Evidence-gated digital twins for energy systems: field-validated heat-pump flexibility with online conformal bounds

A research project regarding the role of DT and AI in Energy Systems. In this research, I used data in 1 minute to analyze heat pumps in buildings. The research investigates how heat pumps with hot-water cylinders can provide flexible electricity demand by shifting heating to cheaper or more convenient periods without compromising household comfort or hot-water availability. It develops a layered digital twin that combines machine learning and physical modelling to predict heat-pump behaviour, estimate cylinder temperature and assess how reliably heating can be postponed. The findings show that purely data-driven predictions are not sufficiently reliable, while physically constrained models offer better performance. The study also highlights that monitoring cylinder temperature is more useful than estimating stored thermal energy when only limited sensor information is available. Although uncertainty-based forecasting methods improve the reliability of predictions, maintaining guarantees over extended periods remains challenging. Overall, the paper demonstrates that reliable flexibility requires more than predicting when a heat pump can be switched off; it requires accurately estimating stored heat and accounting for uncertainty over time. It also shows that electricity tariff structures play an important role in determining the financial value of shifting heating demand. This research submitted to Nature Communications on 3 November 2026.