AI Thermal Inspection and Spot Cleaning for Utility-Scale Solar

On a utility-scale solar park, most of the losses that matter start small: one cracked cell, one failed bypass diode, one bird dropping sitting on a cell. Inverter and string monitoring average these defects away, while a drone with a thermal camera sees each of them as a temperature difference on the module surface. My position is that AI thermal inspection, flown to a recognized standard and connected directly to the work-order system, together with targeted spot cleaning of localized soiling, belongs in the core O&M scope of every utility-scale plant, not in the list of optional extras. The value comes from how quickly a finding becomes a closed repair.

What drone thermography finds on a PV plant

When part of a module stops converting light into electricity, the absorbed energy stays in the module as heat, so defective areas appear warmer than their healthy neighbors ^[1]. The IEA PVPS Task 13 review of infrared and electroluminescence imaging groups the main thermal abnormalities as hot cells (from broken glass, external shading or internal cell problems), heated bypass diodes, heated string fuses, and heated cables and connection points ^[1]. Task 13’s 2022 O&M guidelines add open- or short-circuited strings and tracker mismatch ^[2].

The energy impact varies widely between anomaly types:

• Bypassed substrings and diode faults. In a module with three bypass diodes, a bypassed substring contributes nothing, so the module loses about one third of its power ^[8]. Task 13 recommends replacing modules with permanently activated bypass diodes ^[1]. Open-circuit diodes, for example after a lightning strike, leave the cells unprotected, and hot spots can follow ^[8].
• Cracked cells. Task 13 reports that inactive area above about 8% of a cell causes a power loss roughly proportional to that area, while 50% or more activates the bypass diode and again costs one third of the module. Cracked cells with a local reverse current path can also develop hot spots ^[8]. Infrared shows the consequence, not the crack itself. Electroluminescence imaging confirms it ^[3].
• Disconnected modules and string outages. A module, or a series of modules, that is uniformly warmer than its neighbors is probably not connected, and the wiring should be checked ^[1]. String monitoring may report a dead string. The thermal map shows exactly where it is.
• PID. Warmer cells close to the bottom frame in a patchwork pattern are a strong indicator of potential-induced degradation ^[1].
• Hot spots in general. An analysis of 2,580 modules across the UK found that hot-spotted modules had performance ratios between 0.83% and 15.47% lower than healthy modules, with losses growing as more cells were affected ^[9].

Some faults are subtle. Disconnected modules, PID-affected cells and short-circuited bypass diodes produce only small temperature differences that scale with irradiance, so under weak sun they can be hard or impossible to detect ^[1].

Inspection conditions under IEC TS 62446-3

IEC TS 62446-3:2017 is the technical specification for outdoor infrared thermography of PV modules and plants in operation ^[4]. It defines procedures and requirements, but, as a Task 13 expert has pointed out, it contains no pass/fail criteria and was not written specifically for aerial inspection ^[10]. SolarPower Europe’s O&M guidelines state that aerial thermography that does not meet the specification’s minimum requirements is of little value for maintenance, and list the key ones ^[3]:

• Irradiance of at least 600 W/m², measured continuously on site, ideally in the plane of the modules.
• An infrared camera with at least 640 × 512 pixels and a thermal sensitivity of 0.04 K or better.
• A flight distance at which each solar cell is imaged with 5 × 5 pixels.
• Wind speed below about 28 km/h for safe drone control, depending on the aircraft.
• Temperature differences projected to nominal irradiance before anomalies are classified.

Task 13 adds stable conditions (no clouds, low wind, steady ambient temperature) and avoiding reflections ^[1]. The plant should be at its maximum power point, in steady state and free of partial shading ^[10]. A curtailed plant is not at its maximum power point, so I would not accept data flown during curtailment.

IEC TS 62446-3 sorts findings into three classes of abnormality: class 1, no abnormality and no action; class 2, a thermal abnormality whose cause should be checked and, if necessary, rectified within a reasonable period; and class 3, a safety-relevant thermal abnormality that calls for prompt interruption of operation ^[11]. Because the specification sets no numeric limits, the O&M contract has to. The Task 13 review gives a workable starting point: a cell temperature difference below 10 °C is normally unproblematic, 10-20 °C should be watched in later inspections, and above 20 °C is expected to degrade output and justifies replacing the module ^[1]. Task 13 also points to ISO 9712 Level 2 as a guide to analyst qualification ^[1].

Platform choice matters. Aircraft-mounted cameras can cover around 100 MW per hour but struggle to reach the resolution needed for cell- and submodule-level failures ^[2]. Drones are slower, with SolarPower Europe citing a 12 MWp plant inspected in a single day, but can meet the resolution requirement ^[3]. Because the 600 W/m² threshold and clear skies limit flights to sunny middays, I recommend agreeing weather standby days in the contract, not just a flight date.

How AI thermal inspection speeds up the analysis

Flying is no longer the bottleneck. The work is in turning thousands of radiometric images into geolocated, classified module-level faults, loss estimates and a maintenance plan ^[3]. Task 13 notes that many services still rely on semi-manual image processing, and that human error lowers both the accuracy and the consistency of the diagnosis ^[2]. This is the part AI addresses.

The research results are encouraging:

• Bommes and co-authors built a tool that extracts individual modules from drone infrared videos and assembled 4.3 million images of 107,842 modules from seven plants. A ResNet-50 classifier exceeded 90% test accuracy across ten common anomaly types. Including manual steps, the tool supports inspecting 3.5 to 9 MWp per day, and most of its failures came from poor flight trajectories and module segmentation errors ^[5].
• Corley and co-authors at Zeitview describe a deployed pipeline built on aerial infrared surveys of 6,155 North American solar farms. Its detector, covering six defect types including hot spots, bypassed diodes, string outages and misaligned trackers, reached a peak precision of 89.8% ^[6].

The same paper offers a corrective. On the popular InfraredSolarModules benchmark of 20,000 single-module crops, published models report 85-93% accuracy, yet a simple random forest reached 84%. The benchmark may be too easy to predict field performance, and single-module crops lack the neighboring modules an analyst uses to judge whether a warm area is really a defect ^[6].

In practice, AI makes it realistic to analyze the entire plant every year, classify and geolocate anomalies consistently, and compare results year over year. It does not provide certainty. A precision of about 90% means roughly one flagged anomaly in ten may be a false alarm, and a model trained mostly on North American fleets may behave differently on European plants. As SolarPower Europe puts it, “not all the thermal anomalies may be considered failures” ^[3]. Task 13 recommends cross-checking findings with I-V tracing and electroluminescence imaging before corrective decisions ^[2].

Manual I-V curve tracing is a complement, not a substitute. In the field its accuracy is limited to around 5%, and it only reveals defects that already cause a measurable loss at the maximum power point ^[2]. Thermography covers the whole plant and also finds safety-relevant heating. I use infrared to locate problems and I-V and electroluminescence to confirm and quantify them. Comparable public price data per MWp are scarce. Task 13 reports that most service providers justify aerial inspection with preventive O&M costs more than 10% below manual, ground-based inspection, plus recovery of 1-2% of lost capacity ^[2]. Those are provider claims, to be tested against the plant’s own records.

Why localized soiling needs a targeted response

Uniform dust is a yield question. A bird dropping is a reliability question. When one cell is shaded while the rest of the string keeps producing, the string current can drive the shaded cell into reverse bias, and it dissipates power produced by its neighbors ^[1]. Bird droppings are fairly opaque, hit performance much harder than dust, and are not easily washed away by rain, so the loss persists until someone cleans the module ^[12]. Uneven soiling also hurts the maximum power point more than the short-circuit current: it is a mismatch problem ^[2].

The numbers are out of proportion to the soiled area. In a 2024 experiment on a 72-cell, 370 Wp PERC module with three bypass diodes, shading 40-60% of a single cell caused power losses of up to 36% under a solar simulator and 42% outdoors, and the shaded cell reached 85-90 °C outdoors ^[7]. That shading covers less than 1% of the module’s cell area.

Left in place, localized soiling turns into damage. Task 13’s O&M guidelines warn that uneven soiling such as bird droppings can cause hot spots with follow-up failures, including glass and cell fractures, backsheet melting and fire hazards ^[2]. SolarPower Europe notes that long intervals between cleans let droppings etch the module surface ^[3]. Task 13 adds that prolonged high temperatures from droppings may damage modules, and that it may later be difficult to establish whether the damage came from module quality or from missing maintenance ^[1].

The response should therefore be targeted. When the thermal survey and its paired visual image show a hot spot under a visible dropping, a technician should clean those specific modules within days, not wait for the next scheduled wash. SolarPower Europe recognizes that some parts of a site need cleaning more often than others, and that the method must respect the manufacturer’s warranty conditions ^[3]. After cleaning, re-check the cell with a handheld camera. If the hot spot is gone, close the work order. If it persists, the cell is damaged and the finding moves into the module-defect workflow.

Two caveats apply. For acceptance or warranty-grade surveys, soiling should be checked and cleaned before the flight, otherwise it masks or mimics module defects ^{[10, 11]}. On routine O&M surveys, soiling hot spots are findings in their own right and should be reported as a separate class. And whether to wash the whole plant on a schedule is a separate economic decision that depends on labor and water costs, tariffs and weather ^[2], which I cover in a separate article on module cleaning cost-benefit analysis.

Thermal reports as warranty and insurance evidence

SolarPower Europe notes that high-quality infrared images can serve as proof in warranty claims and that inspection reports support insurance claims as well as maintenance ^[3]. Task 13 recommends that, for module warranty purposes, the first thermographic inspection be carried out at commissioning ^[1]. It also offers a useful serial-defect trigger: thermal abnormalities in fewer than 10% of modules do not by themselves indicate a quality issue, while above 10% the cause should be evaluated, whether module quality or planning, installation, maintenance or operation faults ^[1]. The 2022 Task 13 guidelines list end-of-warranty inspections and insurance claims among typical uses ^[2], and after a hailstorm thermography was found to be the easiest way to determine module condition ^[11].

For each anomaly, the evidence chain should include the module ID or serial number, location, radiometric and visual images, irradiance at capture, normalized temperature difference, class of abnormality, action taken and re-inspection result. With 170 MWp under asset management, I treat the thermal report as part of the plant’s contractual record, not just a maintenance aid. A spot-cleaning log is what answers a manufacturer who argues that a burnt cell was caused by neglect.

Inspection-to-action workflow for O&M teams

This is how I would structure an annual campaign so that a good flight does not end as an unused report.

1. Contract the outcome, not the flight. Specify compliance with IEC TS 62446-3, irradiance logging, 5 × 5 pixels per cell, module-level geolocation on the as-built layout, delivery of raw radiometric data, numeric thresholds for each class of abnormality, turnaround time and weather standby days ^{[3, 4]}.
2. Plan the flight. Provide an up-to-date layout with module and string IDs, and use pre-programmed flight paths so every survey is repeatable ^[3]. In the EU, drone operators must be registered, flights beyond visual line of sight fall into the “specific” category, which needs an authorization or a standard-scenario declaration, and geographical zones may require a flight authorization ^[13]. Site work follows EN 50110-1 on the operation of electrical installations ^[10].
3. Check conditions on the day. Plane-of-array irradiance of at least 600 W/m², stable clear sky, wind within the aircraft’s limit, and the plant at its maximum power point, not curtailed ^{[1, 3, 10]}.
4. Run AI processing. Build the orthomosaic, geolocate each module, detect and classify anomalies, and normalize temperature differences to nominal irradiance ^{[3, 6]}.
5. Verify. An analyst reviews every class 3 and class 2 finding and removes artifacts such as reflections or emissivity differences between backsheets ^[1].
6. Triage by severity. Class 3 findings mean prompt isolation of the affected string or component ^[11]. Soiling hot spots go to spot cleaning within days. Confirmed class 2 defects are bundled into planned repair campaigns, and cells with a 10-20 °C difference go on a watch list ^[1].
7. Raise work orders. Each anomaly becomes a work order in the maintenance management system with module ID, location, images, class and due date, cross-checked against monitoring data ^[3].
8. Close and confirm. The technician confirms on site with handheld infrared, I-V tracing or electroluminescence as needed and records serial numbers for the warranty file ^{[2, 3]}.
9. Re-inspect. The next annual flight compares results year over year, follows up the watch list and checks whether abnormal modules are approaching the 10% serial-defect trigger ^{[1, 3]}.

The baseline cadence is a commissioning survey ^[1], an annual full-plant flight as part of preventive maintenance ^[2], an inspection before module warranties expire ^[2], and event-driven flights after hail or storms ^[11].

Conclusion

Drone thermography is mature, and AI has removed most of the analysis bottleneck. The remaining risks lie in how it is used: data flown outside the required conditions, AI output accepted without verification, and findings that never become work orders. Localized soiling shows why speed matters, because a single dropping can cost a large share of a module’s output, overheat a cell and later weaken a warranty claim. Fly to the standard, verify what the model flags, clean and repair quickly, and keep the records.