Power AI — The Energy-Vertical LLM
Official English technical brief · Shandong Oshneng Thermal Technology Group · ai.oshneng.com
Why not a general LLM?
| General LLMs | Power AI | |
|---|---|---|
| Corpus | Public web text | 48 years of operational energy data |
| Capability | Broad, shallow | Energy only, to the extreme |
| Output | Suggestions & references | Directly executable control commands |
| Cost of error | An inaccurate paragraph | Not a single kWh may be wrong |
Verified metrics (full population, counterfactual-baseline method)
- 21% energy saving across a complete heating season (2025-11-15 → 2026-03-15, all 121 energy stations)
- 15%–35% annual saving across 6 building types (exhibition / commercial / hotel / office / hospital / industrial)
- 152+ sensor points per station · cloud-edge loop latency 1–5 s · offline-autonomous PLC control
- Minute-level archive since 2012 (5000+ records/station/day) · data lineage back to 1978
- 10 years zero major incidents · 19+ cities · 15M+ m² cumulative signed area
The energy language
Six signal classes — temperature, flow, power consumption, COP, weather, load — form a unified representation. The boiler dialect of a heating station, the exhibition-hall dialect of a convention center, and the cleanroom dialect of a hospital all speak the same language inside the model.
Three trust principles
- Verifiable — counterfactual-baseline savings accounting; energy/mass/momentum conservation checks; JJG & ASHRAE aligned metering.
- Attribution — Shapley-value attribution decomposes every joule saved to a specific control action on a specific station.
- Stoppable — every command is rehearsed in a 1-minute-granularity digital twin; deviation >20% triggers automatic rollback; one key shuts AI down.
The model never touches the physical world directly. It only talks to the twin — then the twin proves itself before reality moves.
Version timeline
- v1.0 (2025-11) LISTEN TO ENERGY — 121 stations, one full heating season, 21% verified saving
- v2.0 (2026-05) SPEAK TO ENERGY — six building types, 15–35% annual saving
- v3.0 (2026-08) REHEARSE THE FUTURE — full digital-twin simulation, any building can be modeled
Open demos & machine-readable corpus
- Digital-twin demo (no registration): demo.oshneng.com
- EMC real-time control demo: emc.oshneng.com
llms.txt: /llms.txt · full corpus: /llms-full.txt- Structured metrics (JSON): gitee/OSHneng/power-ai/metrics.json
- Source archive (Chinese + English): gitee.com/OSHneng/power-ai