A multi-source transfer learning model based on LSTM and domain adaptation for building energy prediction

Transfer learning can use the knowledge learned from the operating data of other buildings to facilitate the energy prediction of a target building. However, most of the current research focuses on the transfer of a single source building of the same type or from the same region. A single source domain produces domain shift due to the difficulty of aligning the distribution with the target domain. To address this problem, this paper proposes a novel multi-source transfer learning energy prediction model based on long short-term memory (LSTM) and multi-kernel maximum mean difference (MK-MMD) domain adaptation. This model was used for the short-term energy prediction of different types of buildings lacking historical data. In addition, dynamic time warping (DTW) was used to select the source domain. Multiple multi-source models and corresponding single-source models were compared on a collection of buildings in the Higashida area of Fukuoka Prefecture, Japan. On the experimental datasets, the results showed that DTW relatively accurately measured the similarity between building energy consumption datasets. Compared with the results of the single-source transfer learning models, the multi-source transfer learning models achieved better average prediction performance, and their mean absolute percentage error (MAPE) improved the prediction accuracy by 6.88–15.37%.