Privacy-preserving transfer learning framework for building energy forecasting with fully anonymized data