Over the past three months, WP3 has advanced the development of a physics-informed heat digital twin for residential buildings, aimed at improving heating demand prediction under sparse data conditions. Building on earlier work combining physics-based models, machine learning, and PINNs, recent efforts have moved beyond simplified resistance–capacitance (RC) representations toward a more complete whole-building physical description.
A new approach, Heat-Balance Physics-Informed Neural Networks (HB-PINNs), has been introduced, embedding full building heat balance equations within the learning framework. This enables inference of latent physical parameters and a more complete representation of thermal behaviour compared to traditional RC-based PINNs.
Systematic evaluations show HB-PINNs deliver superior accuracy and robustness across diverse scenarios, marking a significant step forward in whole-building heating demand modelling.
Over the next six months, WP3 plans to further explore the capability and practical relevance of physics-informed modelling within the Heat Digital Twin.
First, WP3 plans to explore new research directions in PINN-based modelling, for example through the use of real experimental and operational data, moving beyond simulation-only studies. This exploration aims to improve understanding of the applicability, robustness, and limitations of PINNs under realistic data conditions.
Second, WP3 is planning to develop a simple online heat prediction system focused on rapid estimation of heating demand. The initial concept involves breaking down gas consumption data to provide fast heat demand estimates, offering a lightweight and scalable approach for early-stage heat assessment. This activity is intended to serve as an initial demonstration of how digital twin methodologies could support near-real-time heat demand prediction and user-facing applications.
WP3 – Heat Digital Twin, is building a digital twin to enable real-time control and optimisation of residential heating. Accurate demand forecasting, led by Dr Meng Zhang, is central to its success.
Key Activities:
Simulation: Developed a detailed EnergyPlus model of a Stockport district to simulate thermal behaviour and validate forecasts.
Forecasting Models: Compared Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Transformer models; published findings on data partitioning impacts using London data.
Paper published in Energy and Buildings – Volume 344 – “Machine Learning-based regional cooling demand prediction with Optimized dataset partitioning“
Physics-Informed Neural Networks (PINNs): Integrated physical constraints into neural networks to improve accuracy and reduce data needs.
Benefit to SP Energy Networks:
This work equips SPEN with a high-resolution view of heating demand, essential for unlocking building flexibility. By combining simulation, forecasting, and PINNs, WP3 is creating a robust digital twin to support cost-effective demand-side management, peak reduction, and renewable integration — all while maintaining system security.
Next Steps
Refine PINNs, connect the twin to live pilot data, and explore flexibility services like peak shaving and frequency response. Assess network impacts of heat pump integration and develop control strategies.
