Data-driven models have been widely used in building heating load prediction, but often fail when facing limited data. Previous studies have shown transfer learning can assist model learning of target building under limited data by means of other source building data, however, which is subject to the similarity between source and target building. Selecting similar source building data is not easy, especially when the target building is with limited data. This paper, therefore, proposes a novel meta learning-based framework for building heating load prediction. Using meta learning method, a set of promising model parameters is trained by local and global learning on multiple source buildings data. The obtained model parameters has the ability to get quickly trained with few data in each source building, which is further used as model initialization parameters of target building to assist model learning. Framework validity is confirmed by 550 groups of practical buildings data (50 are as target buildings for testing and 500 are as source buildings). The results showed the proposed framework could reduce the prediction errors by 2.04 %∼61.59 % compared with six common transfer learning methods. The novel meta learning-based framework provides an effective solution for building heating load prediction with limited data.

中文翻译：

---

面向有限数据的建筑热负荷预测方法：一种新颖的基于元学习的框架

数据驱动模型已广泛应用于建筑热负荷预测，但在面对有限数据时往往会失败。先前的研究表明，迁移学习可以通过其他源构建数据来辅助有限数据下目标构建的模型学习，但这取决于源构建和目标构建之间的相似性。选择相似的源建筑数据并不容易，特别是当目标建筑数据有限时。因此，本文提出了一种新颖的基于元学习的建筑热负荷预测框架。使用元学习方法，通过对多个源建筑物数据的本地和全局学习来训练一组有前途的模型参数。获得的模型参数能够在每次源构建中用少量数据快速训练，并进一步作为目标构建的模型初始化参数来辅助模型学习。框架有效性通过550组实际建筑数据（50组作为测试目标建筑，500组作为源建筑）进行验证。结果表明，与六种常见的迁移学习方法相比，所提出的框架可以将预测误差降低2.04%∼61.59%。这种新颖的基于元学习的框架为利用有限数据构建热负荷预测提供了有效的解决方案。