BuildingsBench: Supporting Building Decarbonization (Text Version)

This is the text version of the video BuildingsBench: Supporting Building Decarbonization.

[Music plays]

[Narrator speaks]

Residential and commercial buildings are responsible for nearly 30% of consumed energy, or load, and 27% of greenhouse gas emissions globally. Decarbonization efforts for addressing climate change include gradually transitioning buildings to cleaner sources of electricity and managing how buildings use energy more efficiently.

[Text on screen: Residential and commercial buildings are responsible for ~30% of consumed energy. Residential and commercial buildings are responsible for ~27% of GHG emissions. ]

Introducing BuildingsBench, a new dataset and benchmark developed by NREL for artificial intelligence—or AI—researchers looking to evaluate the efficacy of their approaches on building decarbonization challenges.

[Text on screen: Visualization of BuildingsBench Dataset]

AI researchers can use this dataset to develop and evaluate AI that supplements efforts to decarbonize buildings. BuildingsBench currently challenges AI researchers to develop methods for short-term load forecasting or STLF.

Think of load forecasting like weather forecasting, but with energy usage in place of temperature.

The STLF challenge in BuildingsBench involves predicting a time series of future hourly building load up to a day ahead. This has multiple practical applications. For one example, accurate day-ahead load forecasts can help utilities effectively manage a dynamic energy market that aims to incentivize homeowners to reduce their energy usage during peak periods.

Current AI approaches for STLF are hindered by a lack of data. BuildingsBench provides the first large-scale pretraining data set and benchmark to help researchers study time series foundation models for STLF. Foundation models are trained on broad data, typically learn from self-supervision, and can be customized for a wide range of downstream contexts. This work explores how to train foundation models on time series data produced by physics-based simulation models that are representative of the U.S. building stock to learn shared patterns of energy consumption across many buildings. We then evaluate these foundation models on load data from real buildings to illustrate how AI can help contribute to a challenging problem like building STLF.

[Text on screen: Visualization of BuildingsBench Dataset]

The ability to forecast building loads with AI will enable us to deploy new technologies to buildings faster.

NREL's buildings research transforms energy through building science and integration. BuildingsBench can add to the legacy of this work by helping to bring the buildings and AI research communities closer. There's more work to do—and the BuildingsBench team encourages more research on AI and foundation models to help tackle building decarbonization challenges.

Stay tuned for developments on BuildingsBench and head to www.nrel.gov to learn more about NREL researchers' work in AI and buildings.

[Text on screen: Learn more about BuildingsBench by visiting www.nrel.gov]

[Narration ends, music stops.]