A novel building energy consumption prediction method based on multivariate time series alignment and dual-branch joint learning

Accurate building energy consumption prediction is essential for optimizing building operations, promoting energy conservation, and reducing carbon emissions. While multivariate time series (MTS) fusion methods have advanced prediction accuracy, further gains are often constrained by locally stationary lead-lag relationships among variables. Aiming at this issue, a novel prediction method based on MTS alignment and dual-branch joint learning (MTA-DBJL) is proposed to achieve accurate building energy consumption prediction. First, a novel MTS alignment method based on time offset cross-correlation (TOCMTA) mechanism is developed to address local lead-lag relationships between auxiliary variables and energy consumption variable. By analyzing the correlation between auxiliary variables and energy consumption variable in the time dimension, the time offset steps of auxiliary variables relative to energy consumption variable are accurately estimated. A time shift strategy is then employed to eliminate the offset, realize the time alignment of auxiliary variables and energy consumption variable, and ensure the accuracy of temporal consistency. Second, Seasonal and Trend decomposition using Loess (STL) is employed to decompose the energy consumption variable into seasonal, trend, and residual components, which allows the decoupling of complex temporal patterns into simpler components, thereby highlighting the inherent characteristics of energy consumption variable. Finally, a novel DBJL is designed to extract features from the temporal and the weakly temporal residual variables in parallel, which are then fused to predict future energy consumption. Through a series of experimental comparisons on the AEP and IHEPC building energy consumption datasets, the excellent predictive performance of the developed MTA-DBJL was verified.