Agentic AI for IEC 61850 Digital Substation Commissioning

Digital-substation commissioning changes the shape of the test problem. In a conventional panel, a technician verifies wiring, analog quantities, contacts, and relay logic. In an IEC 61850 system, the protection function also depends on machine-readable configuration, Ethernet services, time, switch behavior, sampled values, GOOSE subscriptions, quality bits, and the relationship between engineering files and the installed system.

Agentic AI can help organize this complexity. It should not be placed inside the trip decision or allowed to rewrite the substation configuration without a controlled engineering process.

The evidence chain in a digital substation

The core evidence is not one file. It is the relationship among:

• the approved system specification and design assumptions;
• IED capability statements and firmware versions;
• SCD, ICD, IID, and related SCL files;
• network topology, VLAN, redundancy, and time-synchronization design;
• GOOSE publisher/subscriber relationships and supervision;
• sampled-value streams, quality, scaling, and merging-unit configuration;
• protection settings and logic diagrams;
• FAT, SAT, commissioning, and maintenance procedures;
• packet captures, relay records, test-set files, and deviations.

IEC 61850-6 defines a configuration description language for IEDs and communication systems. It is valuable to an AI workflow because it is structured context rather than a screenshot, but structured does not mean self-validating. A file can be stale, incomplete, or inconsistent with the installed network.

What an agent can do before the outage

A bounded agentic workflow can perform a pre-commissioning quality gate:

1. Inventory the approved files and hash or version them.
2. Identify which IEDs, logical nodes, datasets, control blocks, GOOSE messages, and sampled-value streams are in scope.
3. Compare the SCL context with the test plan and flag signals that have no test case.
4. Check that each planned case has an expected input, output, binary response, timing requirement, and acceptance criterion.
5. Retrieve the governing manual or engineering standard passage for an ambiguous setting or protocol parameter.
6. Produce a review queue rather than silently filling a missing value.

The assistant’s role is coordination and retrieval. Parsing, schema validation, network checks, timing measurements, packet assertions, and pass/fail comparisons should be deterministic and reproducible.

GOOSE testing: message content is not enough

The GridAPM GOOSE testing guide explains why a GOOSE test must connect the message to the protection behavior. A valid test may need to cover publisher configuration, subscriber mapping, dataset quality, state changes, supervision, retransmission behavior, timing, and the resulting trip or block logic. An agent can help map the test plan to the SCL context and identify a missing subscriber or inconsistent name. It should not infer that a packet seen on a network proves the protection scheme works.

IEEE PES Technical Report 84 and IEC TR 61850-10-3 provide useful anchors for application testing. The exact procedure remains project-specific: the accepted performance, network conditions, redundancy, time source, test equipment, and boundary between FAT and SAT need to be written down.

Sampled values: scaling, quality, and time

Sampled values make analog measurement a communication problem. A test must consider sampling rate, dataset structure, nominal values, quality flags, synchronization, network paths, merging units, and how the IED reacts to loss or degradation. A technically plausible waveform with the wrong scaling can produce a wrong protection decision with no obvious visual clue.

An agentic reviewer can compare the declared configuration, test-case assumptions, and captured evidence. It can flag that a test record has no time-source information or that a selected case does not exercise degraded quality. It cannot turn a prose summary into proof. The evidence must include the actual trace or packet-level record, the deterministic assertion, and the reviewer’s decision.

Commissioning is a lifecycle, not a packet demonstration

CIGRE Technical Brochure 760 describes test strategy for protection, automation, and control functions in a fully digital substation, including certification, interoperability, FAT, commissioning, SAT, and maintenance. CIGRE Technical Brochure 949 likewise reflects implementation experience with process bus. The implication is straightforward: protocol tooling is one layer of a lifecycle test programme.

That is why a “green” AI summary is not enough. The record should identify the test stage, equipment and software versions, network and time conditions, cases executed, cases intentionally excluded, deviations, re-tests, and named approval. A digital substation makes incomplete context easier to hide because many relevant relationships sit in engineering files rather than on the panel face.

How an agentic copilot should be bounded

For an OT review, define at least these controls:

• read-only access to source evidence unless a separate change process is invoked;
• allow-listed tools and data fields rather than unrestricted shell or network access;
• no ability to write relay settings, change SCL, publish GOOSE, or control equipment;
• explicit abstention for missing, conflicting, or stale configuration;
• source-linked draft output and a record of model, prompt context, tool calls, and reviewer edits;
• local processing for deterministic engineering and controlled outbound context for requested AI assistance;
• security review against the real architecture, not a marketing phrase such as “IEC 61850 ready.”

Research using LLMs with human-in-the-loop and HIL testbeds is relevant evidence that controlled AI workflows can assist analysis. It is not evidence that an agent should issue live trips. The distinction must stay visible in the product, the procedure, and the article describing it.

Where ProtectionAI fits

ProtectionAI can be positioned as an evidence-first assistant around the relay-testing workflow: it can help an engineer prepare a test, reason over approved manuals and settings, use a simulator, and draft a report. Its public product boundary should be stated in every serious IEC 61850 discussion: simulator and loopback capabilities are not the same as qualified physical process-bus testing, and AI output remains a draft.

The useful pilot question is whether the workbench catches missing coverage or configuration assumptions earlier. Measure that with a fixed sample of SCL, test plans, and known deviations. If a tool claims improved commissioning, show the baseline, cases, false positives, false negatives, reviewer time, and what still had to be tested at FAT or SAT.

References

• Apostolov, A., & Leitloff, V. (2019). Test strategy for Protection, Automation and Control (PAC) functions in a fully digital substation based on IEC 61850 applications (CIGRE Technical Brochure No. 760). https://www.e-cigre.org/publications/detail/760-test-strategy-for-protection-automation-and-control-pac-functions-in-a-fully-digital-substation-based-on-iec-61850-applications.html
• IEEE Power & Energy Society. (2020). Application testing of IEC 61850 based systems (Technical Report 84). https://resourcecenter.ieee.org/publications/technical-reports/pes_tp_tr84_psrc_120720
• International Electrotechnical Commission. (2022). Communication networks and systems for power utility automation—Part 10-3: Functional testing of IEC 61850 systems (IEC TR 61850-10-3:2022). https://webstore.iec.ch/en/publication/61395
• International Electrotechnical Commission. (2024). Communication networks and systems for power utility automation—Part 6: Configuration description language for communication in electrical substations related to IEDs (IEC 61850-6 consolidated version). https://webstore.iec.ch/en/publication/103863
• Zaboli, A., Choi, S., Song, T.-J., & Hong, J. (2024). ChatGPT and other large language models for cybersecurity of smart grid applications. In 2024 IEEE Power & Energy Society General Meeting (pp. 1–5). IEEE. https://doi.org/10.1109/PESGM51994.2024.10688863