Finding failing solar panels with a digital twin

A solar plant can report normal production while an individual panel is quietly underperforming. The reason is structural: monitoring systems usually observe inverters or strings, so the output of many panels is combined before it reaches an operator. A weak panel may reduce the output of its string without creating an obvious plant-level alarm.

The problem is not a lack of data. Many installations already collect voltage, current, panel temperature, solar irradiance, and timestamps. The harder problem is determining what each panel should have produced under the conditions present at that moment.

My research approached this problem with a digital twin: a physics-informed model calibrated from operational telemetry for each panel. Instead of comparing raw power values directly, the method estimates expected behavior, normalizes performance across the plant, and looks for persistent or changing deviations.

What the digital twin represents

The digital twin is based on a two-diode electrical model of a photovoltaic cell. The model represents the physical processes that shape a panel's current-voltage curve, including photogenerated current, diode recombination, series resistance, shunt resistance, and temperature-dependent behavior.

For anomaly detection, the important point is not the equation itself. It is that the model connects environmental and electrical measurements to a set of internal parameters that describe how a particular panel behaves.

The input telemetry consists of:

• voltage;
• current;
• panel temperature;
• solar irradiance;
• measurement time.

From these inputs, the calibrated model estimates quantities such as voltage, current, and power at the maximum power point, together with internal electrical parameters. A long sequence of raw measurements can therefore be reduced to a compact description of one panel over a defined analysis window.

Step 1: prepare comparable telemetry

Operational telemetry is noisy. Measurements may arrive at slightly different times, and rapid changes in irradiance can make two otherwise healthy panels look different simply because they were observed under unstable conditions.

The first stage aligns measurements into consistent time-based records and filters out unsuitable observations. In the research dataset, this meant retaining physically plausible measurements from stable daylight periods and excluding points with abrupt changes in current or irradiance.

This filtering step is essential. A digital twin fitted to badly aligned or unstable data can explain measurement noise instead of panel behavior.

Step 2: calibrate the physical model for each panel

The parameters of the two-diode model cannot be read directly from routine plant telemetry. They must be estimated by solving a nonlinear optimization problem.

The method combines least-squares fitting with a genetic search. Least squares minimizes the difference between measured and modeled behavior, while the genetic search explores alternative parameter combinations and reduces the risk of settling on a poor local solution.

This produces a calibrated parameter set for each panel and analysis period. The model is then checked against electrical balance constraints, and only solutions within the accepted error range are retained. In practical terms, the process is:

1. Read and align telemetry.
2. Filter unstable and physically invalid measurements.
3. Fit the two-diode model to each panel.
4. Validate the model error.
5. Calculate expected maximum-power-point behavior.
6. Store the panel-level results for comparison.

The result is not one universal simulation of the plant. It is a fleet of panel-level twins calibrated from the conditions each panel experienced.

Step 3: normalize maximum-power-point output

Raw power is a weak anomaly signal because panels operate under different temperatures, irradiance levels, mounting conditions, and nominal tolerances. The digital twin provides an expected maximum-power-point value under comparable conditions, but the results still need a stable fleet reference.

The method calculates a normalized power deviation:

normalized deviation = (panel MPP power - reference MPP power) / reference MPP power

The reference is based on the median behavior of comparable panels. Median values are useful because a small number of badly performing panels cannot pull the reference as strongly as they would pull a simple average.

The comparison can be performed hierarchically:

• across the entire plant;
• within panels connected to the same inverter;
• within a single string.

This hierarchy helps distinguish a plant-wide shift from a problem associated with one electrical group or one physical area.

Static and dynamic anomaly detection

The same normalized signal supports two complementary types of analysis.

Static analysis asks whether a panel is unusually far from its peers during a defined period. In the research, a deviation of approximately 10% or more from the median-based reference was used to flag panels for investigation. Healthy strings generally remained within a much narrower performance band.

Dynamic analysis asks whether a panel's normalized behavior changed substantially from one period to the next. A panel may be consistently weaker than its neighbors, or it may show a sudden decline after months of stable operation. Those are different maintenance signals, and a single snapshot cannot distinguish them.

The threshold matters. A lower threshold detects more changes but also produces many more candidates for review. In the validation data, a 5% month-to-month threshold was overly sensitive, while a 10% threshold produced a smaller and more interpretable set of panel-level anomalies. Thresholds should therefore be calibrated to the installation and the cost of inspection rather than treated as universal constants.

What the method revealed

The approach was evaluated on seven months of operational telemetry from a plant containing hundreds of panels. Several findings were especially useful:

• Normalized maximum-power-point analysis identified substantially more abnormal panels than direct analysis of average power.
• Month-to-month analysis exposed changes that were not visible in a static comparison.
• Grouping results by inverter and string helped localize problems to a specific electrical or physical area.
• Some large deviations were caused by temporary shading rather than hardware failure.
• A persistently underperforming panel was later found to have physical damage.
• One faulty panel reduced the output of its series-connected group by roughly 1.5-2.0%, illustrating why a real fault can remain inconspicuous in aggregate plant metrics.

The research found that the normalized digital-twin method could identify roughly four times as many anomalously operating panels as direct average-power analysis in the evaluated dataset. Dynamic comparison added another layer of detection by identifying meaningful changes over time.

These results should not be read as a universal accuracy guarantee. They show that condition-aware, panel-level comparison can reveal information that is lost when monitoring stops at raw power or aggregate output.

A practical monitoring architecture

The method fits naturally into an existing telemetry pipeline:

1. Collect electrical and environmental measurements.
2. Build clean, synchronized panel-level records.
3. Run the digital-twin calibration on a scheduled window.
4. Calculate normalized maximum-power-point deviations.
5. Apply static and dynamic criteria.
6. Rank panels for inspection instead of issuing an automatic fault diagnosis.

That last distinction is important. An anomaly is evidence that a panel deserves attention; it is not proof of a specific defect. Shading, dirt, sensor errors, wiring differences, and physical damage can produce different forms of abnormal behavior. The twin narrows the search. Inspection and additional electrical tests establish the cause.

What transfers beyond solar

The broader pattern applies to many fleets of similar assets:

1. Model expected behavior under observed operating conditions.
2. Compare each unit with a robust peer reference.
3. Examine both current deviation and change over time.
4. Use system topology to localize the source of the anomaly.
5. Treat detection as decision support, not automatic diagnosis.

For solar plants, the practical lesson is straightforward: if monitoring only sees the aggregate, it will miss some of the failures that matter. A panel-level digital twin turns telemetry already being collected into a more useful signal for maintenance and investigation.

A separate article will cover a complementary neural-network approach to the same monitoring problem.