A commercial floor in Manhattan has four ways to keep itself comfortable: an electric heat pump (efficient when it's warm outside, mediocre below freezing), a gas boiler (constant efficiency, gas commodity price), Con Edison district steam (the world's largest urban steam loop â premium-priced), and an electric chiller. Electricity from NYISO Zone J swings from ~$30/MWh overnight to over $200/MWh at the 5 pm peak. Search for a control policy (linear in zone state) that minimises the daily total energy bill plus comfort penalty, averaged across four real-weather days from a NYC year. 16-D continuous problem.
Pick a policy and watch it dispatch across all four weather days.
16-D policy, each evaluation simulates four 24-hour days.
Each row is the best policy a given algorithm found (lower cost = better).
| Algorithm | $/day | Policies | Detail |
|---|---|---|---|
| â no runs yet â | |||
The zone is a single-RC thermal model â capacitance C = 10 kBtu/°F, overall envelope conductance UA = 0.8 kBtu/h/°F, so the natural time constant Ď â 12.5 h. Stepped hourly, with setpoint 72 °F and a comfort band of 68â76 °F enforced during occupied hours (8 am â 6 pm). The outside temperature follows a smooth diurnal curve with a daily peak around 2 pm.
Four dispatchable sources, with cost structures that pull in different directions:
Each scenario simulates a 24-hour day from a different point of a NYC year â a January cold snap, a mild winter day, a mild summer day, and a July heat wave. The score is the mean of total daily cost (energy + comfort penalty at $1.5/°F-h during occupancy) across all four days.
The 16-D decision variable is the parameters of a linear control policy: four sources, each driven by a sigmoid of (bias + zone error + outdoor demand + electricity-price excursion). The optimum trades off when to use each source against the current state â a "good thermostat" gets to ~$70/day, but the optimum gets close to ~$30/day by riding the heat pump when it's warm and the LMP is cheap, switching to the boiler during the 4â7 pm peak, and barely touching steam.
Single-zone RC simplification of an EnergyPlus-style building model. Real BMS optimization adds latent loads, weather forecasting, occupancy stochasticity, multi-zone coupling, and demand-response tariff signals â but the core structure (find a policy that minimises an expected cost over a distribution of days) is the same.
If your hyper-parameter searches are heating the Earth, drop this in Cursor or Claude:
Read https://raw.githubusercontent.com/microprediction/humpday/main/SKILL.md and create a project skill from it.