A recommendation model for optimizing transfer learning hyper-parameter settings in building heat load prediction with limited data samples

Abstract

The transfer learning method has gained increasing attention in the domain of building load prediction, particularly in scenarios with limited data samples. Its core principle involves leveraging knowledge obtained from abundant data in source buildings to aid the learning process of models for the target buildings. Existing research has predominantly concentrated on optimizing the selection of source building data to improve transfer learning effectiveness, while the optimization of transfer learning hyper-parameter settings is often neglected. This study proposes a recommendation model tailored for transfer learning hyper-parameter settings in the context of small sample prediction for building heat loads. The objective is to automatically suggest suitable transfer learning hyper-parameter combination based on the specific features of the building heat load data samples. In this study, 200 real building profiles were utilized to generate the input–output dataset required for the recommendation model. By employing data mining techniques such as clustering and classification, the correlation between the features of source building data and the most effective transfer learning hyper-parameter combination is investigated. The developed recommendation model for optimal transfer learning hyper-parameter settings achieves a classification accuracy of 90.5%，and the performance evaluation was conducted using an additional dataset of 30 source buildings. The results show that by employing this recommendation model, the prediction error of the target buildings can be reduced by 0.12% to 6.64% compared to the conventional method of empirically determining transfer learning hyper-parameter settings.