FDA-HeatFlex: Scalable Privacy-Preserving Temperature and Flexibility Prediction for Heat Pumps using Federated Domain Adaptation

Heat pumps are a significant source of flexibility in energy systems since they can be operated flexibly, e.g., turned up when electricity is green (low CO2) or cheap, and turned down when electricity is expensive or mainly from fossil sources. However, the indoor temperature has to be kept within a user-specified comfort interval, e.g., 20-24° C, for residents to accept this flexible operation. To estimate the available flexibility, we need to know how the indoor temperature changes depending on the heat pump input power and outdoor temperature. Machine learning (ML) models can learn this given enough historical data, typically at least one year, to account for seasonal variations. However, for new buildings and/or newly retrofitted heat pumps, there is no or little data and users may be reluctant to share such sensitive data. To estimate the heat pump flexibility of such buildings, we propose FDA-HeatFlex (Federated Domain Adaptation Heat Pump Flexibility) framework where we transfer the knowledge from the source domain (a known building) to multiple target domains (new buildings) to accurately predict the indoor temperature of new buildings and derive their flexibility, making the prediction scale easily to many new buildings. Particularly, we leverage the idea of parameter-based transfer learning and adaptive boosting (AdaBoost) techniques for indoor temperature prediction to address the data shift problem, i.e., the discrepancy of data distributions between buildings, and employ the idea of federated learning to address the privacy concerns raised by data sharing between source and target domains. We conduct an extensive experimental evaluation on widely used real-world heat pump datasets which shows that our FDA-HeatFlex outperforms the state-of-the-art training approaches for indoor temperature prediction, and the state-of-the-art baseline for flexibility prediction with and improvement (on average), respectively.