Every solar technology vendor claims AI. Very few deliver it. A proper solar AI evaluation separates platforms that make operational decisions from tools that simply raise more alarms, and it should be run before you sign a multi-year contract for a utility-scale portfolio.
The difference between real AI and simple threshold alerting is not academic. It shows up in lost revenue, wasted time, and unused drone reports.
Short answer: real solar AI makes decisions, not just alerts. It tells you whether to act, wait, or ignore — and ties every issue to financial impact.
If your system cannot do that, it is not AI.
The Real Problem – Most “Solar AI” Is Just a Better Alarm System
The Alert Inflation Trap
Most systems trigger alerts based on thresholds.
- Most alarms are noise that never leads to a work order
- Engineers waste time clearing false alarms
- Real issues get missed
This is not AI. This is automation with alerts.
The Disconnected Drone Problem
Drone inspections often produce PDFs and thermal reports that are never used operationally.
If drone data is not connected to SCADA and work orders, it does not improve performance.
The Model Decay Problem
AI models that don’t learn from your fleet become less accurate over time and increase false positives. Equipment ages, modules degrade, trackers are re-tuned and inverters are replaced; a model trained once at commissioning slowly drifts away from the plant it is supposed to describe.
The 6 Standards Real Solar AI Must Meet
| Standard | What it means |
|---|---|
| Decision-making | AI decides action, not just alerts |
| Financial output | Every issue shown in €/day |
| Continuous learning | Improves from your data |
| Drone + SCADA integration | Full mapping to assets |
| Data ingestion | Handles PDFs, notes, images |
| Data ownership | You own your data |
Q1 – Is Your AI Making Decisions or Just Alerts?
If your engineers are clearing hundreds of alarms but acting on only a few, your system is not AI.
Real AI filters noise and escalates only high-value issues.
A useful test: ask the vendor for the ratio of alarms raised to work orders created on a comparable site, and how that ratio changed over the first six months.
Q2 – Can Drone Data Link Directly to SCADA?
Without integration, drone insights remain unused.
With integration, anomalies are directly mapped to strings and converted into work orders.
In practice this means the drone imagery is geo-referenced to an as-built layout, each anomaly is tagged with its module, string, combiner and inverter ID, and that ID is matched to the SCADA channel that measures the same equipment. A string that shows a bypass-diode pattern in the thermal image and a matching current drop in SCADA is a confirmed fault; one that shows neither is not worth a truck roll.
Q3 – How Accurate Is Your Expected Yield Model?
Expected yield is the baseline every alert is measured against, so its error matters more than any dashboard feature. Ask the vendor to show monthly expected-versus-actual error by site, and how quickly the model updates after a change on site.
A credible model uses plane-of-array irradiance and module temperature, follows the performance-ratio definitions in IEC 61724-1, and attributes losses to specific causes. It must also separate curtailment (output limited by a grid or offtaker instruction) from clipping (DC power above the inverter’s AC rating), because one is a commercial loss and the other a design choice, and neither is an equipment fault.
Q4 – Are You Flying Drones Too Often or Too Rarely?
Drone inspections should be condition-based, not scheduled blindly.
- Trigger inspections based on anomalies
- Use targeted flights
- Avoid unnecessary costs
A sensible pattern is one full-site thermal survey per year, flown in conditions that meet IEC TS 62446-3 (stable, high irradiance on the module plane and a radiometric camera), plus targeted flights over blocks where SCADA shows unexplained underperformance.
Q5 – Can Your AI Use Unstructured Data?
Technician notes, PDFs, and photos contain valuable insights that most systems ignore.
Advanced AI converts this into structured intelligence: a technician’s note that a combiner fuse was replaced, or a photo of a damaged connector, becomes a dated event on that asset’s history, so repeat failures become visible across the fleet.
Q6 – How Do You Prioritise Drone Defects?
Not all defects are equal.
Prioritise using:
- Severity
- Yield loss
- Risk
- Recurrence
For thermal findings, severity starts with the temperature difference (ΔT) between the anomaly and comparable healthy modules, measured with a radiometric camera, and with the anomaly classes described in IEC TS 62446-3. Severity alone is not enough: a mild anomaly repeated across a whole string can cost more energy than a single hot cell. The IEA PVPS report on degradation and failure modes is a good reference for which defects tend to worsen over time.
Q7 – Does Your AI Learn From False Positives?
Real AI reduces noise over time. Poor systems repeat the same errors. Ask how an engineer marks an alert as a false positive, and whether that feedback retrains the model for that site or only hides the alert.
Q8 – Are You Ready for BVLOS Drone Operations?
Large solar sites benefit from beyond-visual-line-of-sight (BVLOS) flights, because one flight can cover blocks that would otherwise need several launch points. BVLOS needs regulatory approval: a waiver or authorisation from the FAA in the United States, or an operation in the EASA specific category in Europe. Check whether the vendor’s drone partner holds the approvals your sites need.
Q9 – Does Your Digital Twin Deliver Financial Insights?
A real digital twin connects operational data with financial modelling: every open defect carries an estimated energy loss and a revenue figure at the site’s actual tariff or PPA price, so the backlog can be sorted by value.
Q10 – Is Your “Machine Learning” Actually AI?
| Feature | Threshold System | Real AI |
|---|---|---|
| Detection | Rule-based | Pattern recognition |
| Context | None | Asset-specific |
| Learning | Static | Continuous |
| Output | Alerts | Financial insights |
| Action | Manual | Automated |
How to Run a Solar AI Evaluation in 30 Days
Claims are easy to make in a sales demo. A short pilot on one of your own sites shows what the platform really does.
- Pick a representative site. Choose one with known issues and at least a year of SCADA history.
- Record a baseline. Count alarms per week, work orders raised, and mean time to repair (MTTR) under your current tools.
- Connect the data. SCADA, the latest drone inspection, work-order history and technician notes. Note how much manual mapping the vendor needs.
- Compare outputs. For each issue the platform raises, check whether it is real, whether it was already known, and whether it carries an energy and revenue estimate.
- Test the feedback loop. Mark false positives and check whether the same alert stops recurring.
- Check data ownership. Confirm you can export raw and processed data in open formats if you leave.
At the end, compare alarm volume, confirmed issues and MTTR against the baseline. Those three numbers answer the evaluation better than any feature list.
External Resources
Frequently Asked Questions
Check the false positive rate, whether the system recommends an action rather than just raising an alert, and whether every issue carries an energy and financial impact. Real AI also improves when engineers give it feedback.
Only when the inspection data is integrated with SCADA and maintenance workflows. Anomalies need to be mapped to specific modules and strings, confirmed against production data and turned into work orders.
Combine thermal severity, measured as the temperature difference against healthy modules, with estimated energy loss, safety risk and recurrence. Rank the backlog by financial impact rather than by defect count.
Yes, if it measurably reduces alarm noise and mean time to repair and improves yield on your own sites. A short pilot against a recorded baseline is the most reliable way to find out.
Alarm volume, the share of alerts that are confirmed issues, mean time to repair, the accuracy of the expected-yield model, and whether you can export your data in open formats.
Conclusion
Most solar AI tools create more data. Very few create decisions.
If your system does not prioritise, decide, and learn — it is not AI.