A hierarchy learning strategy for probabilistic power demand prediction for buildings

Accurate power demand prediction is essential for maintaining supply-demand balance, reducing energy costs, and supporting proactive building energy management. Machine learning approaches have demonstrated strong performance in predicting energy and power demand. However, they typically rely on abundant high-quality historical data, which are often unavailable for newly monitored buildings. To address the challenge of low prediction accuracy under data scarcity, this study explores the potential of hierarchy learning (HL) to improve overall and peak power demand prediction with limited data. First, a similarity analysis method, Dynamic Maximum (DM), combining dynamic time warping and maximum mean discrepancy, is developed to identify the source domain data. The selected source and target domain data are then used within an HL framework to train a decomposition-based prediction model, namely a multiple seasonal-trend decomposition using LOESS combined with a Bayesian additive regression tree (MSTL-BART), forming the proposed DM-MSTL-BART-HL framework. The framework is evaluated using real-world data from residential buildings. Our results demonstrated that the framework significantly reduced CV-RMSE error with only one week of target domain data, achieving predictive accuracy improvements ranging from 2.58% to 19.18% compared to the baseline model, ASHRAE requirements, and other state-of-the-art methods. For peak demand prediction, the framework achieved CV-RMSE values ranging from 2.96% to 21.08%, with corresponding coverage width criterion values between 0.24 and 1.23, demonstrating a strong balance between prediction interval width and coverage. This method highlighted the potential of DM-MSTL-BART-HL as a practical solution for building power demand prediction when sufficient historical data are not available.