Large-scale transfer learning for data-driven modelling of hot water systems

Hot water systems represent a substantial energy draw for most residential buildings. For design and operational optimization, they are usually either modelled by domain experts or through black-box models which makes use of sensor data. However, given the wide variability in hot water systems, it is impractical for a domain expert to individually model every hot water system. Likewise, black-box systems typically require an enormous amount of data to converge to a usable model. This paper makes use of transfer learning, a novel machine learning tool, to completely automate the learning process while substantially accelerating the performance of comparable black-box systems. Using real world data from 61 houses employing two different types of hot water systems, the proposed system is shown to work on both homogeneous and heterogeneous hot water systems. Convergence to a reliable model with transfer learning is on the order of a few weeks, as opposed to months or years without transfer. By presenting a detailed account of how transfer learning can be used in different contexts, we hope that it will become a widely used tool in the building modelling and simulation community.