How we changed our roof and cut 1.5 tons of CO2e

Factory

In 2018, my wife and I bought an old shoe factory in Paris. We entirely demolished and revamped the space.

We were conscious of the importance of energy insulation. We used hemp-based insulation for walls and a double-glass window for the facade. One thing we did not replace was the roof. We loved the industrial-style "verrière" and it was an expensive thing to do back then.

Roof

But soon after, we realized it was a bad decision. Insulating the entire flat except the roof led to severe thermal bridging. During winter, as exterior temperature fell, the glass roof was exposed to cold air from the outside and hot air from the inside leading to condensation.

In July 2021 we finally were able to replace our old single-glass roof. After one year looking for a capable artisan, we installed a double glass ceiling roof with thermal break steel from Jansen. The difference in confort was mind boggling. It was like moving in a new flat.

Data

A question I asked myself was: how much energy and CO2 are we saving?

In February 2019, I acquired a Netatmo thermostat that controls the heating system. The device collects temperature, boiler activity on a 30min step. It has an API we can query. See code source in Python here.

In July 2020, I also got installed Gazpar, the smart gas meter from GRDF (French gas operator). While far from perfect (the data portal is often down) it allows me to get an accurate daily measure of my energy consumption.

Below is a line graph plotting internal temperature, exterior temperature (°C) and energy in Kwh between October 2020 and April 2021. Few notes:

• It's pretty chilly in our flat! Average interior temperature is 17.2 °C.
• The colder outside the more energy we consume (duh).

Modeling

In order to measure how much energy we are saving, we need a model. We could for instance compare our energy usage with the previous year. But since exterior temperature varies or we could have been away for our house, comparing with the previous year is imprecise.

A good model would be like a "twin" flat. It would show how much energy we would have consumed had we kept the old roof.

To do so I run a good ol' linear regression with the following inputs:

• Internal temperature (°C)
• Exterior temperature (°C)
• Whether or not the flat is occupied (boolean)

The model is trained from October 2020 to February 2021 and has a nice fit! We are able to capture 90% of the daily energy variance. The graph below displays the model's predictions (dotted line) and the true energy consumed (solid line). Notice that the model "holds well" after the training period.