Smart Grid Transformer Integration & Automated Scenario Generation Pipeline

[Rohith-Kanathur]

Overview

This PR delivers two major contributions to AssetOpsBench:

1. A new Smart Grid Transformer asset class with four FMSR diagnostic tools grounded in IEC standards
2. A multi-phase automated scenario generation pipeline (ScenarioGeneratorAgent) that scales benchmark creation to new asset classes without manual authoring

---

1. Smart Grid Transformer Integration

CouchDB Data

• Added Smart Grid Transformer asset documents to CouchDB, including dissolved gas analysis (DGA) readings, winding temperature profiles, and electrical load profiles as new data sources.

Four New FMSR Tools

Tool	Standard	Description
predict_health_index	—	Supervised regression model trained on the Mendeley Transformer Health Dataset; predicts a 0–100 health score across five condition bands (Very Good → Very Poor)
interpret_dga	IEC 60599	Applies the Rogers Ratio method to classify transformer fault type from dissolved gas concentrations with confidence rating
assess_winding_temperature	IEC 60076-7	Computes insulation ageing rate and thermal risk from winding/oil temperature inputs
assess_load_profile	IEC 60076-7	Derives three-phase apparent load and classifies loading status against cyclic loading limits

Tests

• Added unit tests for all four FMSR tools covering valid inputs, edge cases, and schema compliance.

---

2. Scenario Generation Pipeline (ScenarioGeneratorAgent)

A three-phase automated pipeline that generates physically plausible, causally consistent, and tool-reachable benchmark scenarios for any onboarded asset class.

Phase 1: Asset Profiling

Discovers live asset instances, sensors, and failure mappings from CouchDB, retrieves and synthesizes domain literature from ArXiv or Semantic Scholar, and merges everything with MCP tool schemas into a single structured AssetProfile that grounds all downstream generation.

Phase 2: Budget Allocation

Distributes the total scenario count across focusses (iot, fmsr, tsfm, wo, vibration, multiagent) proportionally to the asset's available data modalities and tool coverage, with multiagent capped at 75% of total budget to preserve lane diversity.

Phase 3: Scenario Generation & Validation

Generates per-focus scenarios conditioned on the asset profile, runs each candidate through an LLM-based repair step.

Output

Each run produces a timestamped directory with scenarios.json. Each scenario object contains an id, type, text, category, and characteristic_form.

CLI

# Closed-form generation
uv run python -m scenarios.generator "Transformer" --num-scenarios 50

# Grounded open-form with live CouchDB data
uv run python -m scenarios.generator "Transformer" --data-in-couchdb --num-scenarios 50

[DhavalRepo18]

@Rohith-Kanathur @Sagar-CK

• Standardize LiteLLM params: thinking toggle, max_tokens, and generation settings #304
• Add smart grid transformer data file and load into CouchDB (Git LFS-managed) #303

We request a separate PR.

[DhavalRepo18]

Pls See Comment

[ShuxinLin]

Hi @Sagar-CK @Rohith-Kanathur I would like to know why only 5 data points are in the bulk_docs_transformer.json

[Rohith-Kanathur]

Hi @ShuxinLin

The current bulk_docs_transformer.json only contains 5 data points because it was initially added as lightweight mock/test data to verify if the CouchDB ingestion was happening correctly.

The intent wasn’t to model a full production level smart grid dataset but rather to validate the loader + CouchDB integration flow first. We can definitely expand the dataset with additional transformer documents.

I was thinking about using this dataset since it contains around 400 transformer data points: https://data.mendeley.com/datasets/rz75w3fkxy/1

Please let me know what you think.

Thank you.

[DhavalRepo18]

@Rohith-Kanathur You needed these sample data for what? These PR is for generating scenarios? or testing Scenario?
Any advise on how to collect the data and etc will be useful.

[Rohith-Kanathur]

@DhavalRepo18 I was talking about this github issue: #303
(I was NOT referring to this PR which is about scenario generation)
To load smart grid transformer data into couch db, we will need a transformer dataset (in json or csv format).
This is similar to how it is done for the Chiller asset.
I suggested using a transformer dataset I found online (https://data.mendeley.com/datasets/rz75w3fkxy/1) since there are good number of data points and it contains relevant sensor readings.