Awesome post. Who was the cranky researcher who developed the climate-controlled window system with the Linux box and python scripts? I want to do the same thing.
I want to put a 120mm fan in a small opening next to it that'll turn on when it's more than 3C cooler outside, and more than 24C inside.
Seems like an Arduino would do it.
Two analog temp switches will be far more reliable, easier, and adjustable.
(I get the fun with an Arduino, but it is hard to beat century old tech)
- Comfort is a function of multiple factors, not just the dry bulb temperature difference. It includes occupant metabolism, insulation from clothing, radiant heat exchange, air speed, and humidity. This is trivial to solve in the case of a single-person office (just give them control of the window), but more complex once you have multiple occupants with different needs, who are in different locations in the office. I wonder if there's value in using reinforcement learning to customize natural ventilation to optimize the comfort of multiple occupants.
- Predicting the temperature the next day, and using a fan to draw in cold night time air into your HVAC ducts when a hot morning/afternoon is expected. This way you can have a reservoir off cool air you can access the next day.
- Once you have multiple rooms, and a tall building, there's multiple effects that are worth attempting to coordinate. For example you could coordinate the ventilation of the building air with the stack effect and wind-driven air through the windows, or coordinate the cross-ventilation horizontally through multiple rooms. This would not just modulate temperatures, but also satisfy indoor air quality needs traditionally achieved through the HVAC.
Typically a lot of fine-tuning of natural ventilation isn't done because the actual energy/comfort impact for one of these ideas is small relative to complexity of building the system, and the large uncertainty associated with the dynamics of airflow. But my general hypothesis is that with the integration of cheap prediction or RL methods, we can reduce that uncertainty which would permit a more long-tail approach where we can combine multiple small improvements.
Not checked on paper (especially the first) but on principle both should work and also would be extremely cheap.
I don't think it's overkill within the context of his full work. Based on my brief skimming, he's coupling the window with a personal device to modulate his environment through reinforcement learning.
https://resenv.media.mit.edu/#Projects#personalized-building...
He published his thesis 2010, which aligns with when Ethan arrived at his office, and if you watch the video at that link, you can see the automated window control system at 1:47.
(Also the video is kind of hilarious and awesome in its own right).