projects / gridsight
GridSight
From transformer telemetry to maintenance actions: simulated SCADA readings, IEEE C57.91 and C57.104 threshold rules, and env-gated LLM recommendations.
- 8 Simulated substation assets Telemetry simulator
- 5 s Telemetry cadence over SSE Telemetry simulator
- 5 Hand-authored fixtures Recommendation fixtures
Overview
Problem
Transformer telemetry is plentiful. Deciding what to do about it is the hard part.
Approach
An Angular 18 dashboard over simulated SCADA telemetry on Server-Sent Events, with threshold rules from IEEE C57.91 and C57.104 turning readings into alerts.
Result
Recommendations come from schema-enforced tool use when an LLM key is set and from 5 hand-authored fixtures when it is not, through the same JSON contract.
Same JSON contract, two paths. The deployed demo never holds the key.
On the env-gated LLM
Decisions · why this stack
-
Env-gated LLM
With a key, the route calls the model with a tool-use schema that enforces the response shape. Without one it serves a fixture matched to the alert rule.
-
Standards as rules
Thresholds follow IEEE C57.91 (oil 95 and 105 °C, winding 130 °C) and IEEE C57.104 for dissolved gas, so every alert points at a rule a utility engineer recognises.
-
Pure simulator and alert engine
Both are pure functions over one reading type, and vitest covers the threshold matrix, including compound failures that raise several alerts.
-
Stateless and portable
Env-driven config and no server state mean the Vercel demo moves to Azure App Service with a Bicep file and a connection string.
Stack
- Angular 18
- TypeScript
- RxJS
- Node.js 20
- Express
- Zod
- SQLite
- OpenAPI 3.1
- LLM APIs
- vitest
- Docker
- Vercel
Demo
Recorded snapshot: Råsunda T11 TR-RASUNDA-11
A reading captured mid-fault. Same data shape that flows over the SSE stream and into the AI recommendation endpoint.
Telemetry reading
- Oil temp
- 102.4 °C alarm 105 °C
- Winding temp
- 118.7 °C warn 110 °C
- Load factor
- 93% warn 100%
- Voltage
- 0.984 pu
- DGA H₂
- 64 ppm
- DGA C₂H₂
- 0.60 ppm
AI recommendation
demo fixtureTop-oil temperature has crossed the IEEE C57.91 alarm threshold (105 °C) while load factor remained near rated capacity. The combination indicates the cooling system is under-performing relative to thermal demand, most commonly a fouled radiator surface, degraded fan operation, or low oil level.
- now Reduce loading to ≤ 70% of rated kVA until oil temperature returns below 90 °C.
Sustained operation above 105 °C accelerates insulation aging by ~2× per 6 °C (Arrhenius). Load reduction is the only corrective lever available remotely.
- soon Dispatch field crew to verify cooling fan status, radiator surface cleanliness, and oil level.
Three of the four most common causes of high-oil-temp events are cooling-system mechanical and can be diagnosed in a single 30-minute site visit.
- planned Schedule infrared thermography sweep of all radiator panels during the next available outage window.
Thermography reveals partial blockages and uneven flow that are not detectable from oil-temp telemetry alone.
The integration code in api/src/domain/recommender.ts calls the model with a
tool-use schema enforcing this exact shape. The deployed demo doesn't ship the API key. It
serves five hand-authored fixtures matched to the alert rule. Flip LLM_API_KEY in env
to switch paths.