The alarm stack, triaged before the first truck rolls

It is 6:15 in the O&M office and the lead is staring at a weekend of alarms — inverters offline, strings underperforming, comms drops that fired forty times and cleared — with one real production fault buried somewhere in the noise.

Source: pv magazine, "AI platforms split on how far to push solar O&M automation," Jun 1, 2026 — https://www.pv-magazine.com/2026/06/01/ai-platforms-split-on-how-far-to-push-solar-om-automation/

The problem

A solar O&M lead starts the day staring at a wall of alarms that all claim to matter: inverters offline, strings underperforming, comms drops, tracker faults, nuisance alarms that fired and cleared. Buried in the noise is the one fault actually costing production.

Sorting it by hand is slow, skilled work. Each alarm has to be checked against the asset’s history, the weather, and what the plant actually produced. Do it in the wrong order and a crew drives two hours to a gateway that needed a reboot while an underperforming string sits unvisited for another week. The triage is judgment — but most of the hour is clerical: grouping duplicates, pulling history, writing the same summary again.

How the agent workflow works

1. Read the overnight feed in context. The agent pulls the monitoring platform’s alarm feed — timestamps, asset IDs, fault codes — and lines it up against the asset registry, each asset’s repair history, and production and weather data for the same window.
2. Group by asset and probable cause. Forty repeats of one fault arrive as one item, not forty rows. Related alarms on the same inverter or combiner are bundled, each with its evidence attached.
3. Sort the stack before anyone rolls. Each group is ranked production-losing, safety-relevant, watch-and-see, or noise — with when it started, what production did, and whether the asset has done it before.
4. Draft, do not dispatch. For the top items the agent drafts the root-cause note and the work order, showing the data it used and its confidence. Nothing dispatches on its word.
5. The lead works the brief top-down. The O&M lead approves, reorders, or kills each item. Detection, analysis, and drafting belong to the agent; decision and execution belong to the operator — a boundary at least one platform writes explicitly into its architecture and its contracts, per the source reporting.

What you need to set it up

• Monitoring data the agent can actually read — an export, an API, or a scheduled report. A dashboard only humans can click through is a wall.
• A clean asset registry. If the monitoring platform, the as-builts, and the CMMS name the same inverter three ways, fix the names first; the agent inherits the confusion perfectly.
• Sorting rules written by the lead: which faults are safety-relevant, what production loss earns a truck roll, what gets watched for a week. Thresholds are site-specific.
• A work-order destination — the CMMS or a shared board — so a drafted order lands somewhere real. Cost, honestly: this runs on tools the operation already pays for; the real cost is the registry cleanup and the lead’s time writing the rules.

Where it breaks

Telemetry lies politely. A dead communications gateway makes a healthy array report like a failed one, and sensor drift makes a failed array report fine. The agent cannot know the data itself is the fault. Comms alarms deserve permanent suspicion.

Weather normalization cuts both ways. Comparing output to irradiance catches real underperformance — and can explain away a genuine fault on a cloudy week. A string “fine for the conditions” three weeks running needs a human to notice.

Grouping can merge separate faults. Two problems on one asset sharing a symptom arrive as one item with one confident story. The lead’s review catches the second problem hiding inside the first.

It never sees the physical cause. An alarm does not say whether the fix is a fuse, a rodent, or a failing inverter. The root-cause note is a hypothesis, not a diagnosis.

No remote hands. The agent must never reset, re-energize, or clear equipment. The moment triage software can actuate, a software error becomes a field safety event.

The bottom line

Run it in recommendation-only mode for a month: the agent sorts, the lead sorts as usual, and the two orders get compared every Friday. Where they agree, the morning gets faster. Where they disagree is where your alarm data was lying to somebody.

Related use cases

• Weather go / no-go for rooftop work Safety
• Commissioning paperwork that tracks itself Quality
• The 6 AM superintendent’s brief Management