“Great, so the agents see the drone data.”
I hear that all the time when I explain ClearSpot’s architecture. And every time, I know we have not explained it yet.
That line is the 10-second version of how AI agents and drones work together in utility-scale solar O&M. It is technically true, but operationally incomplete. What matters is the sequence: who decides when to fly, what happens to the imagery, how quickly the system produces a work order, and what happens after the work is done.
Why the gap matters
Solar operations have historically lived in two separate worlds.
The first is electrical intelligence: SCADA, string monitoring, and performance analytics. This world processes inverters, current sensors, irradiance meters, and temperature data. It is fast, continuous, and rich in numbers, but it cannot see physical defects.
The second is visual intelligence: drone inspection, thermal imaging, and RGB orthomosaics. This world finds micro-cracks, soiling, bypass diode failure, and overheating connections. It sees what electrical systems miss, but it is episodic and disconnected.
The gap between those worlds is where a large share of O&M value is lost. SCADA can spot an anomaly, but it cannot tell you whether the cause is soiling, hardware damage, or an early fault. Drone imagery can show the defect, but without operational context it stays isolated. ClearSpot’s agentic AI platform closes that gap by making electrical and visual intelligence part of the same operational context.
The five layers
The architecture works in five layers, and each one has a specific job.
1. Electrical perception
The SCADA integration agent ingests data from your existing SCADA system through standard protocols such as Modbus, IEC 61850, and REST. It normalises the data, checks for drift, and feeds the result into the performance agents.
No SCADA replacement is required. The existing hardware stays in place and becomes the sensor network for the mesh. That is why the integration is non-disruptive: the intelligence layer changes, but the hardware investment is preserved.
2. Anomaly detection
The String Performance agent and Performance Monitoring agent continuously analyse the normalised SCADA data. They do not rely on a single threshold crossing. Instead, they look for rate-of-change patterns, cross-string correlation, and persistent underperformance.
A string that drops to 94 percent of expected output may be noise. A string that declines by 0.3 percent month after month for six months, while nearby strings show the same pattern, is a physical investigation signal. That is where AI performance benchmarking becomes more useful than classic alarm logic.
3. Dispatch decision
When the Performance Monitoring agent sees a signal that warrants physical investigation, it creates a drone inspection request in the shared operational context. The request includes the affected string or zone, the anomaly type, and the recommended imaging mode.
The O&M Operations agent then checks weather, drone availability, and pending inspections. If conditions are good, it dispatches a targeted solar panel drone inspection. If the site uses drone-in-a-box hardware, the flight can be autonomous. If it uses a service provider, the request is sent with a precise flight brief.
4. Physical perception
The Drone Vision agent receives the flight brief, manages the mission, and processes the imagery after landing. This is more than image review. It includes orthorectification, geo-referencing, thermal classification, severity scoring, and spatial matching to the electrical anomaly that triggered the flight.
Classification only works on radiometric thermal data, where every pixel carries a temperature value. The agent compares each module with healthy neighbours under the same irradiance and scores severity on the temperature delta (ΔT), following the capture conditions set out in IEC TS 62446-3 for outdoor infrared thermography of PV plants. A non-radiometric image can show that a hot spot exists, but not how serious it is.
That is why the full technical workflow matters, as our guide to radiometric thermal inspection of solar farms explains. The output is not a PDF report. It is a structured defect register that the operations layer can use immediately.
5. Action
The defect register flows into the O&M Operations agent. Each defect is classified, scored for severity, tied to production impact, and turned into a work order specification.
The system checks inventory before creating the job. If parts are available locally, technician scheduling begins at once. If parts are missing, procurement starts at the same time. The result is a direct path from detection to action, without manual handoffs slowing the process.
An illustrative example
Here is what the loop looks like in practice. This is an illustrative example: “Site Alentejo” and the timings below are a representative scenario, not a client record.
Monday 07:23: the String Performance agent flags String C7 at Site Alentejo. It has declined by 0.31 percent month on month for seven straight months. That is a rate-of-change signal, not a threshold alarm.
Monday 08:15: the O&M Operations agent raises a drone inspection request. Weather is clear and wind is below 12 km/h, so dispatch is approved.
Monday 08:40: the drone launches from the site. It surveys rows 12 to 19 with thermal and RGB imaging.
Monday 09:15: the drone returns and uploads the imagery.
Monday 11:30: the Drone Vision agent finishes processing. It identifies seven hotspots linked to bypass diode failure, two hotspots linked to soiling shadows, and one connection overheat at junction box C7-J14.
Monday 12:05: the O&M Operations agent generates two work orders. One is urgent cleaning and re-torquing for the junction box. The other is a repair job for the seven affected modules, with parts procurement triggered.
Monday 12:05: the Warranty Management agent also flags the module batch against warranty terms. Because the batch is still inside coverage, a claim package is initiated.
Total elapsed time: less than five hours from anomaly signal to work order and warranty action.
That is the operational difference. It is not just faster inspection. It is faster decision-making, faster dispatch, and faster recovery.
Why this matters at scale
On a 100 MW portfolio, faster fault detection, shorter repair cycles and better warranty recovery add up. ClearSpot reports 99% fault-detection accuracy, a 30–50% reduction in O&M cost, a 65% reduction in downtime and 97%+ plant availability when agents and drones run as one loop. Targeted flights also change inspection economics: an AI drone inspection costs a fraction of a conventional drone survey with manual image review or a manual inspection, and ClearSpot is priced as a platform subscription, from $50 to $1,000 per MW per year depending on portfolio scale, site size and the depth of automation and monitoring.
The point is not that drones replace SCADA. The point is that drones extend the operating layer beyond what SCADA can do alone.
Conclusion
AI agents and drones work together because they solve different parts of the same problem. The agents decide, the drones perceive, and the operations layer turns that combined context into action.
For grid and operational context, the IEA PVPS report on active power management and SolarPower Europe provide useful background on why solar systems need more dynamic operational control.
Frequently Asked Questions
Agents watch SCADA and string data for anomalies, decide when a physical check is needed and dispatch a targeted drone flight. The drone imagery is classified into a defect register, and the agents turn each defect into a work order, parts request or warranty claim.
No. The SCADA integration agent reads existing SCADA data over standard protocols such as Modbus, IEC 61850 and REST. The hardware stays in place and becomes the sensor network for the agents.
A signal that needs physical investigation, such as a persistent rate-of-change decline on a string or cross-string correlation, rather than a single threshold alarm. The operations agent then checks weather, drone availability and pending inspections before dispatch.
A structured defect register rather than a PDF: each defect is geo-referenced, classified from radiometric thermal and RGB imagery, scored for severity and matched to the electrical anomaly that triggered the flight.