Stacking deep transfer learning for short-term cross building energy prediction with different seasonality and occupant schedule

Development of accurate predictive models is essential for establishing optimal operation strategies reflecting future energy profiles of buildings. As a result, most precedent data-driven models are widely used and typically customized for individual buildings with a huge amount of training samples and data quality. However, buildings may be limited to gather measurements due to the lack of advanced measurement devices and acquisition time in both newly built and old building stocks. To mitigate this research gap, we proposed the deep transfer learning integrated with stacking ensemble (SDTL) for 1h-ahead building energy predictions by reusing the pre-trained model with datasets of other buildings even though there exist insufficient sample sizes. To validate the generalization performances, the experiments are divided into three different cases with different building characteristics and climates (i.e., different occupant schedule and season) according to deep learning (DL)-based feature extractors and data availability of the target building. As the results through three evaluation metrics such as mean absolute error (MAE), root mean squared error (RMSE), and mean absolute percentage error (MAPE), SDTL based on long-term short memory (LSTM) can reduce the MAE by approximately 6.4%–26.36%, RMSE by 2.54%–26.75%, and MAPE by 3.71%–30.5% in all experimental cases than other DL models. It was also shown that SDTL was relatively superior by achieving more generalized and stable results than baseline models. From the results of this study, we can provide valuable insights for development of advanced transferable prediction models with insufficient datasets.