomaly detection with grid sentinel framework for electric vehicle charging stations in a smart grid environment,” Sci. Rep., vol. 15, no. 1, p. 15774, May 2025, doi: 10.1038/s41598-025-00400-z.

8. Y. l(jm and S. Kim, “Forecasting Charging Demand of Electric Vehicles Using Time-Series Models,” Energies, vol. 14, no. 5, p. 1487, Mar. 2021, doi: 10.3390/enl4051487.

9. K. C. Akshay, G. H. Grace, K. Gunasekaran, and R. Samikannu, “Power consumption prediction for electric vehicle charging stations and forecasting income,” Sci. Rep., vol. 14, no. I, p. 6497, Mar. 2024, doi: I0.1038/s41598-024-56507-2.

[II] V. Kaur, D. Sharma, S. K. Dwivedi, M. K. Yadav, A. K. L. Singh, and P. Shanna, “Prediction of Electric Vehicle Charging Using Machine Leaming,” SSRN Electron. J., 2024, doi: I0.2139/ssm.4495876.

1. T. Van Etten, V. Degeler, and D. Luo, “Large-Scale Forecasting of Electric Vehicle Charging Demand Using Global Time Series Modeling:,” in Proceedings of the I0th International Conference on Vehicle Technology and Intelligent Transport Systems, Angers, France: SClTEPRESS – Science and Technology Publications, 2024, pp. 40-51. doi: 10.5220/0012555400003702.

2. Z. Yi, X. C. Liu, R. Wei, X. Chen, and J. Dai, “Electric vehicle charging demand forecasting using deep learning model,” J. Intel/. Transp. Syst., vol. 26, no. 6, pp. 690-703, Nov. 2022, doi:

   10.1080/15472450.2021. l 966627.

3. A. Hussain, Q.-C. Lu, S. S. Rizvi, S. Wang, and S. J. Kwon, “Short term demand forecasting of electric vehicle charging stations using context aware temporal transformer model,” Sci. Rep., vol. 15, no. I, p. 36652, Oct. 2025, doi: I0.1038/s41598-025-20557-x.

4. F. Gon9alves, L. De Abreu Borges, and R. Batista, “Electric Vehicle Charging Data Analytics of Corporate Fleets,” World Electr. Veh. J., vol. 13, no. 12, p. 237, Dec. 2022, doi: l 0.3390/wevj I3120237.

5. M. Manai, B. Sellami, and S. Ben Yahia, “Accurate Recommendation of EV Charging Stations Driven by Availability Status Prediction:,” in Proceedings of the I9th International Conference on Sofiware Technologies, Dijon, France: SCJTEPRESS – Science and Technology Publications, 2024, pp. 351-358. doi: l 0.5220/0012752600003753.

6. A. Hafeez, R. Alammari, and A. Iqbal, “Utilization of EV Charging Station in Demand Side Management Using Deep Learning Methnd,” IEEE Access, vol. 11, pp. 8747-8760, 2023, doi: I0.1109/ACCESS.2023.3238667.

7. A. Kumari, “Machine Learning-Based Prediction for Ev Charging Statinn Availability and Wait-Time Estimation,” 2024, SSRN. doi: I0.2139/ssm.5035693.

8. Z. Tang, Y. Cui, Q. Hu, M. Liu, W. Rao, and X. Liu, “Electric vehicle charging load prediction based on graph attention networks and autofonner,” J. Eng., vol. 2024, no. 9, p. e70009, Sep. 2024, doi: I0.1049/tje2.70009.

9. Y. Zhang, C. Liu, X. Rao, X. Zhang, and Y. Zhou, “Spatial-temporal load forecasting of electric vehicle charging stations based on graph neural network,” J. Intel/. Fuzzy Syst., vol. 46, no. 1, pp. 821-836, Jan. 2024, doi: l 0.3233/JIFS-231775.