Picture a 400 kW rooftop system on a logistics warehouse near Utrecht. On a Friday in June, one of its four inverters trips and stays off. The fault shows up in the vendor portal, but nobody is looking at that portal, because the asset manager responsible for the site also covers 280 others across six inverter brands.
The fault turns up three weeks later, in the monthly report. By then the site has lost roughly 10 MWh, about €1,500 at a typical C&I PPA rate. Nobody would call that a disaster, but in portfolios of that size, we see it happen dozens of times a year.
Opinion: Europe’s battery boom now runs on data
For me, that scenario sums up the operational problem of distributed solar. A 100 MW utility-scale plant has a control room, a SCADA monitoring system and people on site. A 100 MW portfolio of commercial rooftops can mean 300 sites, each with its own data logger, meter and vendor cloud. Portfolios grow quickly through acquisitions and new installations, while operations teams can only grow by hiring, which is slower and more expensive. The result is that attention has to be rationed.
Where AI already earns its place
I use a simple test to decide what to hand over. If a task involves reading, comparing and explaining, AI can already do it well. If it involves making a decision, it is a different matter.
Every night, a software agent can compare each site’s output with its budget, adjusted for measured or satellite irradiance, and work out what caused any gap. A grid outage shows up in the meter data, while a curtailment setpoint, a limit placed on output, appears in the inverter logs alongside any fault codes, and a slow decline that tracks the weather points to soiling. The agent can also rank recurring alarms by the output that they actually cost, so an inverter that throws 200 harmless warnings a week no longer buries the one that tripped once. It can even read an operations and maintenance (O&M) contract, pull out the availability guarantee and response times, and check them against ticket timestamps, a job that many teams never assign to anyone in particular.
Solar Investors Guide #6 – Minimising risks
Reporting and billing are next. A monthly performance report or a PPA invoice is mostly arithmetic on data the system already holds: metered yield, the tariff and the contract terms. In many portfolios, billing already runs automatically. AI adds the explanation around the numbers, such as why June came in 6 percent under budget, which three sites drove the shortfall and why one bill uses an estimate because a meter dropped out for four days. It can also draft the investor report, commentary included, so nobody has to write it up from a spreadsheet. None of this is new analysis. Portfolio operators have always done it, but they could never do it for 300 sites at once.
The data foundation AI needs
None of this works without a sound data model, because an AI is only as good as the data it is built on. Inverter data from five brands arrives in five formats, with different units, sign conventions and gaps. Site metadata is often wrong: a capacity entered in kW where the system expects watts, or a meter wired the wrong way round. If irradiance comes only from an on-site sensor that has been dirty since spring, every performance ratio is fiction. In the portfolios we monitor, most of the effort before any automation goes into normalising data and fixing metadata. Skip that step, and the AI will confidently explain losses that do not exist.
Where people are still needed
Anything that triggers a site visit, touches safety or moves money still needs a person. Whether a 4 percent underperformance justifies a site visit is a commercial call that depends on travel time, the customer relationship and what else is on the schedule. Telling shading from soiling or module degradation often needs eyes on the roof, or at least someone who knows that the neighbour built an extension last spring. A dispute with an O&M contractor over a missed response time is a conversation between people. And commercial documents such as invoices and investor reports still need a human sign-off, because contracts and customers carry nuances a calculation can miss: a tariff change agreed by email, a curtailment the customer asked for, a meter reading under dispute.
A bankable battery needs a digital exit strategy
My rule of thumb is that the AI drafts, ranks and flags, while the person who is accountable, whether an asset manager, an O&M lead or a CFO, approves. People stay in charge, and they win back most of the hours that they used to spend clicking through portals.
The losses that add up
For most financial stakeholders, whether asset owners, asset managers or investors, the chronic losses matter more than the dramatic ones. Think of a string that has been down since March, a curtailment limit left in place after grid works, clipping because the DC/AC ratio was pushed too far, or soiling on a low-tilt roof after a dry summer. Each of these costs a percent or two of yield – often for months – before anyone notices.
Most operators cannot say how much of their budget gap comes from fixable causes and how much from weather, and that is the first number to find. The second is the time to detect a fault, meaning how many days pass between an inverter tripping and someone acting on the alarm. If the answer is measured in weeks, look at who is watching the portals. Usually nobody is, because no human can keep track of 300 portals. An AI co-worker can. (Andreas Lehner/hcn)
