HVAC digital twin: PPO controller + LoRA explainer
Research prototype on synthetic Melbourne-style weather, VIC1-style prices and occupancy. Not validated on real buildings or real market data.
What is here
adapter/: LoRA adapter forQwen/Qwen2.5-3B-Instructthat turns an HVAC agent's decision (JSON facts from an RC thermal twin) into a short plain-English explanation.ppo_hvac.zip: PPO cooling-setpoint agent (Stable-Baselines3, 9-input observation, 1 action: cooling fraction).
Results (held-out synthetic days)
PPO vs a tuned thermostat (optimal start, P-control):
| Days | Cost/day | Comfort violation (°C·h/day) | Occupied steps outside band |
|---|---|---|---|
| Normal (20 days) | $8.22 vs $16.13 | 0.47 vs 0.35 | 5.1% vs 3.9% |
| Heatwave + price spike (40 days) | $13.58 vs $38.05 | 0.80 vs 4.18 | 8.5% vs 29.9% |
Explainer, share of answers that state the real cooling % and invent no numbers: tuned 100%, untuned base 3-11%.
Limitations (read these)
- The explainer was trained on rule-generated reference text, so it reproduces those rules. It does not reason about HVAC control.
- It cannot answer free-form questions: in a probe it ignored the question and sometimes contradicted itself, while never inventing a number. The faithfulness metric only checks that numbers appear in the input, not that statements are logically correct.
- Synthetic data only. Real AEMO prices and weather are untested.
- Downloads last month
- -