Probabilistic Electricity Demand Forecasting for Buildings Using a Variational Mode Decomposition-Attentive Quantile Network

Abstract

Accurate and reliable forecasting of building electricity demand is essential for energy management, demand response, and operational planning in smart buildings. However, building load profiles are influenced by interacting factors such as weather conditions, occupant behavior, and operational schedules, leading to nonstationary and multi-scale temporal patterns with inherent uncertainty that traditional deterministic models struggle to represent. To address these challenges, this study proposes a Variational Mode Decomposition (VMD)-Attentive Quantile Network for probabilistic building load forecasting, designed to jointly capture complex temporal dynamics and predictive uncertainty. The framework leverages VMD to separate load signals into frequency-specific components, enhancing robustness against nonstationary effects. A sequential GRU–LSTM encoder models complementary short-term fluctuations and long-term dependencies, while Multi-Head Attention dynamically weights influential time steps. An IMF-dependent gating mechanism suppresses high-entropy noise components. Quantile regression is incorporated to directly estimate conditional demand distributions and generate prediction intervals at multiple confidence levels. The model is evaluated on hourly electricity consumption data (2019–2023) from three functionally distinct campus buildings at National Taiwan University. Results indicate consistent outperformance of VMD-based benchmark models in point forecasting, with R² improvements of 8.4%–12.9% and reductions in MAE and RMSE of 16.9%–40.8% and 13.5%–39.6%, respectively. In probabilistic forecasting, the model produces more compact and reliable prediction intervals, reflected by lower PINAW and AIS values across confidence levels. Overall, the proposed approach offers a robust and interpretable solution for short-term building load forecasting across heterogeneous load patterns.

@article{huang2026probabilistic,
  title={Probabilistic Electricity Demand Forecasting for Buildings Using a Variational Mode Decomposition-Attentive Quantile Network},
  author={Huang, Chun-Hao and Hu, Yu-Shin and Hsieh, I-Yun Lisa},
  journal={Energy \& Buildings},
  volume={366},
  pages={117696},
  year={2026},
  month=jul,
  doi={10.1016/j.enbuild.2026.117696}
}