A hybrid deep transfer learning strategy for short term cross-building energy prediction

Highlights: • A deep transfer learning strategy is proposed for cross-building energy prediction. • The proposed strategy is employed to overcome the data shortage problem. • Domain adaptation can well overcome the domain shift between two domains. • The proposed strategy can enhance the building energy prediction performance. • LSTM can extract temporal features better than CNN and FC layer. To overcome the data shortage problem of model training in building energy prediction, this study proposes a novel hybrid deep transfer learning strategy for short term cross-building energy prediction using long short term memory (LSTM) and domain adversarial neural network (DANN). The proposed strategy can utilize knowledge learned from relevant building data to assist the energy prediction for target buildings with limited historical measurements. LSTM based feature extractor is used to extract temporal features across source and target buildings. DANN attempts to find domain invariant features between the source and target buildings via domain adaptation. Then, the domain adaptation based transfer learning model (i.e. LSTM-DANN) trained with data from source buildings can be applied to assist in predicting the target building energy without prediction performance degradation caused by domain shift. Experiments are conducted to evaluate the performance of the proposed transfer learning strategy in different models. Results demonstrate that the proposed strategy can significantly enhance the building energy prediction performance compared to models trained on the target only data, the source only data, both the target and source data, but without transfer learning. This work can provide guidance for the effective use of existing building data resources.