Engineering software for protection and transformer reliability

Request pilot

Transformer asset performance management

Agentic AI forpower transformer reliability

GridAPM turns DGA, partial discharge, and inspection records into engineer-approved maintenance decisions — on your infrastructure, with every recommendation traceable to its evidence.

Request a pilot See how it works
  • IEEE C57.104
  • IEC 60599
  • IEC 60270
  • IEEE C57.149
  • CIGRE
  • NIST AI RMF
  • Local-first — runs on your workstation

TX-47 · 230/69 kV

Load 78 % OK

Agent note

  • H2 trend exceeds IEEE C57.104 condition 2; correlated with load steps. Recommend confirmatory sample.
  • PD amplitude on the B-phase bushing tracks load steps, not humidity — pattern reads as surface discharge. Drafting an inspection task.
  • Hot-spot estimate stays within the IEEE C57.91 loading guide at present load; no accelerated ageing flagged this cycle.

Engineer review pending

Two products

One company, two engineering products

GridAPM builds software for the two places a substation engineer needs evidence they can defend: the protective relays that clear a fault, and the power transformers that must not fail. Separate products, separate licences, the same discipline — the AI drafts, a named engineer signs.

  • Relay testing

    ProtectionAI

    Plan, run and document protective-relay tests, with a tool-using AI copilot that reads your relay manuals and drafts the report.

    Explore ProtectionAI

  • Transformer APM

    AgenticGrid Pro

    Turn transformer fleet evidence — DGA, partial discharge, SFRA, thermal — into an engineer-approved work package with a retrievable audit trail.

    Explore AgenticGrid Pro

Compare both products

Time-based maintenance

The status quo runs on the calendar

Most transformer fleets are still maintained by the calendar, not by their condition.

Fixed intervals miss real faults
A unit that starts gassing in month three waits nine more months for its scheduled inspection.
Evidence sits in silos
DGA lab results, PRPD captures, SFRA sweeps, and inspection notes live in systems nobody has time to correlate.
Fleets at design life
US large power transformers average about 40 years in service — the typical design life — and more than 70% are over 25 years old.¹
Replacement takes years
Lead times for a large power transformer now commonly run 36 months — up to 60 — versus under a year before the pandemic.¹

1 U.S. Department of Energy, Large Power Transformer Resilience — Report to Congress (2024).

TBM to CBM

Schedule to evidence

Condition-based maintenance replaces fixed intervals with interventions the evidence actually calls for.

Every 12 months

Fixed calendar intervals: healthy units are opened anyway, and developing faults wait for the next slot.

When the evidence says so

Condition-triggered interventions: a DGA trend shift or PD onset opens one targeted work package.

Wasted work Fault missed DGA trend shift PD onset Targeted intervention
25–30%

reduction in maintenance costs

35–45%

reduction in downtime

8–12%

additional savings over preventive-only maintenance

US DOE Federal Energy Management Program, O&M Best Practices Guide (PNNL-14788) — cross-industry predictive-maintenance program averages, not GridAPM measurements.

The workflow

From raw evidence to a signed-off work package

One pipeline, five stages. AI drafts; a named engineer decides.

  1. IntakeApproved evidence streams — DGA, partial discharge, SFRA, thermal, inspection records — are read from your systems.
  2. Correlate & quality gateSignals are aligned per asset and screened for gaps, unit errors, and stale data before any reasoning starts.
  3. Agent reasoningAgents draft a condition assessment, citing each piece of evidence and flagging contradictions instead of hiding them.
  4. EngineerEngineer sign-offA qualified engineer approves, edits, rejects, or escalates the draft. Nothing ships without a named reviewer.
  5. Work package & reportApproved decisions become maintenance work packages and audit-ready reports for your CMMS and stakeholders.
An evidence token travels the pipeline and pauses at stage 4 until an engineer signs off.

What the agents read

The evidence

Every recommendation is built from named diagnostic signals — inspectable at any time.

Dissolved gas analysis Multi-gas trends screened against IEEE C57.104 and IEC 60599 context, with rate-of-change flagged per gas.
  • H2
  • CH4
  • C2H4
  • C2H2
Dissolved gas analysis Line chart of dissolved gas concentrations in ppm over 13 oil samples. H2 ends at 148 ppm, CH4 ends at 74 ppm, C2H4 ends at 58 ppm, C2H2 ends at 11 ppm. H2 watch level: 80 ppm. Gassing event marked at sample 9. H2 watch level Gassing event
Read the method
Partial discharge PRPD patterns tracked over time to separate persistent defect activity from noise and load artifacts.
  • AC reference
  • Amplitude (pC)
Partial discharge Scatter plot of 170 partial-discharge pulses across one 360-degree power cycle, plotted against a sine voltage reference. Pulses cluster on the rising flanks of both half-cycles, the signature of an internal discharge source. Amplitude scale: ±500 pC.
Read the method
SFRA comparison Frequency-response sweeps overlaid against baseline, with deviation bands highlighted for review.
  • Baseline (factory)
  • Measured
SFRA comparison Overlaid SFRA magnitude curves (Baseline (factory) and Measured) from 20 Hz to 2 MHz. The measured curve deviates from the baseline inside the highlighted band: Deviation 10–200 kHz. Deviation 10–200 kHz
Read the method
Duval triangle Gas ratios plotted against fault zones to type a developing fault — thermal or electrical — before it escalates.
Duval triangle Duval triangle 1: ternary plot of relative CH4, C2H4 and C2H2 percentages with fault zones PD (Partial discharges), T1 (Thermal fault below 300 °C), T2 (Thermal fault 300–700 °C), T3 (Thermal fault above 700 °C), DT (Mixed thermal and electrical fault), D2 (High-energy discharges), D1 (Low-energy discharges). Plotted sample: CH4 62 %, C2H4 35 %, C2H2 3 %, falling in zone T2 (Thermal fault 300–700 °C). PD T1 T2 T3 DT D2 D1 % CH4 % C2H4 % C2H2 Sample
PD
Partial discharges
T1
Thermal fault below 300 °C
T2
Thermal fault 300–700 °C
T3
Thermal fault above 700 °C
DT
Mixed thermal and electrical fault
D2
High-energy discharges
D1
Low-energy discharges
Read the method

Illustrative data — synthetic values for demonstration, not measurements from a customer fleet.

Published CBM targets

What condition-based maintenance has documented

Cross-industry targets published by the US Department of Energy's Federal Energy Management Program.

25 –30% 1

reduction in maintenance costs

70 –75% 1

fewer equipment breakdowns

35 –45% 1

reduction in downtime

$3.3M 2

average insured loss per major transformer failure (1997–2001 claims, nominal dollars)

Trust & governance

Built for control rooms, not demos

Local-first deployment, no autonomous control actions, a named engineer on every decision, and an audit trail from raw record to final report.

IEEE, IEC, and NIST references provide context for our methods. Context is not certification.

Capabilities

What the workbench does

Evidence intake
Read-only connections to DGA, PD, SFRA, thermal, and inspection sources — provenance preserved.
Quality gating
Gaps, unit errors, and stale data are caught before reasoning, and logged when rejected.
Cited drafts
Agent assessments cite every input and flag contradictions instead of smoothing them over.
Sign-off workflow
Approve, edit, reject, or escalate — each action recorded with reviewer and timestamp.
Work packages
Approved decisions export as structured tasks for your CMMS, with the evidence attached.
Audit-ready reports
Every report traces to raw records, review comments, and the engineer who signed it.

Questions

What engineering and procurement teams ask

Direct answers for the review that precedes any pilot.

Where does the software run?

On your infrastructure. GridAPM is a local-first workbench: operational records and deterministic engineering stay inside your boundary. OpenAI-powered features send only the approved context needed for the requested operation through a controlled outbound connection.

Does the AI act autonomously?

No. GridAPM issues no control actions and changes nothing in your systems. Every output is a draft until a qualified engineer approves it at the sign-off gate, and every approval is logged.

What data do we need for a pilot?

Typically DGA history exports for the pilot units, plus whatever you have of PD, SFRA, thermal, and inspection or maintenance records. Gaps are acceptable — the quality gate makes them explicit rather than hiding them.

How does this relate to IEEE C57.104 and IEC 60599?

GridAPM applies these documents as screening context: thresholds, gas ratios, and interpretation frameworks that engineers already use. That is context, not certification — final diagnostic conclusions remain with your engineers.

How long does a pilot take?

A typical pilot is scoped in weeks, not quarters: a defined set of units, agreed evidence streams, named reviewers, and success metrics fixed before the start. You receive the full evidence pack whatever the outcome.

How is procurement and security review handled?

A procurement pack covers security posture, data handling, deployment boundaries, and RFP-style answers in one place. Your security team can review the local-first architecture before any data is discussed.

Controlled scope

Tell us about your fleet.

Share your fleet profile and diagnostic data. We will propose a focused pilot plan: evidence streams, review steps, security boundaries, and deliverables.

OT boundary

Site network

Workstation

GridAPM

Controlled OpenAI connection

Evidence in

Operational evidence and deterministic processing stay inside your OT boundary. Requested AI assistance sends only approved context through the controlled OpenAI connection.

This form collects contact details only — no fleet data. Pilot evidence stays on your infrastructure.

Local-first — runs on your workstation

More options

We use your information to respond to your pilot request. Nothing else.

Put your fleet's evidence to work

Scope a controlled pilot: your units, your data, your reviewers.

Request a pilot Learn more

Type to search research, platform pages, and tools.