As part of the AI for Climate Change hackathon, we wanted to develop an application of reinforcement learning that could help households in Montreal reduce their carbon footprint. This wasn't obvious, since so much of our household energy is low-carbon electricity from hydropower. However, residential heating still predominantly relies on fossil fuels while accounting for the lion's share of Canadians' energy consumption: making it an ideal problem to tackle for this hackathon. In fact, natural gas for residential space heating in Canada produced ~24,000,000 tC02 in 2016.
What it does
To address this, we built a simple 2D heat transfer model to simulate houses in Montreal's winter, and trained a reinforcement learning agent to optimize heating schedules so as to reduce waste and maximize the comfort of simulated occupants with their own schedule and temperature preferences.
Challenges I ran into
Difficulties with heat transfer modelling, reinforcement learning agent stability and reward shaping.
What's next for It's getting hot in here
Investigate use of reinforcement learning to encourage humans into heating their homes less (1% savings per degree per 8 hours!)