Autonomous Building Energy Optimization -- Physical AI PoC

A closed-loop system where an open-source LLM (Ollama, qwen2.5:7b) controls a real DOE Commercial Reference Building (Small Office, Chicago) simulated in EnergyPlus, via MCP tools -- proving quantifiable energy savings while maintaining occupancy-aware thermal comfort.

Full architecture, prompt strategy, and design trade-offs (including a real bug we found and fixed in the comfort metric): docs/ARCHITECTURE.md

Project Layout

models/baseline.idf     Real DOE Small Office reference building (5 conditioned zones)[cite: 4]
models/Energy+.idd      IDD matching EnergyPlus 26.1.0
weather/chicago.epw     Chicago O'Hare TMY3 weather
tools/idf_tools.py      Core EnergyPlus bridge: run, parse metrics, edit setpoint schedules[cite: 4]
mcp_server.py           MCP server exposing those functions as LLM tools[cite: 5]
agent/llm_client.py     Ollama tool-calling client (+ offline mock for wiring tests)[cite: 2]
agent/agent_loop.py     The closed-loop orchestrator (system prompt + iteration control)[cite: 1]
dashboard/generate_dashboard.py   Baseline-vs-AI comparison chart[cite: 3]
runs/                   Simulation outputs, agent transcript, dashboard (generated)[cite: 3]
docs/ARCHITECTURE.md    System architecture document

Setup (Windows)

You've already done steps 1-4 if you're reading this after our chat -- listed here for completeness / for judges reproducing your results.

1. EnergyPlus 26.1.0 -- already installed at C:\EnergyPlusV26-1-0

2. Python packages

cd C:\building-ai-poc\project
venv\Scripts\activate
pip install -r requirements.txt

3. Point at your EnergyPlus install

set ENERGYPLUS_DIR=C:\EnergyPlusV26-1-0

4. Ollama + model -- already pulled qwen2.5:7b

Running the Closed Loop for Real

set LLM_BACKEND=ollama
set OLLAMA_MODEL=qwen2.5:7b
python agent\agent_loop.py

Watch it print every tool call and result live: reset -> baseline run -> propose a schedule -> simulate -> read real EnergyPlus feedback -> refine -> repeat, for up to 6 iterations. Full trace saved to runs\agent_transcript.json.

Then build the savings dashboard:

python dashboard\generate_dashboard.py

Writes runs\dashboard.png and runs\savings_summary.json.

Key Results & Performance

• Headline Savings: A verified 4.45% reduction in total energy consumption (1,107.52 kWh vs. 1,159.06 kWh baseline) achieved while maintaining zero thermal comfort violation hours across all conditioned building zones.

Testing Without Ollama (wiring check only)

set LLM_BACKEND=mock
python agent\agent_loop.py

Exercises the exact same MCP/EnergyPlus pipeline with a scripted (non-LLM) policy -- useful for confirming the plumbing works, but it is not a real optimizer and typically performs worse than baseline (we verified this ourselves -- see docs/ARCHITECTURE.md Section 4). Only LLM_BACKEND=ollama produces a real result to report.

Key Building Facts (verified against the actual IDF)

• 5 conditioned zones (Core + 4 Perimeter) + 1 unconditioned Attic, each conditioned zone on its own PSZ-AC system.

• Natural gas heating, electric DX cooling.

• All 5 zones share two setpoint schedules (HTGSETP_SCH, CLGSETP_SCH) -- the agent's control surface.

• Baseline schedule is the DOE reference's built-in weekday/weekend setback (15.6C/21C heating, 26.7C/24C cooling) -- already reasonably good, so beating it is a genuine optimization challenge, not a strawman.

• Comfort is checked only during occupied hours (via the building's real BLDG_OCC_SCH), so the baseline correctly shows 0 violations rather than being penalized for intentional night setback.