Data augmentation strategy for short-term heating load prediction model of residential building

Data-driven models are widely used for short-term heating load prediction of buildings due to the advantages in mining actual load characteristics and improving prediction accuracy, which often require significant quantities of training data to ensure the strong generalization ability. However, insufficient data often exist in practice, which will seriously affect the prediction performance of data-driven models. This paper, therefore, proposes a data augmentation strategy to facilitate the training of data-driven prediction model under the condition of limited data. This strategy integrates a data augmentation source generated by calibrated simulation model of target building and a transfer learning-based data fusion method. Validity of this strategy is confirmed by practical case and data. The results suggest that under four different conditions of limited training data, the proposed data augmentation strategy could reduce short-term prediction errors of heating loads by 4.2%–18.14% compared with the models without data augmentation. Moreover, the proposed data augmentation strategy achieved the best result among four different data augmentation strategies. Through the contrastive analysis of different strategies, it can be concluded that the calibrated simulation model could provide high-quality augmented data and the transfer learning-based data fusion is more effective than direct data fusion.