Load forecasting of buildings using LSTM based on transfer learning with variable source domain

Insufficient historical data severely affects the accuracy of the long short-term memory(LSTM) in predicting building loads. Transfering and using the energy consumption data of other similar buildings in the source domain can improve the prediction accuracy of LSTM processing of the buildings in the target domain. However, the existing methods do not take into account the change of the source domain matching relationship caused by the increase in data during the prediction process. In view of this, an LSTM-based building load prediction method guided by transfer learning with variable source domain is proposed. During the execution process, according to the correlation between the load of the source domain building and that of the target building in new windows, the source domain building and its energy consumption data to be selected are adjusted in real time to ensure that the source and target domains always remain high similarity. Finally, the application on several typical examples shows that compared with the traditional fixed source domain transfer learning method, the proposed LSTM-based load prediction method guided by transfer learning with variable source domain can always maintain a higher prediction accuracy. © 2021, Editorial Office of Control and Decision. All right reserved.