Optimus: Autonomous EnergyPlus Building Controller

Problem

Building HVAC control typically relies on fixed schedules or reactive thermostats. The challenge was to create an autonomous controller that observes live simulation state, plans bounded supervisory actions via local LLM or deterministic fallback, validates safety (PMV/ITE constraints), and injects approved schedule overrides — all with full auditability.

Architecture

EnergyPlus baseline → Interface discovery → Live PyEnergyPlus callback loop → Bounded LLM planner (Ollama/Qwen) or deterministic fallback → Hard safety validator (PMV for comfort, ITE inlet temp for data centers) → One-hour bounded schedule override with automatic native restoration → EnergyPlus SQL KPI comparison + JSONL audit trail.

Challenges

• Integrated EnergyPlus Python API for live state observation
• Designed bounded LLM planning with deterministic fallback for offline operation
• Implemented hard safety validation: PMV/temperature guards (comfort), ITE inlet temperature (data centers)
• Built one-hour bounded cooling relaxation with automatic native schedule restoration
• Delivered Docker/Streamlit deployment with JSONL audit trail and EnergyPlus SQL-derived KPI reports
• Handled unsupported interfaces gracefully (baseline-only with explanation)

Benchmarks

Validated on EnergyPlus models: comfort buildings (PMV safety) and data centers (ITE inlet headroom). Frozen v1 controller produces auditable comparison reports (baseline vs AI energy, peak, comfort) with per-action validation reasons.

Lessons Learned

• Physical-AI systems require hard safety boundaries, not just ML optimization
• Deterministic fallback enables evaluation without external LLM dependencies
• EnergyPlus SQL + JSONL audit trail provides full traceability for building operators
• Bounded one-hour overrides with automatic restoration prevent runaway control
• Interface discovery determines actuation capability per model — unsupported models run baseline-only