Indoor Temperature Prediction for Residential Heat Pumps: A Physics-Informed Machine Learning Approach

Abstract

With space heating accounting for a large proportion of UK energy demand, air-source heat pumps have become central to building decarbonisation. Successful deployment depends onaccurate prediction of indoor temperature, which is critical for maintaining thermal comfort. This study presents a novel, hybrid approach for temperature prediction, embedding the interpretability of a Resistance–Capacitance network with a Long Short-Term Memory neural network to capture unmodelled dynamics.Archetypes represented building geometry and envelope characteristics for each housing typology, supplying baseline parameters that were subsequently calibrated for individual dwellings using variational inference. The calibrated Resistance–Capacitance model achieved acceptable accuracy (RMSE = 1.38 °C), while the hybrid model reduced error further (RMSE = 0.44 °C).Clustering households and applying transfer learning reduced error by an additional 30%, highlighting the potential of physics-informed models to enable demand-response control of residential heating.

More information

Identifiers

ORCID for Stephanie Gauthier: orcid.org/0000-0002-1720-1736

Catalogue record

Export record

Altmetrics