AI Claims in Solar Inverters: Evidence Checklist

“AI” on an inverter datasheet or dealer brochure is a label, not a specification. It can describe three very different things, and only one of them involves a model that learns anything about your plant. This guide separates them, shows the arithmetic behind a learned baseline, and lists what to verify on the exact model before you pay for the label.

Three things the label can mean

The first row is useful and common. The third row is real engineering, but it is control logic rather than “intelligence” about your site. Only the middle row changes what you learn about the plant, and it is the one that depends most on data quality. The smart inverter glossary entry covers the grid-support side, and the smart solar inverter guide covers what else a datasheet can evidence.

How a learned baseline works, with numbers

A baseline model needs an estimate of what the plant should have produced. The standard yardstick is the performance ratio, defined in IEC 61724-1 as the ratio of the energy actually delivered to the energy an ideal plant would deliver for the irradiation received. In formula form:

That last condition is the weak point. The model needs measured irradiance from a sensor, or a satellite estimate for the location, and a good estimate of the plant’s own PR when it was healthy. Without them the system is guessing. See the performance ratio glossary entry and the plane-of-array entry for the definitions, and the irradiance guide for sourced yield figures by state.

What a model can and cannot flag

Plausible findings from a baseline comparison include one MPPT input producing less than its neighbour in good light, a steady shortfall that suggests soiling or shading, output clipping on a hot afternoon, and a communication gap. These are observations of reported data, and each one still needs a person to look at the plant.

A model cannot:

• repair or replace a failed component, or compensate for an undersized inverter or poorly sized strings;
• prevent damage from lightning, a grid surge, flooding or physical impact, which are handled by surge protection devices, earthing and mounting;
• fix wiring, earthing or ventilation faults left by the installation;
• keep working when the router, SIM or cloud account is down. Monitoring stops with the link, so ask what happens during an outage and whether the logger buffers data.

Verify the claim before you pay for it

Use the same short test for every shortlisted model.

The alerts and offline behaviour comparison shows how to run this test across five exact models, and the monitoring app comparison covers the apps themselves. For messaging channels, see the WhatsApp monitoring guide.

Control features that sit under the AI label

Some functions marketed as smart are fixed engineering that has existed for years, and they are worth judging on their own evidence rather than as AI.

• MPPT tracking. The tracker adjusts the operating voltage to extract the most power from each string, and a better algorithm recovers more energy under partial shading or fast-changing cloud. The test is the MPPT efficiency measured to a standard (IEC 62891, adopted in India as IS 17980) and the number of independent inputs, both of which are datasheet items. See the MPPT glossary entry.
• DC arc detection. Some inverters detect and interrupt a DC arc. The relevant international standard is IEC 63027. Ask whether the feature is fitted on your model, not whether it exists in the range.
• Anti-islanding and grid support. Disconnection when the grid fails is a safety and grid-code requirement tested under IS 16169 / IEC 62116, not an optional intelligence feature. The anti-islanding entry explains it.

Data, privacy and cost

A cloud model needs your plant’s data to leave the building. Before you accept it, ask who the data controller is, where the data is hosted, how long history is kept, whether you can export it and delete it, and what happens to the account when ownership of the plant changes. Ask also whether the analysis is included in the product price or sits behind a subscription, and what you lose if the subscription lapses. A feature you cannot afford to keep is not a feature of the inverter.

A five-minute acceptance test

At handover, ask the installer to do the following with you present. Open the app and confirm that the plant shows today’s output and that it matches the inverter display. Disconnect the logger from the network for a few minutes, then restore it, and see whether the gap is flagged and the data back-filled. Ask for a sample of each alert type the model supports and where it is delivered. Record the plant ID, login and recovery route. If any step cannot be shown on the day, write it into the handover record as open.

What Qbits documents

Qbits’ public datasheets list Wi-Fi on the on-grid series with RS485 or GPRS as options, and the hybrid series differ again, so the datasheet for your exact model is the reference. Qbits publishes monitoring apps on Google Play and the App Store. The datasheets do not describe a machine-learning method, an alert list or a training dataset, so this guide makes no claim about one. If a seller says a Qbits model has AI features, ask for the written model-level description first. Qbits pricing is quote-based through dealers; the datasheet library and the on-grid inverter range show the specifications per model.

Commissioning handover

Whatever the label, the handover should include the plant account, device identifiers, login recovery route, logger details, network requirements and a demonstration with live site data. Confirm that the values on screen match the installed inverter and meter, and record who holds installer or fleet-level access.

The short version

Treat “AI” as a question, not an answer. Find out whether it means thresholds, a learned baseline or control firmware, ask for the line in the datasheet for your model, see a real alert, and confirm what happens when the internet drops. A plain, well-commissioned monitoring setup with a tested alert route beats an unverifiable AI badge.