Soiling Assessment in Large-Scale PV Arrays

How much revenue is a soiled PV array losing, and at what point does it make sense to wash the array?

Owners, developers, bankers and O&M providers all want to know when it makes sense to clean a PV array to recapture revenue that it would otherwise lose due to soiled modules. On the one hand, an overly soiled array represents a loss of money. On the other, a premature cleaning represents a waste of money. While you must consider many variables to reach a definitive washing decision, the economics of module washing are not complex: If having a clean array saves more money than it costs to wash the array, then washing it probably makes sense.

This article shares some of our analyses and observations on array soiling drawn from many years of operational experience. We have had successes and failures, which have led to interesting discoveries and some dead ends. We have based most of our research on utility-scale PV plants with high dc-to-ac ratios in sunny, arid locations. These plants are subject to a unique set of circumstances: They spend a lot of time at full power, have relatively steady soiling rates and are rarely exposed to enough rain to significantly clean the modules.

Energy Recapture

It is difficult to assess soiling and to determine when to wash an array because doing so requires a multi-variable equation. Every analysis is unique, based on a host of project-specific mitigating factors such as technology choices, racking configuration, inverter loading, PPA rates, time-of-day profiles, interconnection agreements and so forth. This means that there is no single right answer when it comes to the economics of washing. The methods for soiling analysis are as varied as the business model behind the PV plant, so each solution uses a unique combination of people, tools and number crunching. What all effective soiling analyses have in common, however, is that they distinguish between percent soiling and percent energy loss due to soiling. While the former is easier to quantify, it may not correlate to unrealized revenue.

For the purposes of this article, we define percent soiling as the reduction of expected output power between soiled dc source circuits (modules, strings, arrays) compared to the same source circuits under clean conditions. In field terms, percent soiling describes the ratio of dirty to clean IV-curve traces, extrapolated to nameplate power under standard test conditions (STC). Meanwhile, we define percent energy loss due to soiling as the difference between the metered energy for a given time period compared to the energy that could have been harvested over the same time period with a fully clean array. This term describes the energy that is available for recapture, which correlates directly to unrealized revenue. To differentiate between these two concepts, we need to quantify the amount of time that a PV power plant spends at or near full power.

Power limiting in PV arrays. It is common practice to deploy PV systems with a high array-to-inverter power ratio in an attempt to capture more energy and revenue. As a result of these high dc-to-ac loading ratios, many inverters spend a lot of time operating at full power, which forces the array off its maximum power point.

Extended periods of power limiting result in a characteristic flat-topped power curve, which people commonly refer to as power clipping. The more time a PV system operates at full power, the less concern is warranted over soiling. Soiling abatement is effective only if you can recapture the lost energy, which requires unused inverter capacity. The returns are diminished in PV systems with chronically clipped power profiles, because an inverter operating at full power cannot increase its output power based on an incremental increase in irradiance. If soiling is viewed as an effective reduction in plane-of-array (POA) irradiance, then a 5% increase in irradiance can overcome a 5% soiling level. For example, if a given inverter hits maximum output at a POA irradiance of 800 W/m2 under clean array conditions, then it follows that power clipping will start at 840 W/m2 in the 5% soiled case. Above 840 W/m2, the percent soiling literally becomes a moot point.

Figure 1 illustrates this point by comparing seasonal POA irradiance and plant production curves for the same PV system. The flat-topped curves on the left, labeled “Day 1 (August),” illustrate how the plant operates at full power for extended periods of time under high POA irradiance typical of summer. The curves on the right, labeled “Day 2 (November),” illustrate how the array operates below full power all day long under partially overcast conditions in the autumn. To compare the percent energy loss due to soiling for Day 1 versus Day 2, we first have to filter out the time spent at full power, as no energy is available for recapture during these hours.

Table 1 presents these filtered results. Compared to baseline values for a clean array, the percent soiling is roughly the same on Day 1 and Day 2 (3.7% versus 3.6%). However, we can recapture energy only during hours when the PV plant is not power limiting. This leads to a slightly counterintuitive result: Even though the incident energy on Day 1 is nearly twice that on Day 2 (10.4 kWh/m2 versus 5.3 kWh/m2), the percent energy lost and the net energy lost due to soiling are greater on Day 2. This means that Day 2 presents the better opportunity for revenue recapture via washing, even though the available solar resource value is lower.

The challenge associated with soiling assessment is that we need to extrapolate this analysis to the near operational future for a PV power plant. The estimate concerning the future mix of clear, cloudy or overcast days is what determines the economics of module washing. A host of models and methods are available to predict and back-calculate the energy available for recapture, including hourly energy models, exceedance probability calculations and regression analyses. Regardless of the methodology used, you must account for inverter power limiting and have an accurate estimate of percent soiling.

Direct Soiling Measurements

The best way to estimate percent soiling is to measure it directly: Test the array, wash it, and test it again. While the process is time-consuming, there is no disputing the results. Soiling sensors and IV-curve tracers are proven tools for getting an accurate answer to the question “How dirty are my modules?” It is also possible to use other devices, such as short-circuit testers, to get a general estimate of soiling levels. Just keep in mind that additional data analysis and filtering is required to extrapolate from percent soiling to percent energy loss due to soiling.

Soiling sensors. Soiling sensors are essentially stand-alone evaluation tools that compare the actual output of a naturally soiled PV reference module to the expected output of a clean PV reference device. Some soiling sensors use short-circuit current (Isc) as the basis of comparison; others incorporate a microinverter and compare maximum power point values (Vmp, Imp, Pmp); some devices use a hybrid technique that compensates for temperature and normalizes results to STC. All of these approaches yield a high-quality data stream that you can easily use to assess the soiling level of the modules in the test rig.

IV-curve tracers. To get the best possible in situ soiling measurements, put a good IV-curve tracer in the hands of a competent technician. Curve tracing is slow but definitive. You can compare PV source-circuit curve traces to STC or use a dirty versus clean approach. As long as technicians capture a representative set of IV-curve traces under roughly the same conditions, the results of the study will be accurate and useful. While it is quick and easy to analyze these IV-curve data, it is incumbent on the technicians to choose representative strings to test in the field.

Other devices. Another option that works well is to use instruments that measure short-circuit current or operating current, or that can extrapolate measured data to a baseline condition—such as PVUSA Test Conditions (PTC) or STC—to estimate percent soiling. Since these devices are not explicitly intended to perform soiling measurements, the correlation process is left to you. However, the process does not need to be complex. A simple multimeter with a current loop sensor is sufficient to get a general idea of soiling conditions. If necessary, you can assess soiling with a Fluke meter, a few gallons of water and a squeegee.


Soiling stations, IV-curve traces and other assessments that compare “before” (dirty) and “after” (clean) conditions give an excellent indication of the soiling conditions on a specific set of modules or test array. The trick is to take data from these devices and extrapolate it twice: once to generalize the entire plant’s soiling condition, and once more to infer how much the measured soiling will affect energy production or performance. We call this the soiling transfer function. Direct soiling measurement is a great start, but it is a rare instance where the estimated percent soiling value will reflect an equal (or even proportional) percent decrease in production. As illustrated in Table 1, percent soiling does not correlate directly to energy lost due to soiling when PV plants spend a lot of time operating at maximum power.

To complete the soiling transfer function from percent soiling to percent energy loss due to soiling, you need to filter the operational data strategically. The data filtering process can be as simple as removing power clipping points, which has the effect of constraining the evaluation to periods of MPPT operation. You can also apply additional filters to remove spurious data points that may muddy the results, such as measurements associated with low POA irradiance, unstable irradiance or excessive wind speeds. Once you have obtained field measurements and filtered the operational data, you just need something with which to compare these to estimate percent energy loss due to soiling.


The best way to estimate the impact of soiling is to compare operational data to plant performance under clean conditions, which we refer to as the plant baseline. Obtaining a performance baseline is a process of characterizing the electrical performance of source circuits, combiners, inverters or an entire plant and isolating these data for frequent comparison. The goal of establishing a baseline is to understand how the system or subsystem performs under known operating conditions when the array is free of faults and unsoiled. Generally speaking, a rough plant baseline is good enough.

Establishing a clean plant baseline is more of a process than an event. The logical opportunity to obtain a baseline for an entire plant is at the time of initial back-feed, testing and commissioning. If you want to get two detailed answers at once, you can perform a full-plant baseline characterization in parallel with performance testing, which is ideal. However, you can establish a baseline at any system level, over any duration of time and under any operating conditions. Nothing is lost if you are unable to characterize some parts and pieces at commissioning. You can always revisit and recalibrate these parts later and make sure that they fit the general performance trend once they are up and running. As long as you restore malfunctioning blocks to operation and characterize their performance using the same measurement methods, the baseline will be accurate and useful despite its piecemeal assembly.

There are various means of applying the baseline. The simplest form—comparing dirty versus clean performance—is effective for both long- and short-term analyses. By characterizing the plant according to its big pieces, such as inverters, skids or ac collection circuits, you can compare these results to one another, normalize dirty results against the clean baseline and make informed decisions about soil abatement. You can express the baseline in whatever terms best suit your goals, such as specific yield (kWh/kW) or energy output in relation to POA irradiance. The latter is useful if you need to tie actual performance back to expected performance based on an energy model.

Since assumptions, data resolution and as-built conditions constrain energy models, we strongly recommend that you use operational data rather than modeled plant behavior as the basis of comparison. Whereas an energy model describes how the plant is supposed to behave, measured data describe how the plant actually behaves. In broad terms, energy modeling software applies soiling assumptions as an effective monthly reduction in POA irradiance and essentially stops there. One-month averages for soiling levels can shore up production and revenue models, but they have little to say about soiling events, differential energy impacts or soiling rates in general. As a result, the input/output resolution for an energy model is far less precise than it is for most operational datasets.


End use and accuracy drive the baseline characterization method. Production losses can be very subtle, typically only a few percentage points, before they become noticeable, so accuracy is vitally important to tying production losses specifically to soiling.

The simplest characterization method is to catalog plant production at the meter as well as measured irradiance in the plane of array. Since this obviously ignores thermal differences within the array, for increased accuracy you may need to apply a temperature compensation to account for deviations from weather station conditions. You also need to remove or ignore performance issues that are not related to soiling, such as module degradation, equipment failures and configuration differences. Soiling analysis has to quantify or transcend these factors to reach a reasonable conclusion.

To illustrate the challenge: A POA irradiance sensor might have an accuracy of ±1.0%; ac power measurement transducers are typically accurate within ±0.2%; dc transducers are rarely better than ±1.0% accurate; secondary measurements, such as temperature and wind speed, have ±2% accuracies at best. These measurement errors typically compound rather than cancel one other. Compounded, these uncertainties suggest that isolating a few percentage points of performance loss using gear with measurement errors of a few percent can produce dubious results.

The net result is that a thorough soiling analysis could very well estimate that modules are 4.5% soiled, plus or minus 2%. Given these uncertainties, module washing may or may not be cost effective. While no one likes this type of answer, it is often the case that soiling analysis results have a high degree of uncertainty.

Practical Application

We recommend a relatively simple five-step approach for isolating the effects of soiling on energy production based on measured data from operating PV plants. The methodology uses a comparison to a baseline as a means of assessing the production that the array might have achieved if it had been completely clean and operating perfectly. The specific implementation of this methodology depends on plant type, capacity and the monitoring solution. However, you can apply this method at almost any plant level using similar techniques.

Step 1: Catalog all IV-curve traces and other string-level commissioning tests to establish source-circuit behavior with respect to nameplate power. This step provides a consistent reference dataset that you can revisit when using periodic string testing for performance assessments.

Step 2: When commissioning the array and conducting energy performance tests, establish plant-level and inverter-level baselines using high-resolution data. These baselines should isolate trend data for clipping and nonclipping production as a function of POA irradiance and should be normalized to dc capacity by inverter. You can complete this step in pieces, if need be, updating the baselines as more datasets become available. The key is to characterize a clean, fully operational plant.

Step 3: Track plant performance using trend data from the time of (clean) commissioning through operations. Using the same filters employed to establish the baseline, determine approximate soiling levels while the plant operates (as time, data and weather allow). 

Step 4: If you suspect excessive soiling, perform a series of string-level field measurements before and after washing, and compare these results to the commissioning data. Next, compare these measured results to the soiling estimates generated from trend data with the appropriate clipping filters applied. Establish the correlation between the measured and modeled results for future use.

Step 5: When field measurements and data analysis align—and when the comparison to baseline indicates that energy recapture will be cost effective—then it is time to schedule a wash. Over time, take advantage of these full-array washing opportunities to recalibrate the baseline, the energy model and so forth.


The following examples illustrate how you can use baseline comparisons to isolate soiling conditions. We have taken all examples from utility-scale plants with multiple central inverters in sunny, arid locations. We have summarized and annotated each case to show how you can apply the same methodology at various scales.

Plant level. Figure 2 shows an example of a long-duration soiling analysis. We cataloged these data over an 8-month period, and they capture a few isolated rain events as well as a complete array cleaning. We have filtered the datasets from each for clipping and reported them as percent of baseline. Although these daily values have quite a bit of variance and error, the soiling accumulation trend is undeniable. While the rain events mitigated soiling only marginally, the wash effectively rehabilitated the arrays to full potential. 

With any macro-level assessment, especially on larger plants, you must level out or ignore some asymmetries and performance issues with strategic math. The end result is an accurate model of how the plant turns photons at the modules into energy at the meter. You can parse this type of baseline into subsections, perhaps by combiner, inverter, skid or ac collection circuit. Regardless of the scale, the concept is the same and provides an adequate assessment of performance in an ongoing manner. You can employ and repeat this dirty versus clean comparison to baseline under any circumstance and recalibrate the whole process after a full array cleaning.

Inverter level. Inverter-level assessments are a subset of whole-plant characterization but with higher data resolution. The key to this level of analysis is to establish a unique baseline for each inverter under clean and fully operational conditions. Inverter-level comparisons are useful for identifying the impacts of differential soiling across the whole plant.

For example, Table 2 compares inverter-level data, reported as “percent inverter-specific energy compared to baseline,” for a large-scale PV plant with differential soiling. Most, but not all, of the arrays at this site are subject to rapid soiling from an adjacent road and farm field. By tracking inverter-level data, we can isolate soiling by location or overall contribution to lost energy. In this particular case, the soiling was profound enough to trigger a full wash cycle. If the differential soiling analysis had indicated that soiling affected less of the plant overall, we could have focused our maintenance activities more selectively, perhaps electing to wash only arrays associated with specific inverters.

Combiner level. We can further increase data granularity and resolution by evaluating dc input current at the subarray level, which effectively facilitates combiner-level assessments. While this approach makes it easy to diagnose the effects of differential soiling on an individual inverter, the real beauty of combiner analysis is that it provides a built-in method of validation. If all of the subarray inputs are showing the same thing, as in Figure 3, our confidence in soiling assessments improves. The increased granularity also makes it easier to track incremental changes from the baseline.

String level. Because it provides the highest-resolution data possible, string-level analysis is the alpha and the omega—the first step and the final step—of an effective performance assessment. Since most large-scale PV systems do not have string-level monitoring, cataloging source-circuit performance generally requires field tests. Though string-level testing demands high-quality tools and competent technicians, the data produced are effective for establishing a baseline or calibrating the energy metrics and assumptions used at all other levels of analysis.

You can use these string-level data to calibrate independent soiling sensors. You can also apply string-level dirty versus clean results, such as those shown in Figure 4, to historical data or to a before-and-after cleaning analysis. In this figure, the raw trace data, based on in situ irradiance, are shown in green; the curves in red correct these field measurements to STC; the blue curves, meanwhile, show the ideal I-V curve for the source circuit at STC. These dirty versus clean traces provide a good indication of the energy available for recapture at the string level, which we can extrapolate to larger performance blocks.

An ideal use for field measurements is to calibrate soiling analyses in relation to operational data. This process involves comparing IV-curves to soiling station data and other soiling metrics. To the extent that we can draw correlations, we can triangulate these datasets and better inform our washing decisions. This process of continuous improvement is essential to effective soiling assessment.


Dust storms, intermittent construction activity, unusually heavy traffic and sporadic agricultural activity are examples of event-based soiling. When soiling gets very bad—or when it gets a lot worse in a hurry due to a soiling event—strange things start to happen in terms of plant behavior. Module soiling can reach a point where the fundamental electrical characteristics of the dc array change dramatically, so much so that it sometimes forces inverters out of maximum power point tracking. These results are most common in neglected PV plants where extreme soiling causes blocking diodes in the modules to engage, which can completely confuse the inverter.

Really bad soiling almost precludes analysis. The electrical behavior of a PV plant becomes less predictable and performance suffers, but it can be difficult to quantify how bad the problem is and how much energy the plant is losing. Such conditions combine significant energy shortfall with chaotic behavior. While we can measure the lost energy, we cannot directly discern the reasons for the loss. This complicates the process of troubleshooting any problems not related to soiling.

Soiling events are a constant source of panic. Everyone wants to know how bad the problem is, but making even a rough estimate takes at least a day. Rather than rushing to get a washing crew in place based on incomplete information, the best approach to soiling events is to send technicians to the site to assess the problem via dirty versus clean testing. These strategic test results will quickly provide the answers needed and frequently trigger a wash cycle.

Soiling events can also be localized, a situation we call asymmetrical soiling. This occurs when some arrays get a lot dirtier than others. Exterior arrays next to dirt roads or agricultural activity are the most common culprits. Differential soiling across the whole plant skews bulk numbers, especially when you take the soiling assessment measurements from a relatively clean or dirty array.

Since soil detection is intended to generalize soiling conditions, you cannot trust the numbers it yields when you are adapting a general model to an asymmetrical problem. We call this phenomenon forced mismatch, meaning that uneven soil deposition creates an imbalanced electrical condition. Here again, the best response is to send out a crew to assess the situation, and then back up the findings by comparing filtered operational data to a clean baseline. Asymmetrical soiling may make selective module washing a viable option.


The next case studies represent rigorous analyses using high-resolution data applied to fully operational plants that all ended up with dubious results. Some may call these war stories; we call them analytical head-scratchers. We present them here to illustrate the chaotic nature of soiling measurements and the unpredictability of the results.

Case 1. After measuring overall soiling of a PV plant at around 4%, the owner scheduled washing. Before the wash, a short-duration rain event occurred, so the owner asked us to investigate to see whether the rain had cleaned the modules enough to justify delaying the capital expense of a full wash. By our calculations, the rain event actually increased soiling to more than 5%, calling the entire chain of decisions, as well as our analytical approach, into question.

Case 2. In an attempt to quantify soiling, we conducted a series of before-and-after IV-curve traces across a plant. Our strategic plan called for washing selected strings of modules across a representative set of arrays on assorted inverters to quantify a measurable difference. The curve traces showed less than 1% soiling on some strings and more than 7% on others, with a relatively even distribution between these extremes. We recommended a full cleaning, and the net performance results after washing showed a similar distribution of results. However, the overall performance increase was only about 33% of the expected result, netting a 1.9% increase in production. We had a hard time trusting the results, the analysis approach and the wisdom of our recommendation to wash.

Case 3. Cleaners fully washed a plant at night to prevent production losses, which is a reasonable approach. The next morning, while the modules were still cool and wet, the farmer on the upwind side of the plant starting tilling fields, which spread a thick dust cloud onto an otherwise clean array. In this case, unforeseen farmwork forced another wash cycle.

These case studies illustrate that attempts to isolate the effects of soiling can be elusive. Soiling effects are design dependent; geographically varied; simultaneously localized and vastly different between arrays; dependent on geometry, orientation and array racking configuration; and variable based on the weather or off-site activities. In addition, rain does not necessarily clean modules very well, if at all. These factors are not necessarily bad news. Rather, they are limiting assumptions that you need to categorize, isolate, quantify and remove from the analysis to begin a valid assessment. Once you accept that soiling is a chaotic phenomenon, you can begin to see patterns and to learn from the more predictable parts of the problem.


Sanjay Shrestha / SOLV Performance Team / San Diego, CA /

Mat Taylor / SOLV Performance Team (retired) / San Diego, CA /

Article Discussion