Power Transformer Diagnostic Evidence Models for Agentic AI

Agentic AI for transformer asset performance management is only credible if it can reason over an evidence model, not a pile of files. A transformer recommendation should answer: which asset, which component, which observation, which method, which standard context, which timestamp, and which confidence level?

That evidence contract is the difference between a reviewable reliability workflow and a black-box score.

Start with asset hierarchy

Transformer evidence should attach to the component that can fail. A useful hierarchy includes fleet, site, transformer, tank, winding, bushing, tap changer, cooling system, sensors, and protection context. Utility data standards such as IEC 61850 and CIM-oriented work such as the PNNL guide to IEC CIM are useful references for thinking about consistent asset identity and interoperability.

For GridAPM, the practical rule is simple: every record must know what it belongs to. A DGA result belongs to a transformer and oil system. A bushing test belongs to a component. An SFRA trace belongs to a winding configuration, tap setting, and measurement setup. Without hierarchy, AI has no reliable place to attach evidence.

Provenance before prediction

AI reasoning is fragile if the input records are not trusted. A transformer evidence model should capture:

• Source system or file.
• Measurement method.
• Timestamp and sample time.
• Unit and normalization.
• Calibration or test setup notes.
• Data quality flags.
• Transformation history.
• Reviewer or system that created the record.

The NIST AI Risk Management Framework emphasizes governance, transparency, and accountability themes that map directly to this requirement. In transformer APM, provenance is not paperwork. It is the foundation for deciding whether a recommendation deserves attention.

Model the diagnostic channels

A transformer evidence model should represent each diagnostic channel on its own terms:

• DGA: gases, ratios, generation rates, sample source, oil processing context, and interpretation history. See IEEE C57.104 and IEC 60599.
• PRPD and partial discharge: apparent charge, phase reference, measurement circuit, sensor type, noise notes, and pattern family. See IEC 60270.
• SFRA: trace file, instrument, lead placement, grounding, tap position, baseline, and comparison event. See IEEE C57.149.
• Thermal loading: load, ambient, cooling state, top-oil, hot-spot estimate, and aging context. See IEC 60076-7.
• Monitoring equipment: sensor application, data quality, communication, and risk-benefit context. See IEEE C57.143.

The agent should not flatten these into one number too early.

Contradictions are signals

In a well-modeled system, contradictions are not noise to be averaged away. They are diagnostic prompts. Rising combustible gases with normal thermal history may deserve a different review than thermal aging without chemical markers. A PRPD pattern without DGA evolution may suggest measurement context or a developing insulation issue. An SFRA shift after a through-fault may change the confidence of an otherwise moderate risk score.

Agentic APM should ask follow-up questions when the evidence conflicts. That is more valuable than forcing every record into a single blended health score.

The agentic workflow

Once evidence is modeled, AI agents can perform bounded work:

1. Retrieve records attached to the asset and affected component.
2. Check provenance and stale data.
3. Compare observations with baselines, standards context, and prior decisions.
4. Form hypotheses and uncertainty notes.
5. Request missing records or retests.
6. Draft maintenance options for engineer review.

This links directly to DGA generation-rate analysis, PRPD review quality, and SFRA winding movement analysis.

A practical evidence contract

GridAPM’s transformer diagnostic evidence contract can be summarized as:

asset/component ID + observation + method + provenance + standards context + confidence + interpretation + next action + audit trail.

That contract helps transformer teams see exactly how an AI-assisted recommendation was formed. The winning claim is not that AI predicts every failure. The stronger claim is that AI can make transformer diagnostic reasoning more traceable, cross-modal, and reviewable.