Transferability and robustness of a data-driven model built on a large number of buildings

Abstract

Data-driven energy prediction models have shown a great importance in building energy management. However, these models require sufficient operational data to ensure prediction accuracy, posing great challenges for buildings with scarce data. Transfer learning has emerged as a key strategy to overcome this issue, enabling cross-building prediction even with limited data availability. While existing studies have mainly focused on a single or a few specific buildings to train models, this study aims to explore the transferability of building energy prediction models across a large number of buildings. Initially, energy consumption data retrieved from 327 buildings were selected as source domain to create a pre-trained model. Different volumes of data in the target domain were then utilized to fine-tune the pre-trained model, and the resulting accuracy distribution and accuracy improvement achieved by transfer learning were examined. The study also evaluated model robustness by conducting transfer procedures at 20 different time nodes. The occurrence of negative transfer was also monitored. The results show that transfer learning can significantly improve prediction accuracy when compared with the baseline model, with a median MAPE value improving from 18.31 % to 7.76 % when using only 7 days data. Meanwhile, transfer learning using 7 days data outperformed direct prediction using 180 days of data. However, negative transfer may occur, although at a low rate, and is not related to data volume. In addition, there is a possibility that a model with high general accuracy yield biased results at certain time nodes. This work provides valuable insights into the advantages and limitations of transfer learning in building energy prediction model, which facilitates the exploitation of existing building data source for advanced data analytics.