Reading substation drawings into standard data

Utilities hold decades of single-line diagrams as images and PDFs. This prototype turns one into IEC 61850 SCL, the standard data format for substations, and measures how well each vision model does it. A model reads the drawing; plain code checks every answer against the standards before anything counts.

The approach

1. An assembled ontology

The vocabulary comes from standards organizations, not from us: IEC 61850 SCL for the substation, IEC CIM for what SCL doesn't cover, IEC 81346 and IEEE C37.2 for equipment classes and relay numbers.

What a drawing shows but no standard carries goes into a reviewed gap register, and the ontology grows one reviewed change at a time.

2. A checking harness

The model only fills in a fixed form. Code checks the form, checks that the wiring is consistent, sends errors back once, builds the SCL file and validates it against IEC's own schema.

How it works →

3. Measured, then reviewed

Every run is scored item by item against a hand-checked answer key. In the full tool, an expert reviews each drawing, fixes what's wrong and signs off; each signed-off drawing becomes a new answer key.

Results in detail →

Results so far

Loading results…

Charts, findings, time and cost → · Inspect every run → · How these models compare with other vision models →

Status: a prototype

The accuracy figures here are early and still work in progress. They come from one synthetic drawing in two renderings, a few runs per model and one answer key. Before drawing conclusions we need more sample runs, harder and real drawings, and better detection of errors.

Next steps

Harder drawings: a real, published drawing scored against a provisional answer key, and degraded copies (lower resolution, noise, rotation) of synthetic drawings with exact keys.
More runs per model, so differences between the strongest models can be measured rather than tied.
Better error detection where models fail most: wiring (which ends meet) and protection relays. An agent that re-reads doubtful regions with a second model is built; its effect on accuracy is not measured yet.
A formal ontology export to IEC CIM in RDF (CGMES), so the assembled ontology can be used by standard tools. Planned, not built.

For research partners

A drawing generator with exact answer keys: five substation layouts, degraded copies, scored per equipment type (precision and recall).
An evaluation that separates the model's errors from the pipeline's: nothing our code drops goes unreported.
A gap register of what IEC 61850 SCL doesn't cover for drawings, with the standard each gap could use.
Open questions: how far synthetic drawings predict real ones; whether checking rules can serve as a reward signal for training smaller models (outlined on How it works).