A two-stage hierarchical clustering and transformer-BiLSTM hybrid framework for electric vehicle charging load forecasting
Liang Yuan, Jiayi Zhong, Yonglu Liu, Xubin Liu, Yunqi Wang, Zhao Yang Dong
Abstract
The global transition toward sustainable energy systems has positioned electric vehicles (EVs) as a cornerstone of decarbonized transportation. However, the inherent spatiotemporal variability and uncertainty of EV charging loads continue to challenge power grid stability and operational efficiency. To address these challenges, this paper introduces a novel forecasting framework that integrates a two-stage hierarchical clustering technique with a Transformer-BiLSTM hybrid deep learning model, aiming to enhance both prediction accuracy and interpretability. First, the framework employs a two-stage clustering algorithm to categorize heterogeneous user charging behaviors and meteorological factors, identifying distinct charging modes and their sensitivity to weather dynamics. Subsequently, the hybrid framework synergizes the Transformer’s global attention mechanism and BiLSTM’s localized temporal modeling to capture multi-scale dependencies across clustered user groups and dynamic weather features. Validated on real-world EV charging datasets, the framework demonstrates superior performance, achieving reduction in MAE and improvement in RMSE compared to benchmark models. By disentangling user-weather interactions and enhancing prediction adaptability, this approach provides actionable insights for grid operators to optimize load management and support sustainable energy transitions.