YES

Inspiration

Buildings are a major source of global energy use and carbon emissions, and organisations can only cut that waste if they know which buildings to fix. The WiDS Datathon 2022 gave us about 75,000 real building-year records, and we wanted to answer two questions: can we predict a building's energy use, and what actually drives it?

What it does

Given a building's characteristics and local weather, our model predicts its Site EUI (energy used per square foot per year) for a year it has never seen.

How we built it

We built everything in MATLAB Online.

1. Cleaning: handled heavily missing weather columns and fake zero values in year_built, and typed the categorical columns.
2. Exploring: we plotted EUI against climate, building type, build year and Energy Star rating, and turned what we saw into features (facility groups, construction eras, Energy Star bands, and size and rating relative to building type).

3. Validating: we held out the latest training year, since the test set is a later year, and scored every model with RMSE:

4. Modelling: guessing the average gave an RMSE of 57.2. A single tree gave 49.7, a tuned tree 46.6, and boosting 46.5. Our best was bagged trees, which average many trees:

This reached about 43.7 in validation, roughly 24% lower error. The final model used 100 trees, trained on all the training data.

What we learned

• What a building is matters more than where it is. Buildings in the same climate ranged from about 50 to over 900 EUI.
• Energy Star only makes sense within a building type. At first it looked odd overall, but inside each type, higher ratings meant lower EUI. For example, offices went from about 231 at ratings under 10 to about 37 at 95+. Buildings rated under 10 used about 2 to 6 times the energy of 95+ peers in the same type.
• Our hand-made features barely helped, because tree models can already split on the raw columns. Tuning and averaging helped far more.

Challenges

• MATLAB Online froze on our first large boosted model, so we scaled up gradually.
• Missing toolbox functions and mismatched data files cost us time, so we added checks to make sure everyone loaded the same data.
• The model is weakest on rare, high-energy buildings. Labs and industrial sites are under 2% of validation buildings but cause about 12% of the error.

What's next

Gather more data on rare building types, and test whether predictions guide retrofits that actually save energy.

Built With

• matlab