Nonprofit journalism dedicated to creating a Human Age we actually want to live in.

Detailed home carbon audits are hard to come by. But AI can prioritize retrofits with little data

DAILY SCIENCE

Detailed home carbon audits are hard to come by. But AI can prioritize retrofits with little data

Researchers trained a deep learning model to make decisions based on widely available open-source building images
November 7, 2023

Let the best of Anthropocene come to you.

Artificial intelligence (AI) can identify so-called ‘hard-to-decarbonize’ houses with greater than 80% overall precision, according to a new study. The method could help pinpoint where energy-efficient retrofits are most needed and which elements of a home should be prioritized for replacement, researchers say.

Hard-to-decarbonize houses are properties that require difficult or expensive retrofits in order to make them green and sustainable. They are thought to represent one-quarter of homes globally, and be responsible for more than one-quarter of all housing-related carbon emissions.

Dealing with hard-to-decarbonize houses is therefore crucial for cities around the world to reach their net-zero carbon emissions goals. But there is no way to easily identify these houses and prioritize them for improvement.

In the past, public policy has tended to focus on decarbonizing buildings in general, or targeting specific high-emission technologies. But homes can be hard to decarbonize for multiple reasons.

“Policymakers need to know how many houses they have to decarbonize, but they often lack the resources to perform detail[ed] audits on every house. Our model can direct them to high priority houses, saving them precious time and resources,” study team member Ronita Bardhan of the University of Cambridge in the UK said in a statement.

 

Recommended Reading:
To upgrade energy efficiency, retrofit old buildings—with plants

 

In the new study, Bardhan and a graduate student, Maoran Sun, used a computer algorithm to identify hard-to-decarbonize properties quickly and cheaply. They fed open-source data into the model from a variety of sources, including Energy Performance Certificates, Google street view and aerial view images, satellite land surface temperature data, and building footprints from a UK-wide database.

Energy Performance Certificates are based on detailed inspections of properties built, sold, or rented in the UK since 2008. The certificates include data on features that make a home hard to decarbonize, such as solid walls or walls that are otherwise difficult to fill with insulation, flat roofs that cannot accommodate standard insulation, and traditional buildings constructed before 1919.

Now that the model has been trained using Energy Performance Certificates, it can identify hard-to-decarbonize houses even without such data, which are often unavailable, the researchers say.

The researchers gathered information about more than 1,300 homes in in Cambridge, UK. They trained the model using data from a subset of these homes, then used another subset of the homes to test the model.

“The overall precision for the model is 82%,” the researchers report in the journal Sustainable Cities and Society. If the model identifies a home as being hard to decarbonize, there is an 89% probability that it really is, and if it identifies a home as not being hard to decarbonize, there is a 71% probability that this is accurate.

The model is also able to find 84% of all non-hard-to-decarbonize houses, and 79% of hard-to-decarbonize ones.

Given image data, the model zeroes in on features like roofs and windows where the greatest loss of heat tends to occur, the researchers say. “The deep learning model makes decisions based on meaningful regions of the building facade image,” they write. “This proves the ability of machine learning to extract relevant information from image data.”

The model could help policymakers assess decarbonization needs in a given city, and develop policies for neighborhood- and even building-scale decarbonization, the researchers say.

The researchers also evaluated which data source is most important in securing accurate results. Street view images rose to the top of that analysis. Street view images are available for much of the globe and aerial images have even more widespread coverage, so this makes the method feasible even in low-resource countries or regions where building stock data are sparse.

The researchers are now training the model on data from other UK cities. They next plan to fold in additional data on energy use, poverty levels, and thermal imagery to improve its precision even further.

Source: Sun M. and R. Bardhan “Identifying Hard-to-Decarbonize houses from multi-source data in Cambridge, UK.” Sustainable Cities and Society 2023.

Image: ©Anthropocene Magazine.

Our work is available free of charge and advertising. We rely on readers like you to keep going. Donate Today

What to Read Next

Anthropocene Magazine Logo

Get the latest sustainability science delivered to your inbox every week

Newsletters

You have successfully signed up

Share This

Share This Article