Solar Plant O&M in Practice: From Scheduled Inspections to AI Detection

1. Day-to-day operations: six things that must be done on a schedule

Operations management rests on one foundation: observability. These six items form the baseline framework.

1. **Observe inverter start/stop behaviour** — abnormal cycling is usually the first signal of a DC-side or grid-side problem.

2. **Measure DC operating voltage and per-string current** with a voltmeter and clamp meter to confirm output is normal.

3. **Inspect the connections and insulation** of module leads, connectors, combiner boxes and inverters.

4. **Observe and record daily, monthly and annual generation** — without a baseline there is no diagnosis.

5. **Maintain contact with the utility** to keep generation and transmission routes open.

6. **Walk the site with no blind spots** to confirm modules and arrays are undamaged, not missing, and free of hazards.

Of these six, item 4 is the one most often treated as a box-ticking chore. It is in fact the only thread that ties the other five together. **Generation is the outcome; everything else is a cause.**

2. Invisible losses: dust, bird droppings and shading

Dust accumulating on module surfaces directly reduces transmittance, and bird droppings create localised shading. The damage from the latter goes far beyond "blocking a bit of light" — **a shaded cell flips from a generating element into a consuming load**, its temperature spikes, and a hot spot forms.

The usual answer is "clean on a schedule". The real question is **how long that schedule should be**. Fixed-interval washing either erodes returns by over-cleaning or lets losses accumulate by under-cleaning.

**This is precisely where data pays for itself.** Back-calculate the theoretical output from historical irradiance data, divide it into the plant's actual output for the month, and you get an actual-to-theoretical ratio. Months persistently below the median (say, under 0.7) point to soiling, shading or equipment efficiency problems. Cleaning cost is a known expense; soiling loss is a measurable variable — **the optimum sits exactly where the two are equal**. The cleaning interval stops being a rule of thumb and becomes a loss-threshold trigger.

> Note: the theoretical side of this ratio is derived from irradiance data and an assumed performance ratio (PR). It is an **estimate**, affected by weather, module temperature and equipment condition — useful for spotting trends, not as a settlement basis.

3. Temperature rise: the most overlooked failure precursor

Temperature is the most honest indicator of an electrical fault. Points to verify regularly: inverter operating temperature, temperature at string connections and connectors, contact temperature of fuses and switches inside combiner boxes, and contact temperature at HV switchgear and the grid connection point.

For this kind of "invisible heating", **infrared thermography is the most effective investigative tool**. Plants that can afford it should carry IR cameras — it eliminates a large share of failure precursors at the budding stage.

When temperature rise runs away, the consequence is not lower efficiency but **fire**. In Germany, numerous residential fires occurred at PV-equipped homes where **firefighters refused to intervene because the modules keep producing voltage even after the mains are cut**. The root cause in nearly every case is the same: **contact resistance at a joint quietly climbing through years of thermal expansion and contraction**.

Worth noting: IR thermography depends on specialist equipment and on how often a human walks the site. **Within the visible spectrum, a further set of anomalies can be found automatically by camera plus AI** — see section 9.

4. Current and voltage consistency: the 5% line

String branch currents and total output current at each combiner box or inverter should be measured on a schedule. **In clear weather, branch currents should be essentially identical, with a relative deviation no greater than 5%**; likewise for branch voltages at the combiner box.

This 5% line matters because it converts "one of the modules has a problem" from **invisible to quantifiable**. Any branch exceeding 5% is worth narrowing down to a specific string, connector or junction box.

5. Module-level failures: from junction box to hot spot

Burned lead terminals and destroyed bypass diodes in junction boxes are common failures in string plants, and in severe cases they char the module backsheet and shatter the module.

Module connections and combiner boxes are the highest-frequency failure zones. MC4-type connectors must be **kept out of direct sun and rain** — otherwise the connector ages, the internal metal corrodes, contact resistance rises, and the outcome ranges from lost efficiency to arcing and fire. During assembly, confirm the terminal is fully seated (there is an audible click).

A module's operating environment swings **-30°C to +70°C** in daily range. If cell spacing is poorly designed or ribbon quality is poor, the ribbon endures a bending cycle of thermal expansion and contraction every single day, ultimately **fracturing from metal fatigue** and scrapping the module. The same day-night delta also causes solder joints between ribbon and cell to de-bond.

A hot spot is a cell driven into abnormal operation by an internal defect or local shading, running far hotter than its neighbours. It sets off a vicious cycle: **high temperature → EVA discolouration and dark patches → solder de-bonding → cell burnout → whole-module failure**.

One point is widely misunderstood: hot spots do not only occur on shaded modules. **Even at desert plants with no shading whatsoever, hot spots are remarkably common.** Removing shading therefore does not solve the problem — it has to be found by inspection.

IR thermography is the standard detection method. At utility scale, **AI batch screening of imagery** upgrades "sampling" to "census" — a difference of an order of magnitude. But note the boundary: **a hot spot is a thermal signature and requires IR hardware to image directly**. What visible-light AI covers is appearance anomalies, which is the subject of section 9.

Silver-grey network patterns appearing on module surfaces are a quality issue, caused mainly by oxidation and corrosion of the silver grid electrodes by air and other corrosive gases. In severe cases the silver grid lines fracture and generation suffers.

6. Degradation and lifetime: let measured data speak

Crystalline silicon modules lose roughly **2–3%** efficiency over the first **50–100 days** of exposure, then stabilise and decline **0.5–0.8% per year thereafter**, averaging roughly **10–20% over 20 years**. Amorphous silicon follows a different curve shape but likewise has a steep initial decline. On top of this, EVA discolouration and calcium deposits on the glass cut transmittance further.

Industry consensus puts **crystalline silicon module lifetime at 25–30 years**, amorphous silicon at 15–20 years or more, and other materials at roughly 20–30 years. This figure is not merely theoretical — a **teardown test on four 45 Wp modules manufactured by Japan's Nippon Seiki after 29 years in service** recorded power losses of **16.8% / 12.9% / 19.6% / 14.7%, averaging 16.0%**.

**For the investment side, this data matters far more than for the operations side.** It shows that "25-year lifetime" is a distribution, not a constant safe to apply across the board. Individual variation within a single batch can reach several percentage points. And **transaction pricing, asset valuation and insurance underwriting all rest on that curve.** Whoever holds genuine, plant-level degradation data holds the advantage in pricing.

7. Foundations, mounting structures and trackers

- **Foundations and structures**: keep them stable and reliable, with no corrosion on fixing bolts and sound grounding of the structure.

- **Safety systems**: monitoring and fire-safety equipment on site must be operational and on standby.

- **Agrivoltaic / aquavoltaic systems**: irrigation and tilling disturb the ground heavily, so pay particular attention to **foundation settlement and displacement, structure corrosion, and insulation damage to electrical equipment and cabling**.

- **Tracking systems**: the addition of electronic control and mechanical drive means maintenance demand is **far higher than for fixed-tilt structures** — scheduled inspection and maintenance is mandatory.

There is an easily missed cost structure here: **O&M complexity differs by an order of magnitude across plant types.** Agrivoltaic O&M costs are inherently higher than ground-mount, and tracker systems higher than fixed tilt. Applying a uniform O&M rate in investment appraisal systematically overstates the returns of the former and understates the risks of the latter.

8. Extreme weather and human risk

Storms, floods, theft, animal gnawing, fire — these low-frequency, high-severity events are routinely omitted from O&M budgets entirely. A storm can destroy the gearing of a dual-axis tracker; flooding soaks foundations and dampens electrical equipment.

Against these risks, **the cost gap between reactive repair and pre-emptive warning is an order of magnitude.** Camera-based perimeter security and weather-linked alerting are low-cost means of avoiding high-loss scenarios.

9. Handing the checkpoints to AI: what works today, and what does not

The checkpoints above are nothing new — the industry has run them for twenty years. What is changing is **how they are executed**. Below, split into "already automatable" and "still needs a human". **That line matters more than any feature list.**

**① Perimeter intrusion detection (people / vehicles)**

From video frames captured by on-site cameras, an object-detection model automatically identifies people and vehicles entering the site, raises an alert and stores the frame. It replaces the blind spot that exists whenever nobody is watching.

**② Panel appearance anomaly screening**

Captured images are analysed by AI to detect appearance anomalies such as **snow accumulation, soiling, damage and obstruction**. Camera positions shoot on a fixed interval, analyse automatically, and log anomalies to a timeline with the frame attached.

**③ Health management of the cameras themselves**

Automatic offline detection, automatic recovery attempts, and notification on recovery — preventing the situation where you believe you are monitoring a site that has in fact been disconnected for weeks.

**④ Inverter data health analysis**

Upload the data exported from your inverters and receive rule-based health analysis: efficiency grade (A–E with a 0–100 score), power and efficiency performance, degradation trend, overload and under-utilisation, zero-output records, and data-quality gaps — in Chinese, Japanese and English.

**⑤ Historical-weather generation diagnosis and a 7-day forecast**

Theoretical output is back-calculated from historical irradiance and compared against actual generation to flag anomalous months; an ensemble forecast then gives the coming seven days as P10 / P50 / P90 bands.

**⑥ Periodic health reports**

Five-dimension health scoring, actual versus expected generation, regional ranking and maintenance recommendations, exportable to PDF.

- **Hot spots and contact temperature rise** — thermal signatures requiring IR hardware; outside the scope of visible-light AI

- **Internal degradation of MC4 connectors and junction boxes** — needs opening the enclosure or IR re-inspection

- **Quantitative assessment of foundation settlement and structure corrosion** — needs on-site measurement

- **Bulk screening of IR / drone imagery** — still predominantly manual interpretation

AI detection replaces the **frequency and coverage of inspection**, not human judgement. It turns "look at it once a quarter" into "look automatically every few minutes", and "sample a handful of points" into "run every frame". But **conclusions still require engineer review** — particularly before escalating an anomaly into a repair action.

Closing: turning inspection data into an asset

The end point of O&M is not repairing equipment. It is **being able to explain every single kilowatt-hour that should have been generated and was not**.

Once losses are attributable and quantifiable, they become priceable and optimisable — and at that moment O&M data stops being a cost line and becomes an asset. The value of AI detection is that it **lowers the cost and raises the frequency** of making losses attributable.