Modularized neural network incorporating physical priors for future building energy modeling

Area of Science:

• Building Science
• Artificial Intelligence
• Sustainable Energy

Background:

• Traditional building energy modeling (BEM) is complex, requiring extensive data, expertise, and time for each building.
• This complexity limits the scalability of BEM for large-scale applications like urban planning and energy management.

Purpose of the Study:

• To develop a scalable and efficient building energy modeling approach.
• To overcome the limitations of traditional BEM by integrating physical principles with data-driven methods.

Main Methods:

• Developed a modularized neural network (ModNN) that incorporates physical priors, including heat balance equations and physically consistent constraints.
• Designed a data-driven modular structure enabling model sharing and inheritance for multiple-building applications.

Main Results:

• Demonstrated the scalability and effectiveness of ModNN across four diverse applications: load prediction, indoor environment modeling, building retrofitting, and energy optimization.
• The ModNN approach significantly reduces the need for extensive case-by-case calibration and expert knowledge.

Conclusions:

• ModNN offers a novel framework for large-scale building energy modeling by combining physical insights with data-driven flexibility.
• This facilitates widespread adoption of advanced BEM for energy management, resilient retrofits, and building-to-grid integration, contributing to sustainable urbanization.