The Powerwall Died Three Days Before the PSPS Shutoff. Your Portal Logged the Drift for Six Weeks.

4:12 on a Friday in October, and the utility has already sent the notice

The customer calls at 4:12 p.m. The public safety shutoff notice landed on her phone that morning, the wind event starts Sunday, and her app says the battery is at 11% and will not charge. She bought storage for exactly this weekend. Both service techs are ninety minutes away finishing a commissioning, the warehouse closes at five, and the only person who can pull a gateway log is on a roof.

You will send somebody. It will cost a Saturday at overtime, and if the unit needs a warranty replacement you are not fixing it before the wind arrives. Here is the part that should bother you: the system had been telling you for six weeks. Round-trip efficiency slipping, one module group running warmer than its neighbours, three thermal derate events in a month where there had been none all summer. All of it sat in the portal. Nobody was looking, because nobody has time to look at 1,800 systems.

Predictive maintenance in a solar and storage business means watching the drift in your own fleet's monitoring data until you can put the repair on a scheduled route day instead of an emergency Saturday. That is the entire idea, and in 2026 it finally became practical for a company your size.

The failure that costs the most is the one the customer finds first

Run last year's service tickets and sort them by how you found out. The ones your office found — a monitoring alert, a soft trip, a communication dropout — cost a scheduled visit. The ones the customer found cost a scheduled visit plus a phone call, an apology, a second visit because the tech did not bring the right part, and a real chance at a one-star review with your company name and the words "still not working."

In a residential fleet the expensive unplanned failures cluster tightly: a battery degrading or refusing to charge, a run of optimizers dropping out, an isolation fault after water gets into a conduit, a rapid shutdown device failing so the array will not wake up, and connector failures where two brands of MC4 got mated on a rushed install four years ago. On commercial roofs add combiner fuses and bypass diodes cooking a string in July.

All of these degrade before they stop. The battery does not go from healthy to dead on Friday afternoon. It loses a little capacity and efficiency for two months, and then it fails on the day it is asked to work hardest — which, because your customers bought storage for outages, is the day you least want to be explaining yourself.

What changed: 2026 stopped scheduling by the calendar

Predictive maintenance is now the most-adopted AI use in manufacturing, at around 64%, because the newer tools fuse different kinds of evidence instead of watching one number against one threshold. That matters here more than in a factory, because a solar fleet produces several unrelated streams: per-optimizer or per-microinverter production from Enphase Enlighten or the SolarEdge portal, battery state-of-health and cell balance from the Powerwall, FranklinWH or PWRcell gateway, fault and trip logs, and thermal images from a drone pass or a technician's last visit.

The 2024 version was a threshold alarm: production below X, send an alert. It generated so much noise from cloud cover, snow and a customer's new router that everyone muted it. The 2026 version compares a site to itself over time and to nearby sites on the same weather, reads the fault log alongside the production drift, and looks at the thermal image of the same string from your last inspection. Hyundai's autonomous inspection drones at its Georgia plant cut inspection time by roughly 90% doing exactly this kind of fusion; on a 300 kW commercial rooftop, one thermal drone pass plus the optimizer data finds your hot cells and dead diodes in an afternoon instead of a week of hand-tracing strings with a clamp meter.

flowchart TD
  A["Nightly production data, per optimizer"] --> D["Fleet watch scores every site"]
  B["Battery state-of-health and trip logs"] --> D
  C["Thermal photos from the last drone pass"] --> D
  D --> E{"Drift past the trip line three weeks running?"}
  E -->|No| F["Leave it alone, keep watching"]
  E -->|Yes| G["Service coordinator adds it to the next route day in that ZIP"]
  G --> H["Tech carries the right part on the first visit"]

The signals that show up weeks ahead

Ask your service manager which of these she can see today without opening four portals. In most shops the answer is none.

• Round-trip efficiency drift on a battery. Energy out divided by energy in, tracked weekly. A steady one-point slide over six weeks is the clearest early warning in storage, and it precedes the "will not charge" call by a month or more.
• One string or one optimizer group falling behind its twin. Not below a fixed threshold — below its own neighbour on the same roof plane, on the same day. Three per cent widening to seven is a connector or a diode.
• Soft trips that clear themselves. Arc-fault and ground-fault events that reset and never generate a ticket. A rising count over a month is water in a conduit or a chafed conductor, and it will eventually stop clearing.
• Communication dropouts with a pattern. A gateway that loses contact every afternoon is a heat problem or a failing modem; one that lost contact once when the customer changed internet providers is nothing. The pattern is the tell.
• Thermal signatures that moved. A warm cell in last year's image that is hotter this year on the same module is a real finding. A warm cell with no history is a maybe.

Tuesday route day, rebuilt

The service coordinator opens one page on Monday afternoon. Eleven sites with drift, ranked, each with the reason in plain language and the portal reading behind it: this Powerwall's efficiency has slid four weeks straight, this array has one string running 6% under its neighbour since 2 June, this gateway drops out every afternoon over 96 degrees. Six are within twenty minutes of each other.

She builds Tuesday's route from that list instead of from whoever shouted loudest, pulls the parts Monday evening — an optimizer, a pair of connectors, a spare gateway — and the tech leaves with them on the van. Four of the six are fixed on the first visit. Two turn out to be nothing, which is worth knowing. The customer whose battery was drifting gets a call before she noticed a problem, which is the best marketing your service department will ever do.

The arithmetic on 1,800 systems under warranty

Assumptions, illustrative: 1,800 systems under workmanship warranty, a 14% annual service event rate (252 events), an unplanned emergency visit costing $410 all-in once you count drive time, overtime and the crew day it disrupts, a planned visit added to an existing route costing $180 marginal, and 31% of unplanned visits needing a second trip for parts.

Net movement is roughly $20,800 a year on service cost alone, before you count the warranty production credits you do not owe and the reviews you do not get. Two things make or break that number: whether your techs actually record what they found so the scoring improves, and whether your coordinator has the authority to add a site to a route without the customer having complained yet. The second one is a management decision, not a software feature.

Where the data will lie to you, and where the tech still decides

Snow and soiling will fool any watcher for weeks. So will a shade sail, a neighbour's tree, and any roof where somebody added a satellite dish. Every winter you will chase production drops that are simply weather, and the fix is a human who knows the region reading the list before the truck rolls.

Current sensors installed backwards during commissioning make a healthy system look sick permanently. So do consumption readings on a home where the customer added an EV charger and never told you. Both will sit at the top of your ranked list forever until somebody clears them.

And the diagnosis stays human. The list tells you a string is 6% under its twin. Whether that is a connector, a diode, shading from a chimney the surveyor missed, or a module that needs a warranty claim against the manufacturer is decided on the roof with a meter and a set of eyes. Do not let anyone sell you a system that says "replace the optimizer" — the useful output is "look at this string, here is what moved and when."

Frequently asked questions

My monitoring alerts already fire constantly and everyone ignores them. Why is this different?

Because a threshold alarm compares a site to a fixed number and this compares a site to itself and to its neighbours on the same weather day, over weeks. The output is a short ranked list on Monday, not a stream of emails at 6 a.m. If it produces more than about a dozen items a week for a fleet your size, it is tuned wrong and your techs will mute it exactly like they muted the last one.

Do I need drones?

For residential, no — the optimizer and battery data is enough for most of the value. For commercial rooftops above roughly 100 kW, a thermal pass once or twice a year pays for itself the first time it finds a hot string you would have hand-traced. Rent the service before you buy the aircraft, and remember the pilot needs a Part 107 certificate.

How do I keep this from becoming a second job for my service coordinator?

Cap it. One page, Monday afternoon, ranked, maximum ten sites. Anything below the line waits a week. The whole point is that it fills route days you were driving anyway, not that it creates new ones.

Every one of these events also produces a phone call, and they arrive at the worst moments — a Friday before a wind event, the morning after a storm, the first hot week of July when everyone's air conditioning and their inverter argue at once. CallSphere builds AI voice and chat agents that pick up the service line at any hour, take the address, the system type and what the customer's app is showing, and put the visit on the service calendar so nothing sits in a voicemail box until Monday. The diagnosis is still yours; the intake does not have to be.