Summary
Building-specific energy forecasting is held back by scarce data and high development cost. This study measures time-series similarity with TS2Vec, an unsupervised contrastive encoder, on anonymized electricity data from buildings in Cambridge, UK.
By adjusting the time window used to extract representation vectors, the method finds highly similar data segments better than a plain average similarity. PCA and a moving-window cosine similarity make the similarity of complex multivariate series readable.
- Dynamic similarity captures the complexity of building energy use patterns.
- It can optimize transfer learning by automatically finding the most suitable source data.
- It opens the way to dynamic model selection and ensembles driven by how similarity changes over time.
This work led to the privacy-preserving transfer learning framework published in Applied Energy.
Authors
Wonjun Choi, Sangwon Lee, Max Langtry, Ruchi Choudhary