Spatio-temporal data fusion framework based on large language model for enhanced prediction of electric vehicle charging demand in smart grid management
Yitong Shang, Wen‐Long Shang, Dingsong Cui, Peng Liu, Haibo Chen, Dongdong Zhang, Runsen Zhang, Chengcheng Xu, Ye Liu, Chenxi Wang, Mohannad Alhazmi
Abstract
Accurate prediction of electric vehicle (EV) charging demand is pivotal for effective smart grid management and renewable energy integration. However, predicting spatio-temporal EV charging patterns remains challenging due to complex data fusion requirements arising from heterogeneous temporal, spatial, and contextual features, as well as difficulties in effectively integrating multiple modeling approaches. This paper introduces EV-STLLM, a novel spatio-temporal data fusion framework based on Large Language Model explicitly designed for accurate short-term EV charging demand forecasting through innovative integration of data-level and model-level fusion techniques. At the data level, a multi-source embedding module is developed to seamlessly fuse temporal features (e.g., time slots, weekdays), spatial heterogeneity (e.g., geographical location), and contextual charging behaviors into a unified representation via embedding convolutional network. At the model level, a large language model (LLM) is employed to capture global spatiotemporal dependencies, enhanced with Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning, substantially reducing computational costs while maintaining prediction robustness. Using a comprehensive real-world dataset comprising over 830,000 EV charging records across 16 districts and 331 subdistricts in Beijing, we validate EV-STLLM across multiple forecasting scenarios (district and subdistrict levels, one-step and two-step ahead predictions). Extensive comparative evaluations demonstrate that EV-STLLM consistently outperforms classical, graph-based, and deep learning baselines. Specifically, in one-step ahead district-level forecasting, EV-STLLM achieves up to a 15.41% reduction in MAE and a 53.51% reduction in MAPE compared to the leading baseline, underscoring its potential to significantly enhance data-driven smart grid operations.