Accurate building energy prediction (BEP) is essential for effective energy management. However, current deep learning-based BEP approaches often show poor generalizability across different building types and rely heavily on large, high-quality datasets, resulting in significant performance drops in data-limited conditions. To address these limitations, we present a time-series domain adaptation model (TSDAM) designed to enable cross-building generalization using limited data from the target domain.
TSDAM employs transfer learning (TL) by combining: 1) a hierarchical feature extraction and temporal dependency module for multi-scale and temporally aware energy representation, and 2) a Gaussian stochastic domain adapter (GSDA). GSDA uses adversarial training to learn domain-invariant features, supporting robust adaptation to novel domains.
Extensive experiments conducted on the BuildingDataGenome 2 dataset, which includes six building types under mild, heavy, and extreme data scarcity conditions, confirm the effectiveness of TSDAM. It consistently surpasses baseline and comparison models. Specifically, under extreme data scarcity across six building types, TSDAM achieves an average mean absolute percentage error (MAPE) of 14.0%, outperforming BiLSTM (39.8%), domain-adversarial neural network (DANN) (28.3%), seasonal and trend transfer learning (ST-TL) (17.8%), CEEMDAN-SE-EC-BiLSTM (CSEB) (21.5%), and deep ensemble autoregressive hybrid model (DE-AR) (21.8%).
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号