Multiple Time Scale Deep Expert System for Load Forecasting of Electric Vehicle Charging Stations

• Published in: IEEE transactions on smart grid Vol. 16; no. 5; pp. 4015 - 4031
• Main Authors: Dong, Hanjiang, Li, Shenglin, Wen, Xiyu, Liang, Zipeng, Yang, Haosen, Chung, Chi-Yung, Zhu, Jizhong
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.09.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Cluster analysis; Clustering; Deep learning; elastic net; Electric vehicle; Electric vehicle charging; Electric vehicle charging stations; Electric vehicles; Engines; Ensemble learning; expert system; Expert systems; forecast combination; Load forecasting; Long short term memory; Multivariate analysis; Neural networks; Time of use; Time series analysis; Transformers; Vector quantization
• ISSN: 1949-3053; 1949-3061
• DOI: 10.1109/TSG.2025.3579879

Different types of electric vehicle charging stations (EVCSs) exhibit varying characteristics as affected by time-of-use (TOU) pricing rates for their peak, sharp, and valley periods. This paper develops a multiple-time-scale coordinated deep expert system framework that predicts the load demand of public, highway, and bus EVCSs. In the framework, the expert inference engine generates both the quarter-hourly and monthly predictions. For quarter-hourly load predictions, a stack ensemble learning strategy is proposed to combine heterogeneous neural network methods in a scalable, modular manner. Then, the multivariate data, regularized according to EVCS types and TOU rates, is processed via seasonal and trend decomposition using Loess (STL), and the daily profiles are grouped through multivariate time series K-Means clustering to ensure the consistency of the STL trend. For monthly energy predictions, a multi-matrix elastic net using blockwise coordinate descent is designed to correlate the remainder predictions (offered by the short-term engine and aggregated into monthly intervals) and the trend component (with cluster labels of ECVS types and TOU rates). In the case study, a real-world dataset with volatility is used, such that the proposed framework is implemented to show that forecast combinations can leverage multiple preferences, and short-term predictions can improve monthly performance.