A brief filler post messing around with data on Yale's energy consumption

I originally planned to reserve these write-ups for bigger, more fleshed-out projects. Turns out those take a while, and grinding on the same thing for too long can become a bit much. So here's something in the meantime.

I was poking around the web for project ideas and came across Yale's Facility Energy Usage Explorer tool. The site is a bit outdated, clunky to use, and only lets you export monthly data from July 2020 onwards, but the underlying information was quite unique and I couldn't find a clean, structured dataset aggregating it either on Kaggle or GitHub.

Yale Facility Energy Usage Explorer tool — Screenshot of the landing page for the Yale Facility Energy Usage Explorer.

Inspecting the fetch/XHR network requests revealed a few hidden endpoints which included params for filtering on date ranges. This allowed me to get monthly energy consumption data by commodity for various buildings on campus starting from July 2002 (an extra 18 years of history) along with some metadata on each facility like coordinates and building type.

As hinted at in the post description, I had been meaning to explore DuckDB and Bokeh in general for data analysis and visualization. You can find some of the plots I made below:

Energy usage by commodity type over time across all Yale facilities. There is some very clear seasonality in the summer and winter months. Also some weird spikes in 2013 and 2014.

Year-over-year energy usage across all Yale facilities — Year-over-year change in overall consumption across all Yale facilities.

Share of energy consumption by commodity type. I have no idea what "Gas 2 Equivalent" is but it seems to have taken a significant amount of share.

Energy intensity by commodity type — Overall energy consumption per square foot. As a proxy for efficiency, this has been pretty much flat over recent years. Not surprised given how old a lot of the buildings are.

Map of the top 20 energy consuming buildings on Yale's campus. Admittedly, this is a pretty bad visualization lol I probably should have restricted to central campus but whatever.

Overall, I found bokeh to be pretty clean and flexible, but I definitely had to vibecode quite a bit having used matplotlib, seaborn, and plotly all these years. I'm not sure if it's convinced me to make the switch, but then again I barely scratched the surface in this mini exploration. On the other hand, I found duckdb to be super intuitive to use and it just felt natural to query my CSVs directly using SQL, so I will definitely try to use it in future projects.

As always, the raw and cleaned data as well as my slop code can be found on my GitHub.