Spatiotemporal Prediction of Electric Vehicle Charging Load Based on Large Language Models
Hang Fan 1 , Mingxuan Li 2 , Jingshi Cui* 3 , Zuhan Zhang 1 , Wencai Run 1 and Dunnan Liu 1 · 2025
Read Spatiotemporal Prediction of Electric Vehicle Charging Load Based on Large Language Models by Hang Fan 1 , Mingxuan Li 2 , Jingshi Cui* 3 , Zuhan Zhang 1 , Wencai Run 1 and Dunnan Liu 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-03728.txt.
Introduction
Hang Fan and five co-authors forecast electric vehicle charging demand with a system that pairs graph networks with a large language model. Grid operators need to know when and where cars will plug in, and demand at one station depends on its neighbors and on the weather. Their model, EV-LLM, feeds historical charging records, weather data, and written descriptions into one pipeline.
A graph convolutional network first extracts how stations influence each other across space and time. A fine-tuned language model then generalizes across the mixed inputs. Tests on real records from ten California stations showed the system beating standard deep learning forecasters. The authors say better forecasts of this kind help schedule grid load and support services that let parked cars feed power back to the grid.
Catalog record
| CREATOR | Hang Fan 1 , Mingxuan Li 2 , Jingshi Cui* 3 , Zuhan Zhang 1 , Wencai Run 1 and Dunnan Liu 1 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 23 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-03728
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Spatiotemporal Prediction of Electric Vehicle Charging Load Based on Large Language Models Hang Fan 1 , Mingxuan Li 2 , Jingshi Cui* 3 , Zuhan Zhang 1 , Wencai Run 1 and Dunnan Liu 1 Abstract The rapid growth of EVs and the subsequent increase in charging demand pose significant challenges for load grid scheduling and the operation of EV charging stations. Effectively harnessing the spatiotemporal correlations among EV charging stations to improve forecasting accuracy is complex. To tackle these challenges, we propose EV-LLM for EV charging loads based on LLMs in this paper. EV-LLM integrates the strengths of Graph Convolutional Networks (GCNs) in spatiotemporal feature extraction with the generalization capabilities of fine-tuned generative LLMs. Also, EV-LLM enables effective data mining and feature extraction across multimodal and multidimensional datasets, incorporating historical charging data, weather information, and relevant textual descriptions to enhance forecasting accuracy for multiple charging stations. We validate the effectiveness of EV-LLM by using charging data from 10 stations in California, demonstrating its superiority over the other traditional deep learning methods and potential to optimize load grid scheduling and support vehicle-to-grid interactions. I Introduction The transition to a low-carbon economy has underscored the importance of EVs as a promising solution to reduce carbon emissions and mitigate fossil energy shortages [ 1 ] . In 2024, the global adoption of EVs has increased significantly, with sales surpassing 10 million units between January and August, which accounts for 18\% of total new car sales, and shows an increase of approximately 3\% compared to 2023. However, this rapid growth in EV adoption, coupled with the escalating demand for charging, poses substantial challenges for power grid management. The variability in EV charging loads can disrupt the stable operation of distribution networks, complicating efforts for optimization and scheduling. Moreover, EV manufacturers and charging infrastructure providers are increasingly participating in vehicle-to-grid interactions and electricity markets as load aggregators, as exemplified by initiatives such as Tesla’s Auto-Bidder. Consequently, accurate prediction of EV charging loads is essential to effectively address these challenges. Research on EV charging load forecasting is divided into model-driven and data-driven methods. Model-driven methods involve establishing mathematical models to deduce user behavior under various factors and simulate the spatiotemporal distribution characteristics of charging loads. [ 2 ] proposed a dynamical evolution model for the spatial and temporal distribution of EV’s charging demands based on Agent-cellular automata. [ 3 ] proposed an EV charging load forecasting method that considers the real-time interaction of multi-source information and user regret psychology. The authors used the Monte Carlo method to simulate the travel and charging processes and obtain the spatiotemporal distribution of charging loads within the area. With the development of big data, the data-driven EV load forecasting methods have gradually gained attention. [ 4 ] proposed a novel ISSA-CNN-GRU model to combine the feature extraction capabilities of CNN with the time series forecasting capabilities of GRU and optimized hyperparameters using an improved sparrow search algorithm for EV charging load forecasting. In response to the uncertainty and long-term forecasting issues, [ 5 ] proposed a new algorithm based on dynamic adaptive graph neural networks. The algorithm effectively captures the spatiotemporal fluctuations of charging loads through spatiotemporal correlation feature extraction which enhances the model’s forecasting ability. Experimental results show that the method achieved more accurate load forecasting on the Palo Alto dataset. [ 6 ] introduced a heterogeneous spatial-temporal GCN designed to predict EV charging demand at various spatial and temporal resolutions, which achieved this by constructing heterogeneous graphs that capture the spatial correlations between different charging regions. With the rise of generative artificial intelligence technologies, the application of large language models for time series forecasting has begun to attract attention [ 7 ] . For example, [ 8 ] utilized a pre-trained LLM for wind speed prediction and proved its superiority over other traditional deep learning models which indicates its applicability for energy time series. Although existing works have proposed various EV load forecasting methods, we find that existing methods often struggle to adapt to the dynamic spatial-temporal relationship of user charging behavior, which is influenced by multiple factors, including time, location, and external conditions. To address these limitations, this paper proposes a novel approach combining LLMs and GCNs to enhance spatiotemporal prediction tasks for EV charging loads. By leveraging the capabilities of LLMs, we aim to better model the intricate dynamics of charging behavior compared to traditional methods. Our approach seeks to improve the understanding and forecasting of evolving charging demand patterns, ultimately providing a robust solution for load grid management and facilitating market participation by load aggregators. In summary, the contributions of our proposed EV load prediction method can be summarized as follows: • Adoption of LLMs: To our best knowledge, this paper is the first one to combine LLMs and GCNs to enhance EV charging load forecasting, presenting a novel approach that diverges from traditional and solely data-driven methodologies. • Enhanced Prediction Accuracy: By integrating GCNs with LLMs, our method effectively captures complex spatiotemporal dependencies and multi-variable interactions, addressing the limitations of existing models that struggle with long sequence dependencies. • Reprogramming and Embedding Innovations: We introduce a unique reprogramming layer designed to efficiently map time series data into a format suitable for LLMs processing, thereby enhancing the model’s ability to interpret and predict EV charging patterns. II Problem Description and Model Structure II-A Problem Description Forecasting the load at EV charging stations is fundamentally a time series forecasting problem that can be modeled uniformly. The charging load in the next H time steps can be predicted using historical charging load data from the previous M time steps and numerical weather prediction (NWP) data for the next N steps. This relationship can be formulated as: \displaystyle P_{t+1},\ldots P_{t+H}= (1) \displaystyle f\left(P_{t-1},\ldots P_{t-M},{NWP}_{t+1},\ldots{NWP}_{t+N},\Lambda\right), where P_{t}\in R^{n\times 1} represents the vector of EVs chargining load, and NWP_{t}\in R^{n\times k} denotes the NWP variables. Here, n is the number of EV charging stations, k is the number of the NWP variables, \Lambda is the adjacent matrix about the correlation among stations, and f is the prediction model. II-B Overall Architecture of EV-LLM Model In this paper, we introduce the EV-LLM model, a framework for EV load prediction based on LLMs. This approach enhances the LLM’s reasoning capabilities for time series by utilizing pretrained word embeddings and prompts to improve task comprehension. By employing customized multimodal inputs and performing minimal parameter training on the LLMs, the proposed model demonstrates exceptional abilities in EV load prediction tasks. The architecture of proposed EV-LLM model is illustrated in Fig. 1 and consists of three primary components. While LLMs are computationally intensive, to reduce training cost we focused on training specific components to better align with the requirements of spatiotemporal prediction task instead of tuning the entire model. Fig. 1: Overall Architecture of EV-LLM Model II-B 1 Multimodal Feature Fusion Module This module aligns textual prompts with reconstructed time series data, embedding the time series using projections from pretrained LLMs. By reshaping the data into a format comprehensible to the LLMs, it maps the time series into the same high-dimensional space as the prompt embeddings which enables the LLMs to interpret the prediction task more effectively. Detailed discussions regarding this module can be