Observe
14 inputs describe temperatures, stored energy, weather, prices and the known schedule.
SustainaBuildAI ENERGY LAB / PERTHEnergy decisions.
Made intelligent.
An office needs to stay comfortable. Electricity costs money. Solar comes and goes. Can a learned controller find a better balance?
ONE BUILDING. CONNECTED DECISIONS.
Illustrative system diagram. Open the lab to see the actual simulated actions and power values.
One hour at a time.
The whole day in mind.
Cooling now affects comfort later. Using the battery now changes what remains for the evening. This is a sequential decision problem with competing objectives.
Decide
Choose from 15 actions: three cooling levels combined with five battery requests.
Respond
A thermal model and physical battery constraints produce the next building state.
Evaluate
Measure purchase cost, occupied discomfort, grid exceedance and battery use.
DQN
Learns a future-return score for each action.
REINFORCE
Improves action probabilities using complete-day experience.
A2C
Learns a policy with help from a value-estimating critic.
WHAT THE RESULTS SAYGood research
Good research
shows the trade-offs.
On the 121 held-out test days, all three agents beat the rule-based controller by 25 to 34%. DQN matches the two-step lookahead planner (17.31 against 17.30) and has the lowest mean electricity bill, REINFORCE the lowest mean discomfort, and the planner the lowest grid peak.
DQN is also best in the heatwave test, but the planner is ahead when the office hours shift or solar is halved, and DQN uses the battery about twice as much as the planner. The project does not claim that RL is universally superior or that these are measured real-building savings.
STANDARD TEST / MEAN DAILY OBJECTIVE ↓
Rule-based
26.05
Lookahead
17.30
DQN
17.31
REINFORCE
19.51
A2C
17.50
121 test days · three final seeds per RL agent. Combined penalty score, not the electricity bill. Source: saved executed Colab results.
Watch. Compare. Experiment.
Explore all 213 validation and test days with all five controllers, using the networks from our submitted final run. Try the heatwave, shifted-schedule and low-solar scenarios, or take control of the building yourself hour by hour. Follow the guided tour, inspect hourly decisions and open the training evidence.
Launch SustainaBuild ↗The simulator, both conventional controllers and the three trained networks run in your browser, ported from our Python code and checked against it (identical actions on 6,390 controller-days). The building is simulated; there is no live building connection.