A study on source domain selection for transfer learning-based cross-building cooling load prediction

Transfer learning (TL) can utilize data from information-rich source domain buildings to help information-poor target domain buildings establish cooling load prediction models, thereby enabling cross-building cooling load prediction. However, the prediction performance varies based on the type of source domain building used, making the reasonable selection of source domain buildings crucial. Traditionally, data similarity of source- target building pairs is measured to identify the appropriate source domain building. Yet, this method will result in erroneous judgments if the similarity index used is inappropriate, the target domain building's available data are scarce, or all candidate source domain buildings exhibit weak similarities. To address this issue, this study explores the selection of source domain buildings based on building feature consistency in a data-centric manner. Specifically, 81 buildings with different features and 240 source-target building pairs are designed. The TL method is then used to develop prediction models for each pair. The decision tree method is finally employed to evaluate the correlation between the feature consistency of source-target building pairs and model prediction performance. From this correlation, a selection strategy is derived that specifies the consistency requirements for three key features (scale, function, and climate zone) of source-target building pairs at different accuracy levels. This strategy offers a methodical guide for the selection of the source domain building.