[Paper Review] Deep Spatio-Temporal Forecasting of Electrical Vehicle Charging Demand

Frederik Boe Hüttel, Inon Peled+2 more|arXiv (Cornell University)|Jun 21, 2021

Electric Vehicles and InfrastructureEngineering20 references18 citations

TL;DR

This paper proposes a Temporal Graph Convolutional Network (T-GCN) to forecast electric vehicle (EV) charging demand by modeling spatial-temporal correlations across charging stations using publicly available datasets. The T-GCN outperforms CNN, CNN+LSTM, and traditional baselines, achieving the lowest RMSE across short- and long-term horizons, with a 61±8 RMSE for 1-day forecasts and 161±15 for 30-day forecasts.

ABSTRACT

Electric vehicles can offer a low carbon emission solution to reverse rising emission trends. However, this requires that the energy used to meet the demand is green. To meet this requirement, accurate forecasting of the charging demand is vital. Short and long-term charging demand forecasting will allow for better optimisation of the power grid and future infrastructure expansions. In this paper, we propose to use publicly available data to forecast the electric vehicle charging demand. To model the complex spatial-temporal correlations between charging stations, we argue that Temporal Graph Convolution Models are the most suitable to capture the correlations. The proposed Temporal Graph Convolutional Networks provide the most accurate forecasts for short and long-term forecasting compared with other forecasting methods.

Motivation & Objective

• Address the challenge of accurately forecasting EV charging demand to support green energy integration and infrastructure planning.
• Overcome limitations of existing methods that ignore spatial correlations between charging stations or rely on non-public data.
• Demonstrate the feasibility and effectiveness of using publicly available EV charging datasets for large-scale forecasting.
• Evaluate the performance of graph-based deep learning models, particularly T-GCN, against conventional time-series and convolutional models.
• Provide a reproducible, open-source framework for future research in EV charging demand forecasting using graph neural networks.

Proposed method

• Model the EV charging network as a graph where nodes represent charging stations and edges encode spatial proximity or correlation.
• Apply Temporal Graph Convolutional Networks (T-GCN) that combine graph convolutions with recurrent layers (LSTM) to capture dynamic spatio-temporal dependencies.
• Use a raster map representation as an alternative spatial encoding for comparison, where grid cells aggregate demand from stations within them.
• Train models using historical daily energy demand data from 2012 to 2019, with input sequences of 30 days for short-term and 120 days for long-term forecasts.
• Optimize models using the ADAM optimizer and a loss function with L2 regularization (λ = 10⁻³) over 1000 epochs.
• Implement models using Keras and the Stellar Graph library for graph operations, ensuring reproducibility and open access via GitHub.

Experimental results

Research questions

• RQ1 Can publicly available EV charging datasets support accurate and comparable forecasting of charging demand?
• RQ2 How do graph-based models like T-GCN compare to traditional time-series and convolutional models in forecasting EV charging demand?
• RQ3 To what extent do spatial correlations between charging stations improve forecasting accuracy?
• RQ4 Does modeling the network as a graph yield better performance than raster-based or aggregated time-series approaches?
• RQ5 Can T-GCN effectively capture both short-term (1–7 days) and long-term (30 days) temporal patterns in EV charging demand?

Key findings

• The T-GCN model achieved the lowest RMSE across all forecast horizons, with 61±8 for 1-day, 184±9 for 7-day, and 161±15 for 30-day forecasts.
• Compared to the AR(30) baseline (RMSE 252 for 30-day forecast), T-GCN reduced error by 36.5%, demonstrating significant improvement.
• The T-GCN outperformed the CNN+LSTM model by 28.4% in 1-day forecasting and 7.5% in 30-day forecasting, highlighting the advantage of graph-based modeling.
• The graph-based approach captured subtle spatial differences and inter-station correlations better than raster-based convolutional methods.
• All models, including T-GCN, successfully captured weekly demand patterns, with lower demand on weekends, indicating robust temporal learning.
• The use of publicly available datasets enabled reproducible and transparent benchmarking, supporting broader adoption in EV infrastructure research.