WP3: Heat Sub Digital Twin, has undertaken a focused initiative to build a heating prediction model that leverages physics-based simulations and cutting-edge machine learning (ML) to deliver accurate, scalable, and adaptable insights.
Real-World Context: Data Collection from a Typical Site
The starting point was a typical housing site in Manchester, selected to represent real-life energy use in UK homes. Although data availability was limited, the team gathered essential building characteristics, including:
This baseline information provided the necessary parameters to develop a meaningful and localized energy model.
Detailed Building Simulation with DesignBuilder
To bridge the gap between available data and real-world behaviour, the team used DesignBuilder, building simulation tool. Using it, a detailed physical model of 73 residential units was created, incorporating internal construction details and thermal properties.
A key innovation was the simulation of occupant behaviour. By designing eight representative heating schedules, the team could reflect a broad range of household lifestyles and daily routines, enabling our model to account for variability in human-driven heating patterns.
Data Generation for High-Resolution Forecasting
With the physical model in place, the team generated hourly heating demand profiles for each unit. These profiles were created under different climate scenarios using Met Office Typical Meteorological Year (TMY) data and Solcast’s historical weather datasets, allowing the testing of model performance under a variety of real-world conditions.
Machine Learning Meets Physics-Based Modeling
To accelerate and simplify the heating demand prediction process, two types of deep learning models were applied:
These recurrent neural network models are designed to capture temporal patterns in sequential data, making them ideal for time series prediction like heating demand. Using Bayesian optimization, the models were fine-tuned to enhance performance and reduce training time.
Looking to the future, the team also began exploring Transformer-based architectures — a cutting-edge deep learning approach. In parallel, the team started investigating Physics-Informed Neural Networks (PINNs), which enable accurate predictions even under data-scarce conditions by embedding physical principles into the learning process.
Toward Smarter Heat Demand Management
The WP3 heating prediction model stands at the intersection of physics and machine learning. By combining detailed building simulations with data-driven prediction techniques, the team have developed a framework that can:
This work not only supports smarter energy planning and policy but also sets the foundation for dynamic control strategies and energy-efficient retrofitting in the residential sector.
