This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE JOURNAL OF PHOTOVOLTAICS 1

Review of Statistical and Analytical Degradation

Models for Photovoltaic Modules and Systems
as Well as Related Improvements
Sascha Lindig , Ismail Kaaya, Karl-Anders Weiß, David Moser , and Marko Topic

Abstract—In this work, we investigate practical approaches of applications until the 1970’s and 1980’s. In the beginning of
available degradation models and their usage in photovoltaic (PV) the 1970’s, the dramatic price increase for fossil fuels and an
modules and systems. On the one hand, degradation prediction energy uncertainty because of the oil crisis raised the aware-
of models is described for the calculation of degradation at system
level where the degradation mode is unknown and hence the physics ness for a need of change in the energy supply [1]. Since then,
cannot be included by the use of analytical models. Several statis- the interest in renewable energies and solar energy in particular
tical models are thus described and applied for the calculation of has increased, which in turn led through scientific achievements
the performance loss using as case study two PV systems, installed to a steady reduction in installation costs and performance im-
in Bolzano/Italy. Namely, simple linear regression (SLR), classical provements of terrestrial PV systems. In the late 1990’s, the
seasonal-decomposition, seasonal- and trend-decomposition using
Loess (STL), Holt–Winters exponential smoothing and autoregres- first large-scale PV systems were installed [2]. Nowadays, PV
sive integrated moving average (ARIMA) are discussed. The per- module manufacturers guarantee a performance reduction of no
formance loss results show that SLR produces results with highest more than 20% within 25 years of operation at standard test
uncertainties. In comparison, STL and ARIMA perform with the conditions (STC) (modules tested indoor under TSTC = 25 ◦ C,
highest accuracy, whereby STL is favored because of its easier im- GSTC = 1000 W/m2 , AM 1.5) and even started to guarantee a
plementation. On the other hand, if monitoring data at PV module
level are available in controlled conditions, analytical models can be maximum degradation of 1%/year for the first ten years. Never-
applied. Several analytical models depending on different degrada- theless, the actual performance throughout the lifetime is quite
tions modes are thus discussed. A comparison study is carried out uncertain and unpredictable. Laboratory or field determination
for models proposed for corrosion. Although the results of the mod- of PV modules service life under real environmental conditions
els in question agree in explanation of experimental observations, requires an unacceptable length of time.
a big difference in degradation prediction was observed. Finally,
a model proposed for potential induced degradation was applied PV systems are affected by continuous cycles of temperature,
to simulate the degradation of PV systems maximum power in humidity, irradiation, mechanical stress, and soiling. These en-
three climatic zones: alpine (Zugspitze, Germany), maritime (Gran vironmental mechanisms cause different degradation modes to
Canaria, Spain), and arid (Negev, Israel). As expected, a more take place within a PV module and reduce the performance
severe degradation is predicted for arid climates. of the system. Therefore, it is necessary to develop diagnostic
Index Terms—Degradation models, performance loss, photo- techniques, lower the performance uncertainty, and predict the
voltaic (PV) modules, PV systems, service life prediction. behavior of PV systems with higher accuracy.
Commonly two approaches, statistical and analytical meth-
I. INTRODUCTION ods, are used for evaluating degradation rates of PV modules
and systems. This report describes quantitative degradation and
ECAUSE of high costs and limited efficiencies, photo-
B voltaic (PV) applications were exclusively used for space
service lifetime models currently used for PV modules and rec-
ommends further improvements. A review of available models
and improvements is crucial for accurate life-time calculations
Manuscript received June 1, 2018; revised July 27, 2018; accepted September of future energy PV systems. The first part of this work focuses
5, 2018. This work was supported by the European Union’s Horizon 2020 on available metrics of variables and the most common statis-
Program under GA. No. 721452—H2020-MSCA-ITN-2016. (Corresponding
author: Sascha Lindig.) tical models to retrieve the performance loss based on these
S. Lindig is with the Institute for Renewable Energy, EURAC Research - Viale metrics. The second part deals with analytical models, which
Druso 1, 39100 Bolzano, Italy, and also with the Faculty of Engineering, Univer- pinpoint specific degradation modes and their possible impact
sity of Ljubljana, 1000 Ljubljana, Slovenia (e-mail:,[email protected]).
I. Kaaya and K. A. Weiß are with the Fraunhofer Institute for Solar Energy on the performance of PV systems.
Systems, 79110 Freiburg, Germany (e-mail:, [email protected]; We believe that a more precise prediction of PV sys-
[email protected]). tem performance and the capability of linking performance
D. Moser is with the Institute for Renewable Energy, EURAC Research -
Viale Druso 1, 39100 Bolzano, Italy (e-mail:,[email protected]). losses to relevant degradation modes will increase public trust
M. Topic is with the Faculty of Engineering, University of Ljubljana, 1000 in solar energy. Additionally, it will help stakeholders such
Ljubljana, Slovenia (e-mail:,[email protected]). as investors, PV plant owners, operation and maintenance,
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. and insurance companies as well as other parties involved
Digital Object Identifier 10.1109/JPHOTOV.2018.2870532 to favor more beneficial and accurate business models and
2156-3381 © 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution
requires IEEE permission. See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE JOURNAL OF PHOTOVOLTAICS

to more efficiently operate and maintain PV systems in the Models for degradation are generally either data-driven
future. or derived from physical principles via stochastic processes.
Although data-driven models are more commonly applied to
II. PV MODULES DEGRADATION MODES analyze degradation data, viewing degradation through stochas-
tic processes helps researchers to theoretically characterize the
Degradation modes are effects that irreversibly degrade the
degradation process. Therefore, a coupling of both models could
performance of a PV module/system or may cause safety prob-
enhance the knowledge of what is happening in PV systems.
lems [3]. A great number of different degradation modes are
Data-driven models help to examine the overall performance
observed in PV modules, both under outdoor operation and in-
loss of a system over time and by using analytical models
door testing. The most commonly observed degradation modes
conclusions of what triggered these losses might be derived.
include [3]: light induced degradation (LID), solder fatigue
failure, silver grid finger delamination, bypass diode failure,
delamination, cell cracks, corrosion, polymeric discoloration, A. Data-Driven Models
ultraviolet (UV) degradation of the cell, polymeric mechani- Data-driven models are often empirical employed to estimate
cal failure, and potential induced degradation (PID). Each of degradation rates based on statistical analysis of given data sets.
these degradation modes has different causes and is triggered The goal of the statistical analysis is to calculate the trend of
by different stress factors. Apart from the modes listed, differ- the PV performance time-series and to translate the slope of the
ent technical risks, which affect the PV performance and the trend to an annual loss rate, in units of %/year [7]. Although these
resulting costs, were found by Moser et al. [4]: glass breakage, models can provide consistent performance loss rates (PLRs),
snail track, defective backsheet, hotspot, soiling, overheating, which are useful for data extrapolation and service life predic-
and failure junction box. An occurring degradation mode can tions, they do not directly provide evidence for the degradation
have an increasing impact on the PV performance over time. It modes taking place in the module. Other effects such as diffuse
develops either in isolation or in combination with other degra- soiling, snow, shading or module mismatch have also a direct
dation modes or technical risks and might lead to the failure of impact on the performance trend. Therefore, it is more accurate
a PV module. The term failure for electrotechnical devices is to talk about a PLR rather than a degradation rate.
defined as “the termination of the ability of an item to perform
a required function” [5]. While this definition serves as a clear
B. Analytical Models
guideline for most devices, the failure of a PV component is
somewhat more complex. For example, although a PV module Analytical models are based on the physical/chemical theo-
can still be technically usable, its power output might be too low ries of a specific degradation mode. These models represent the
to verify the continuation of its operation from an economical mechanism involved in complex physical/chemical processes.
point of view. Within the scope of this work, a failure is defined For well-known PV module degradation modes, several analyt-
as the necessity to replace a PV component, because of its ulti- ical models to forecast PV module degradation are available.
mate, economic or safety-related failure. A clear understanding All these models are based on the principle of understanding
of the definition of a degradation mode is also still a challenge the underlying process, but they are still only heuristic models,
and stress should be put on common nomenclature to define which do not include the influence of material parameters.
the same degradation mode with the same terminology. For the In the following chapters, we discuss the most commonly
moment, accelerated aging tests are being utilized in the study used performance loss models.
of some of these modes. However, there is no proof/evidence
that the results from these tests reflect what exactly happens to IV. PERFORMANCE LOSS MODELS
the modules in outdoor conditions.
Before applying any statistical model, the observed data are
generally treated using filtering techniques depending on factors
III. DEGRADATION MODELING like irradiance or standard deviation ranges and subsequently
Degradation models are used to relate a test item’s estimated averaged or added up over certain time periods. This step is
failure time with the wear and tear during its usage period. performed as data preparation to minimize outliers and noise
The failure time is defined as the end of the lifetime of a PV and to remove values corresponding to inhomogeneous irradi-
component because of its failure. Degradation models help to ance conditions on the irradiance sensor and the PV system [8].
quantify the performance loss PV modules and systems are Afterwards, a performance metric can be applied to a pretreated
experiencing under operation. Degradation in PV systems is the data set and the PLR is calculated by using statistical methods.
reduction in efficiency with which a PV system is converting These steps are necessary to minimize seasonal oscillations and
light of the sun into electricity over time [6]. This appears at to eliminate outliers resulting in the reduction of the overall un-
all levels of a PV system, be it at cell, module or system level. certainty in the estimation of the PLR. In the following, a short
To model PV module degradation modes, the knowledge of overview of the most common performance metrics as well as
internal loads like temperature, chemical conditions, irradiance, statistical methods is presented. In Section IV-C, a comparison
and mechanical loads in/on the PV module is required. One very of the statistical models in question is performed on a case study
important part is to convert external loads to internal loads of with data of two PV systems installed at the airport of Bolzano
the module. in Italy.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LINDIG et al.: REVIEW OF STATISTICAL AND ANALYTICAL DEGRADATION MODELS FOR PHOTOVOLTAIC MODULES AND SYSTEMS 3

A. Performance Metrics Tmod − TSTC are considered. An average performance model for
each PV type is considered and the k-coefficients are calculated
Statistical performance loss models need to be applied on cer-
tain PV system performance-rating parameters. The parameters using data from different modules of the same PV technology
[10], [14]. This model creates a matrix instead of a single well-
are expressed through performance metrics, which are mea-
defined value for the maximum power point.
sured or calculated in a specified interval. Performance metrics
are ideal to compare the performance of different systems in Within the PVs for utility scale and applications project, an-
other widely used model, PVUSA, was developed [15], [16].
different climates. They can be categorized into three different
groups. These are: 1) electrical parameters directly taken from While calculating corrected power values, it is assumed that
I–V curves recorded either outdoor or indoor and corrected to the PV system current primarily depends on irradiance and the
voltage on module temperature Tmod . Tmod in turn is strongly de-
STC, 2) empirical metrics such as PVUSA [9], the 6-k-values
performance model (applied in the PVGIS online tool) [10] or pendent on ambient temperature, irradiance, and wind speed. A
the Sandia models [11], and 3) normalized and/or corrected met- regression of the systems maximum power output is performed
against PVUSA test conditions (PTC) by
rics such as performance ratio [7]. Great care has to be taken
when selecting performance metrics. The choice as well as pos- PMPP = GPOA (A + BGPOA + CTam + DuW )
sible corrections such as corrected power for temperature and
irradiance [12] will influence the results. Therefore, the out- (GPTC = 1000 W/m2 , TPTCA M = 20 ◦ C, uW = 1 m/s). (2)
come of a certain performance loss model applied on a specific First, measurements at high irradiance values (G ≥ 800 W/m2 )
performance metric needs to be evaluated and put into context in the plane of array (POA) are selected and fitted to calculate
to understand the validity of the results. monthly values for the coefficients A, B, C, and D, applying
1) Evaluation of I–V Curves: Electrical parameters of the multivariate regression. Afterwards, the coefficients are used to
I–V curve include power, voltage, and current at the maximum receive monthly ratings at PTCs (substituting meteorological
power point, the open circuit voltage, and the short circuit cur- data values). It should be noted that this methodology is op-
rent. With these parameters, it is possible to calculate the fill timized for crystalline silicon PV. An adapted version of the
factor. A PV systems performance loss is observable when com- equation including another coefficient E was developed to con-
paring the values of periodically performed measurements of sider thin-film technologies [7], [9].
systems in operation [13]. If an unexpected decline in one of 3) Normalized and Corrected Metrics: Normalized and/or
the parameters appears, the affected modules can be examined corrected metric parameters are useful when comparing dif-
indoors or outdoors, for example, with electroluminescence or ferent PV technologies in different climates. Here, PV system
thermal imaging cameras. For detailed characterization, indoor performance data are either normalized to comparable, unit-free
measurements can lead to the most accurate results. However, metrics or corrected in respect to outdoor conditions. One of the
removing PV modules from the field is time consuming with most commonly used metrics is the performance ratio (PR),
the possibility of damaging the modules during transport and which is an adequate indicator for the quality of a PV installa-
handling. These considerations need to be taken into account tion. The PR is calculated by dividing the final (or array) yield
to decide which strategy to follow depending on the extent and Yf (a) (depending if ac- or dc-power is evaluated) with the ref-
the complexity of the detected problem and which stakeholders erence yield Yref [17]. The yields are ratios of measured values
are involved. For example, for insurance claims, outdoor mea- of power or irradiance with values obtained under STC
surement may be sufficient while certified indoor measurements
could be required for PV module warranty claims. Yf PAC /PSTC
PR = = (3)
2) Empirical Metrics: The empirical metrics presented are Yref GPOA /GSTC
models, which aim to obtain performance data while taking Ya PDC /PSTC
into account the dependence between the PV system output and PRDC = = . (4)
Yref GPOA /GSTC
prevailing outdoor conditions [9]. The idea is to receive PV
system performance parameter like the efficiency or the maxi- When studying the PV performance, it is advisable to use dc-
mum power through the application of formulas, which consist related performance metrics in order to eliminate possible influ-
of empirical coefficients and weather data. Two widely used ences because of inverter degradation or misbehavior.
models are the 6-k-values performance model and PVUSA. A promising correction method, presented by Belluardo et al.
The 6-k-values performance model describes system perfor- [18], evaluates the irradiance and temperature corrected power
mance through the relative efficiency ηREL correlated to STC as under STC conditions as follows:
a function of in-plane irradiance GPOA and module temperature GSTC 1
PT ,G corr = Pmax (5)
Tmod [14] G 1 + γ(Tmod − TSTC ).
ηREL (G , T  ) = 1 + k1 ln(G ) + k2 ln(G )2 + k3 T  Here, γ is the temperature coefficient of the PV systems power
    2 2 at STC, which is stated on the datasheet. Since γ is retrieved
+ k4 T ln(G ) + k5 T ln(G ) + k6 T . (1)
at 1000 W/m2 and highly temperature dependent, a preliminary
Equation (1) has to be fitted to experimental data to obtain the data filtering, similar to the filter applied in PVUSA, should be
empirical coefficients k1 − k6 . Hereby, the normalized in-plane performed to assure the accuracy of the temperature coefficient
irradiance G = G/GSTC and the normalized temperature T  = in use.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE JOURNAL OF PHOTOVOLTAICS

Methods, which correct absolute values like power, voltage equations below
or current, can be additionally normalized by dividing the cor-
rected value by the nominal installed value under STC. This step ŷ = Tt + St + et , ŷ = Tt × St × et . (7)
simplifies a possible comparison between different PV systems. Here, T is the trend, S the seasonality, and e the remaining part
A clear advantage of these rating techniques is the possibility to of the data [7].
evaluate the performance loss in any desired time resolution. 3) HW Seasonal model: The HW seasonal model contains
a forecast equation and three smoothing equations as shown
B. Purely Statistical Methods below
Statistical analysis methods are used to retrieve trends of per- ŷt+1|t = lt + bt + st−S +1 (8)
formance time-series. These time-series are some sort of perfor-
mance metrics, which are discussed in Section IV-A. The slope lt = A(yt − st−S ) + (1 − A)(lt−1 + bt−1 ) (9)
of a trend function can be interpreted as the PLR. It is possible bt = B(lt − lt−1 ) + (1 − B)bt−1 (10)
to accumulate these ratings for any given time resolution into an
easily comparable annual aging value. The difficulty is to find st = C(yt − lt−1 − bt−1 ) + (1 − C)st−S . (11)
a good estimation of the PLR since the application of a certain Here, lt is the level, bt the slope, and st the seasonal component.
statistical method on a performance metric and a defined filter A, B, and C are smoothing parameters. If monthly data are
determines the result significantly. Statistical analysis methods evaluated, the period of seasonality, S, equals to the value of 12.
can be divided into model-based methods like linear regression, The HW model is either additive or multiplicative, depending
classical seasonal decomposition (CSD), Holt–Winters (HW) on the seasonal behavior. In case of evaluating a PV systems
exponential smoothing or autoregressive integrated moving av- performance, the additive method should be selected because
erage (ARIMA) and nonmodel-based methods such as seasonal the seasonal variations are approximately constant throughout
and trend decomposition using Loess (STL). In the following, the series. The seasonal component is then computed in absolute
these commonly used methods are described. terms and has a mean of around zero. The level equation (9) is a
1) Simple Linear Regression (SLR): Performance metrics of weighted average between the seasonally adjusted observation
any kind are most commonly applied on linear regression be- (yt − st−S ) and the nonseasonal one-step-ahead forecast (lt−1 +
cause of the straight-forward approach. The fitted trend line is bt−1 ). The slope is a weighted average of the level at time t
given by minus the level at t − 1, and the trend at t − 1. The selection
of smoothing parameters determines how fast the exponential
ŷ = at + b. (6)
weights decline over the past observations. The HW method can
Hereby, a represents the gradient and b is the intercept with the be especially useful for computing the future behavior of a PV
y-axis. The SLR-algorithm uses the method of least squares. The system [7], [20], [21].
idea of this method is to sum up squared values of the difference 4) Autoregressive Integrated Moving Average: ARIMA is a
between trend line and actual measurement points and to find model, which can contain several methods in a multiplicative
the minimum value for this sum. Squares are used to add up only way and can be described as ARIMA (p, d, q) (P, D, Q). Here,
positive numbers and to put more weight on more widely scat- p is the auto-regressive, d the differencing, and q the moving
tered residuals. This method overemphasizes outliers as well as average order as well as P is the seasonal autoregressive, D
seasonal variations and can result in large uncertainties. Because the seasonal differencing, and Q the seasonal moving average
of that, performance metrics, which reduce seasonal oscillation order. Due to the flexibility of the model, seasonal variations,
should be applied if the SLR-algorithm is used. errors, outliers, and level shifts can be addressed in a proper
2) Classical Seasonal Decomposition: Another commonly way. ARIMA is applied using the following [7]:
used statistical model is CSD. By using CSD, the seasonality φ(T )φS (T S ) d D S
S yt = φ(T )φS (T )et (12)
and a certain irregular component are separated from a set of
measured time-series data to receive a clear trend over time. T is the delay operator, φ(T ) = (1 − φ1 T − · · · − φp T p ) is an
This technique helps to get a fast idea of a performance loss autoregressive polynomial in T of degree P, φ(T S ) is an au-
of the system in question. The trend is obtained by applying a toregressive polynomial in T S of degree PS , φ(T ) a mov-
centered moving smoothing on a time-series with a certain sea- ing average polynomial in T of degree q, and φS (T S ) is a
sonal period m. When using monthly data, the seasonal period is moving average polynomial of degree QS in T S . Apart from
usually set to 12. Here, the first value is computed by averaging that, d = (1 − T )D is a nonseasonal differencing operator
over the first 12 months. Due to the 12-month centered moving and D S = (1 − T ) is a seasonal differencing operator and
S d

average, 6 months at the beginning and 6 months at the end of grasps nonstationarity in the relevant location in consecutive
the observation period are not included in the computation. To periods [22].
calculate the seasonality, the trend is subtracted from the mea- The stationarity of the time-series determines the optimal
sured data and each month throughout the years of surveillance ARIMA model; a transformation using differencing to achieve
is averaged. What remains at the end is an irregular component stationarity might be indispensable. Stationarity is described by
[19]. Depending on the stability of the seasonal component, a constant mean and variance, resulting in a nonexisting trend
an additive or a multiplicative model is used as shown in the and the graph seems more like white noise. There are different
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LINDIG et al.: REVIEW OF STATISTICAL AND ANALYTICAL DEGRADATION MODELS FOR PHOTOVOLTAIC MODULES AND SYSTEMS 5

ways to difference a time-series, the simplest and most common is done by applying local regression on a data window with a
way being first-order differencing [23] certain width. The regression curve is fitted to the data within
the window. The closer the points are to the center of the win-
ŷ = yt − yt−1 . (13)
dow (higher weight), the greater is the impact on the regression
Here, the differenced value is the change between two consecu- line calculation. The weight is reduced on those points that are
tive values of the original time-series. The resulting time-series furthest from the regression curve. The whole step of regression
has T-1 values. Seasonal or second-order differencing are further and weighting is repeated several times to receive a point on the
examples of how to create stationarity within the time-series in Loess-curve, which is at the center of the window. By moving the
question. window across the data, the complete Loess curve is computed.
The heart of the ARIMA model is the application of autore- What follows is that each point of the Loess curve is the intersect
gression. To perform an autoregression, the desired variable is of a regression curve and the center of the respective window.
computed by applying a linear combination of past values of the
variable. The general form of an autoregressive model of order C. Comparison of Statistical Models
p is
The statistical models presented are applied on the uncor-
ŷ = c + φ1 yt−1 + φ2 yt−2 + · · · + φp yt−p + et (14) rected performance ratio data sets of a mono-crystalline (mc-
Si) and an amorphous silicon (a-Si) system using the software
where c is a constant and et is the remainder.
R. The mc-Si system contains 14 PV modules and has a rated
The moving average model used within ARIMA has a dif-
power of 1960 Watt-peak (Wp). The second installation includes
ferent purpose than the one for CSD. Here, the moving average
12 amorphous silicon modules with a power of 1200 Wp. The
uses past forecast errors in a model similar to a regression. The
observed data have a resolution of 15 min and were averaged
aim of the moving average model is to predict a forecast instead
over whole months. The systems in question were installed in
of smoothing the trend cycle of past values [21].
Bolzano/Italy in August 2010 and are evaluated for seven years
While using the software environment R, the function seas, a
from March 2011 until February 2018 in order to eliminate
function within the R-package “seasonal” which automatically
initial degradation effects such as the Staebler–Wronski effect
performs seasonal adjustments, automatically calculates the op-
or short-term LID. The monitored data are prefiltered to ex-
timal ARIMA (p, d, q)(P, D, Q) variables to apply on the data
clude data with performance ratio values below 1% and above
set [24]. For the application of ARIMA within Section IV-C, the
200% and a POA irradiation of less than 50 W/m2 and more
extracted parameters for both systems are (0, 1, 1) (0, 1, 1).
than 1500 W/m2 . This was done to remove extreme outliers
5) Seasonal–Trend Decomposition Using LOESS: STL is a
and measurement errors. The simplicity of the filter was chosen
continuation of CSD and Loess is a method to estimate nonlinear
to intensify possible deviations among the models. The irradi-
relationships. The centered moving average is replaced by a
ance data are recorded with a pyranometer. For each model,
locally weighted regression to extract the trend [25]. Because
the relative annual PLRs, the corresponding uncertainty and the
of that, the estimates become more robust and are less affected
intercept with the y-axis are given. The PLR of the more sophis-
by missing data and outliers. Similar to CSD, STL decomposes
ticated models are calculated by applying a linear regression to
a seasonal time-series into three components (trend, seasonal,
the respective trend, which was extracted through the statistical
remainder) and is described by
model. To receive the yearly relative PLR and the corresponding
Yt = Tt + St + Rt . (15) uncertainty, the following formulas are used [27]:
STL contains an inner and an outer loop. Every time a run within 12a
PLR = (16)
the inner loop is performed, the seasonal and trend components b
are updated. The number of runs within the inner loop are mostly  
  2  2
equal to 1 or 2. The outer loop includes an inner loop followed  12 −12a
uPLR =  × u2a + × ub
2 (17)
by a calculation of robustness weights. This calculation serves b b2
as an input for the following inner loop to decrease the im-
pact of transient, abnormal behavior on the trend and seasonal where a and b are the fitting coefficients of the linear regres-
parts [26]. sion, u2a,b the variances of these fitting coefficients, and uPLR
To better grasp the idea of Loess, the method is explained the standard deviation of the PLR. This uncertainty calculation
when applied within the software R. Here, two parameters have corresponds to a confidence interval of 68%.
to be chosen, the trend window and the seasonal window. The Two definitions of the PLR can be found in the literature,
seasonal window is either periodic or the span of the Loess win- in relative terms as PLR = 12a/b or absolute with PLR = 12a
dow for seasonal extraction. The smaller the values, the faster [18]. Here, the relative PLR was chosen because it makes it eas-
the trend and seasonal components can change. A high value ier to generalize the findings to the energy yield of the array using
for the seasonal window forces the seasonal part to be peri- the initial yield of the plant. The results of these calculations are
odic, in this case just the means for the monthly values are used less aimed to deliver the best possible combination of filtering
(seasonal component for January is mean of all January val- techniques, performance metrics, and statistical models but are
ues). After calculating the seasonally adjusted data, (measured intended to provide a direct comparison between the presented
data minus seasonality) the trend is Loess-smoothened. This analysis methods. While the uncertainties of the resulting PLRs
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE JOURNAL OF PHOTOVOLTAICS

Fig. 1. Comparison of statistical models on PR-data, circles represent PLR including uncertainties (primary axis), triangles represent initial PR (secondary axis),
on the left is mc-Si system, on the right is a-Si system.

TABLE I
COMPARISON OF STATISTICAL MODELS

were used to rate the statistical models, the remainder, where

applicable, serves as a validation of the parameter fit. The re-
mainder should have Gaussian white noise properties, such as
being uncorrelated and normally distributed.
In Fig. 1, the relative PLRs, uncertainties and initial PR values
are given. The initial value is the intercept with the y-axis.
The stated uncertainty is the uncertainty of the PLR against the
extracted trends, computed by the individual statistical methods.
It can be seen that the performance of the mc-Si system
degrades 0.5%–0.6% per year, the one of a-Si close to 1.8% per
year. Because of cabling and other system losses the initial PR
value is below the theoretical value under STC. As expected, Fig. 2. Trend-lines of CSD- (red-straight) and STL-model (blue-dotted) of
linear regression shows in both cases the highest uncertainty. unfiltered monthly PR-data of a mc-Si system.
For all other cases, a trend was first filtered from the data set
on which a linear regression was performed. This step leads to
an outcome with higher certainties. While all more advanced into account and because of this, the decrease is stronger over
methods show similar results with regards to the uncertainty of time. This in turn results additionally in an overestimation of
the mc-Si system, STL and ARIMA outperform the others when the initial PR value.
applied to the a-Si system. Table I summarizes the models in respect to different char-
In case of CSD, the loss ratio is for both systems higher acteristics of time-series. In general, SLR might serve as a first
in comparison and probably overestimated. In Section IV-B2, it indication of a PLR determination but is not suited for accurate
was mentioned that, when using CSD, the first and last months of calculations due to its simplistic approach. It has been shown
the data set are lost due to the applied centered moving average. that the more sophisticated selected analysis methods perform
This is visible in Fig. 2. Here, the extracted trend-lines of the PR very similar for crystalline systems, both in the estimation of
of the mc-Si system using CSD and STL are shown. Within the the initial PR and the PLR. An exception hereby is the usage
first six months of observation, the trend of the PR has a roughly of CSD for short time-series. The exclusion of the first and last
stable value. When applying CSD, this time period is not taken observations can falsify the final PLR. When a thin-film system
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LINDIG et al.: REVIEW OF STATISTICAL AND ANALYTICAL DEGRADATION MODELS FOR PHOTOVOLTAIC MODULES AND SYSTEMS 7

is subject of the calculations, STL and ARIMA show the best (t = 0) before aging and at time (t = t) after aging or in the
results. field.
When applying HW, ARIMA or STL, it is very important to
perform crucial modeling steps with great care to receive the A. Degradation Models for Corrosion
optimal results. In case of ARIMA, the time-series has to be
Corrosion is one of the most occurring degradation modes
transformed in order to reach stationarity. Since PV time-series
in PV modules [30]. Corrosion is caused by the presence of
are highly seasonal, a seasonal differentiation is essential. For
high temperature and high humidity in the module. Humidity
the same reason, the seasonal window parameter within STL
can enter the module through the backsheet or the layers of
has to be set as periodic and the additive method is preferred
the encapsulant and spread into the module [31], weakening
when using the HW model.
the adhesive bonds between the interfaces. One hypothesis is
Filtering is an integral part when computing PLR values.
that humidity, which catalyzes corrosive processes, leads to a
If performance metrics are corrected using temperature coeffi-
formation of acetic acid through the hydrolysis of vinyl-acetate
cients, which are retrieved at a POA irradiance of 1000 W/m2 ,
monomers present in the EVA [32]–[34]. Corrosion attacks the
an appropriate POA irradiance should be selected. Within this
metallic connections of PV cells and results in a loss of adhesive
work, the preliminary filter are treating outliers and values cor-
strength between the cells and the metallic frame, as well as an
responding to measurement errors not sufficiently. This was
increased leakage current and therefore a loss in performance
done on purpose to amplify the impact of outliers and measure-
[32]. Empirical models based on power at maximum power
ment inaccuracies on the final results. SLR and CSD treat all
point and series resistance as degradation indicators have been
values with similar weights and are therefore strongly affected
proposed to model corrosion according to [35]–[37]. The models
by outliers. HW’s weighted average, STL’s locally weighted
are as follows.
regression, and the combination of similar techniques within
1) The Model of Pan [35]
ARIMA are well suited for outlier handling.
Another statistical method worth mentioning is the year-on- Pmax
= exp(−RD tβ ). (18)
year model developed by Sunpower [28] and later improved by Pmax(0)
NREL [29]. It is implemented within the Python RdTools for
2) Pmax and Rs Models According to Braisaz [36]
the analysis of PV data. This method has a complete different
approach as the here discussed models, as it is using a loss rate 1 − exp(−B)
Pmax = (19)
distribution instead of one single value. The gradient between 1 + exp(RD t − B)
two related data points in consecutive years (hour, day, week,
Rs = Rs(0) + exp(RD t − B) (20)
month) determine a single PLR. The median of this gradient,
the gradients of all remaining data points of that two years, where Pmax and Rs are the output power and series resistance
and all following years determine the final performance loss at time (t), Pmax(0) and Rs(0) are the power output and series
per year. The power loss rates are computed using a 100% per- resistance at time (t = 0), β is the experimental parameter, B
formance baseline value. This approach is excluded within this is a coefficient to be defined, and RD is the degradation rate
study. A comparison is difficult to perform because the initial determined according to (21), (22), and (23).
value is preset and, in contrast with the performed computations, 1) Models for Degradation Rate (RD ) Calculation: Kinetic
the data aggregation is done in an irradiance-weighted manner. models are developed on the primary assumption that the rate
Nevertheless, this approach is of special interest if high quality of degradation is proportional to the concentration of water in
field irradiance data are not available because it can compute PV modules and that the rate constant has a Arrhenius temper-
the PR based on a modeled clear-sky irradiance. ature dependence. Three models according to [19] and [38] are
proposed, namely the Peck’s model, the Eyring model, and the
exponential model.
V. MODELING SPECIFIC DEGRADATION MODES
Peck’s model:
USING ANALYTICAL MODELS  
Ea
In order to explain experimental observations of different RD .Peck = A exp − RH n . (21)
kB T
degradation modes, analytical models are developed based on
the physical/chemical theories of the degradation mode in ques- Eyring model:
tion. These models are environmental stress oriented. The hy-  
−Ea b
potheses of a particular degradation mode are built depending on RD .Eyring = A exp − . (22)
kB T RH
specific environmental stresses applied, and on the assumption
that the kinetics of a specific degradation mode are influenced by Exponential model:
one dominating process. Electrical parameters such as power at  
−Ea
maximum power point (Pmax ), short circuit current (ISC ), shunt RD .Exp = A exp . exp(m × RH). (23)
kB T
(Rsh ), and series resistance (Rs ) are commonly modeled as
degradation indicators. Hence, the environmental stresses and Here, Ea is the activation energy of the degradation process
their interactions with the PV module components are assessed [eV], T the module temperature [K], kB is the Boltzmann con-
based on the reduction of the initial electrical parameter at time stant (8.62 × 10−5 eV/K), and RH is the relative humidity [%].
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE JOURNAL OF PHOTOVOLTAICS

TABLE II
COMPARISON OF MODELS FOR DEGRADATION RATE

Fig. 3. Module 1. Maximum power degradation models of Pan (P) and Braisaz
A, n, b, and m are model parameters. RD [%/h] is the inverse of (B), simulated for three climates: Zugspitze (green), Gran Canaria (blue), and
the mean time to failure at a given condition. In order to obtain Negev (purple).
A, Ea , n, b, and m in (21), (22), and (23), the equations can be
fitted to experimental data or represented on a logarithmic scale
by a straight line, using the following equations: mean values of temperature and relative humidity were used.
  In both cases, the Peck’s model was used for degradation rate
Ea
ln(RD .Peck ) = ln(A) − + n × ln(RH) (24) calculation and both models were fitted to indoor data to extract
kB T the model parameters.
 
Ea b Both models show that temperature is the most relevant fac-
ln(RD .Eyring ) = ln(A) − − (25) tor that influences the PV degradation process, visible by the
kB T RH
  power losses depicted in Fig. 3. This can be seen by a small loss
Ea through lower degradation in power for the module installed at
ln(RD .Exp ) = ln(A) − + m × RH. (26)
kB T Zugspitze, where the climate is characterized by low tempera-
tures and high levels of relative humidity. However, the models
A plot of ln(RD ) versus 1/T [K] gives an Arrhenius plot with
completely differ in degradation predictions. According to the
a slope Ea /kB and an intercept ln(A).
simulations, the model of Pan converges at a relatively fast rate
2) Comparison of Corrosion Models: The models of Peck,
compared with the one of Braisaz. This could be due to the in-
Eyring, and exponential were applied to fit indoor data sets
fluence of the model coefficients such as B and β. Moreover, the
(damp heat at 85C/85% RH), of two c-Si modules, with module
time parameter for the Pan model follows a power law, hence β
1 showing a good performance stability compared with module
might be an accelerating factor.
2. The models were compared based on the extracted parame-
ters (see Table II) as well as the deviation from the fitted data
B. Degradation Models for PID
points. All models are consistent concerning the influence of
the activation energy on the degradation rate and that the pre- PID has been observed in all PV technologies and in almost
dicted values are in the range of literature values for polymeric all operating climates. It does not occur so frequently, but if
materials, which usually range between 0.6 and 2.0 eV [39]. it does, its effect can lead to a severe performance loss within
Nevertheless, the Eyring model shows a significant difference a short period [4], [41]–[44]. In general terms, PID is caused
in the extracted activation energy in comparison to that of Pecks by the difference in potential between the cells and the sup-
and the exponential model. It also has the highest percentage port structure of the module. This difference drives a leakage
deviation of fitted data points in both cases. Therefore, from this current that can lead to power degradation. Different types of
study, we can conclude that the Peck’s model has a better fit PID occur depending on the module technology. For crystalline
compared with the other models. silicon PV, two degradation modes can be identified, PID-p (for
Degradation models are utilized for the simulation of power polarization or passivation) and PID-s (for shunting). PID-p is
output degradation for module 1 to predict its performance a temporary and reversible degradation of the passivation layer,
in three climatic zones: alpine (Zugspitze, Germany), mar- which reduces the performance due to a surface recombination
itime (Gran Canaria, Spain), and arid (Negev, Israel), assuming increase [45]–[47]. PID-s is because of a leakage current in-
the degradation is because of corrosion according to Pan and volving an ionic flow of Na+ from the glass, encapsulant or cell
Braisaz. The module temperature is given for a standard c-Si surface into the cell, diffusing into the silicon stacking faults and
PV module type glass-backsheet construction installed at the shunting the cell [48]. The sodium incorporation in the Si sur-
three test sites and relative humidity was calculated from ambi- face degrades primarily the FF, the Voc , and finally the Isc . The
ent conditions according to [40]. For all the simulations, annual relevant stress factors for PID-s include [42]; high temperature,
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LINDIG et al.: REVIEW OF STATISTICAL AND ANALYTICAL DEGRADATION MODELS FOR PHOTOVOLTAIC MODULES AND SYSTEMS 9

TABLE III
PIDHACKE MODEL

relative humidity, system voltage, light, bias-junction potential,

and injected carriers. The models proposed for PID degradation
according to [36], [43], [44], [49]–[51] are as follows.
1) Pidhacke Model: A parabolic model was proposed by
Hacke et al. [44] to fit the power degradation of c-Si modules
 
Pmax Ea
= 1 − A exp − RH n × t2 . (27)
Pmax(0) kB T
Fig. 4. Simulated normalized maximum power degradation because of PID
The constants A and n are determined by fitting the equation in three climatic zones; Freiburg (green), Gran Canaria (blue) and Negev (red).
to experimental results. The parameters have to be determined
for each module type. This parabolic model is applicable to the
beginning of the degradation phases of PID-s, as it can fit the parameters for the module power. The model is written as
beginning of a sigmoid and does not describe the stabilization
Pmax (U, T, RH, t) = Pmax(0) (1 − P (t)) (30)
phase of the sigmoidal curve. Annigoni et al. [49] used the
indoor data to determine coefficients of the model for distinct 1 − exp t
τ1
aging contributions (temperature, relative humidity, and time) P (t) = P∞ ;
and then applied the model including a voltage term in (28) to 1 − exp t−t 0
τ2
outdoor PID degradation for different climates
  −2
U − U0
  P (U ) = 1 + exp (31)
Pmax Ea Φ
= 1 − A exp − RH n × t2 × U. (28)
Pmax(0) kB T
t0 = a × b × tˆ0 ; τ1 (T ) = b2 × τˆ1 ; τ2 = τˆ2
(32)
In this study, a similar approach is adopted to extract the model  
coefficients in Table III and to simulate maximum power degra- H0 T − T0
a(H) = ; b(T ) = exp . (33)
dation because of PID in three climatic zones; alpine (Zugspitze, H φ
Germany), maritime (Gran Canaria, Spain), and arid (Negev,
Israel). In all simulations, a constant voltage of 500 V was as- The model includes six adaptation parameters: tˆ0 , U0 , τˆ1 ,
sumed. The simulation results, shown in Fig. 4, were consistent τˆ2 , Φ, and φ. H0 and T0 are scaling parameters. The function
with the ones of Annigoni, saying that a more severe degradation P (t) describes the power loss caused by degradation. P∞ (U )
is predicted for arid climates. is its limit for t → ∞, and a(H) and b(T ) are the acceleration
2) Exponential Model: The model was applied by Hacke functions of relative humidity and temperature. For T = 0 and
et al. [43] to predict PID occurrence in thin-film modules in H = 0, they are equal to 1, therefore τˆ1 , τˆ2 are the time constants
the field using accelerated tests. Considering shunting, which is under these conditions. τ2 remains constant for a given module.
the PID mode that occurs first, an exponential model based on To determine the model’s parameters, the power degradation
module temperature T and relative humidity RH was found to is measured as a function of time with the system voltage U
fit well the PID rate for multiple stress levels of a CdTe module as parameters and a fixed humidity H as well as temperature
in chamber tests. The power model is of the form T . The saturating power P∞ is extracted by fitting P (t) to the
measured data.
    4) Taubitz Model: Taubitz et al. [51] proposed a regener-
d Pmax Ea
= 1 − Af (U ) exp − RH n × t (29) ation model for shunt resistance evolution over time because
dt Pmax(0) kB T of PID degradation. The shunt resistance was modeled in three
phases: shunting phase, regeneration phase, and transition phase
where f (U ) expresses the voltage dependency. as follows:
3) Model of Hattendorf: The model of Hattendorf et al. [50] Shunting phase:
is based on a matrix of indoor experiments where modules are  
exposed to varying voltage, module temperature, and ambient −t
Rsh (t) = aS exp . (34)
humidity. The conditions are varied to determine the model bS (t)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE JOURNAL OF PHOTOVOLTAICS

Regeneration phase: 1) Schwarzschild Law: The Schwarzschild Law has been

  applied by Gu et al. [55] to study the effect of intensity and
t wavelength of spectral UV light on discoloration of laminated
Rsh (t) = CR + aR exp . (35)
bR (t) glass/EVA/PPE PV modules. The law is a function of intensity as
Transition phase: k = A(I)p . (41)
2
Rsh (t) = aT (T )(t + bT (T )) + CT (36) Here, k is a constant, I the intensity, and p is the Schwarzschild
coefficient.
where aS , bS , CT , bT , aT (T ), bS (T ), bT (T ), CR , and aR are
Recommendation: When applying this expression in perfor-
constants and have to be determined for a specific module type.
mance (power) prediction models where other loads are also
Some of them are dependent on the module temperature T . The
applied, the parameter p must be calibrated according to the
constants are determined by measuring the times tS , tT , and tR
knowledge of severity ranking [56].
for reaching certain target values.
5) PID Model According to Braisaz et al. [36]: The model
D. Degradation Models for Delamination, Fatigue Solder
is based on shunt resistance Rsh degradation as an indicator, as
Failure, and Cell Cracks
it is the most important parameter for PID. The evolution of Rsh
as a function of voltage, temperature, and relative humidity was 1) Coffin–Manson’s Equation: The model is used to predict
modeled as degradation modes caused by temperature cycling such as en-
Rsh(0) capsulant delamination, fatigue solder failure, and cell cracks.
Rsh (t) = (37) According to Escobar and Meeker [38], the model describes the
1 + aRD t
  number of cycles to failure as
B Ea σ
RD = A × U exp − N= (42)
1 + exp(−C(RH) + D) kB T (ΔT )β 1
(38)
where ΔT is the temperature range and σ and β1 are properties
where Rsh is the shunt resistance at time (t), Rsh(0) is the initial of the material and test setup. The cycles-to-failure distribution
shunt resistance, A, B, C, and D are model coefficients, U is for temperature cycling can also depend on the cycling rate
the applied voltage, and RD is the degradation rate. (e.g., because of heat buildup). An empirical extension of the
Recommendation: The model contains many coefficients Coffin–Manson relationship that describes such dependencies
whose usage is not clearly described, such coefficients might is [38]
affect the physical interpretation of the results when used in-  
σ 1 Ea × 11605
appropriately. Therefore, we recommend that one should have N= exp (43)
a prior knowledge of the impact of the parameters on which (ΔT )β 1 (freq)β 2 Tmax (K)
the coefficients are being applied to the degradation process in where freq is the cycling frequency and Ea is a quasi-activation
question. energy.
2) Crack Propagation Model: The model was suggested by
C. Models for UV Degradation Braisaz et al. [36] and it was applied to simulate the degradation
of the short-circuit current Isc due to the expansion of cell cracks
UV light exposure has been reported to cause PV module
caused by temperature. The model takes the form
degradation in a number of ways. Exemplary, it could result in
discoloration of the encapsulant material [52] or delamination 1
Ca (t) = Ca (t − 1) + m . (44)
at the glass encapsulant or cell encapsulant interface [53]. The x 125
Ta
parameter most impacted by UV exposure is the short circuit
current (Isc ). Braisaz et al. [36] proposed a model for short cir- Ca (t) is the crack activation at time (t), Ca (t − 1) the crack
cuit degradation because of UV exposure over time. They found activation at time (t − 1), Ta is the daily temperature amplitude,
that the degradation curve is not linear but an exponentially de- m a model parameter, and x is the number of thermal cycles.
creasing curve. The short circuit is modeled as a function of The crack activation/propagation model is dependent on the
UV as daily temperature amplitude Ta .
3) Damage Accumulation Model: The model was used by
Isc (t) = Isc(0) − aDUV (t) − b(1 − exp(−CDUV )(t)) (39)
Bosco et al. [57] in order to calculate the solder fatigue damage
t in seven cities investigated in their study and compared with
DUV (t) = E(u) × 5.5% du. (40) FEM simulated results. They found out that the model fits well
0
to the simulated calculations. The model is written as
Here, DUV is the UV dose in MJ/m2 or kWh/m2 , Isc is the  
Q
short circuit current at time (t), Isc (0) is the initial short circuit D = C(ΔT ) (r(T )) exp −
n m
. (45)
kB Tmax
current, and a, b, and C are model coefficients. The multipli-
cation by 5.5% is because the UV radiation (280–400 nm) is In this equation, ΔT is the mean daily maximum cell temper-
approximately 5.5% of the total light spectrum E(u) [54]. ature change, Tmax the mean maximum daily temperature, C a
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LINDIG et al.: REVIEW OF STATISTICAL AND ANALYTICAL DEGRADATION MODELS FOR PHOTOVOLTAIC MODULES AND SYSTEMS 11

scaling constant, Q the activation energy, kB Boltzmann’s con- cal/temperature excursion factor fM is represented by
stant, r(T ) the number of times the temperature history increases ⎡ ⎤
or decreases across the reversal temperature, T the period of a exp G1 T1 − T1b + exp −G2 T1 − T1b
fM = ⎣ ⎦
year, and n and m are model constants similar to those in the D0
Coffin–Manson equation.
4) Backsheet Degradation Model: The model is used to es- × exp (JΔT ). (51)
timate a potential form of the degradation kinetics of the back-
sheet. This kinetic model was applied by Kempe [58] to model The first term in the bracket reflects the stresses arising from dif-
the uncertainty in a 25-year equivalent test for module backside ferences in expansion coefficients of bonded materials. The con-
exposure to irradiance and temperatures in different climatic stants G1 , G2 , D0 , and Tb are chosen to represent the estimated
zones. The degradation rate model is written as magnitudes of these fatigue effects. The factor exp(JΔT ) esti-
mates the effect of the magnitude of the temperature excursion
T −T 0
RD ≈ I X (b + m × TOW) × (Tf ) 10 . (46) ΔT , where J is a constant
 G E − FT
Here, I is the light intensity, X, b, and m are fit parameters, fG = 1 + . (52)
TOW is the time of wetness, T the temperature, T0 a reference Go
temperature, and Tf is a multiplier for the increase in degrada- Here, E and F denote constant parameters and T is the temper-
tion for a rise in temperature in 10 K steps. ature. The frequency of the temperature excursion fω is repre-
Recommendation: As also mentioned by Kempe, the param- sented by
eter that describes the effect of time of wetness has very high  ω P − T
Q

uncertainties, we recommend careful comparison of the relative fω = 1 + . (53)

ωo
change in degradation rate with changes in TOW. In case one
wants to extract thermal parameters such as activation energy, ω is the frequency and P as well as Q are constant parameters.
the multiplier term (Tf ) can be replaced by the Arrhenius term. In a constant temperature test, T is a constant and ω is taken to
be zero. In the cyclic temperature tests, reciprocal temperature
E. Model Based on Multiple Stresses is considered to be a sinusoidal function of time
1
Since degradation of PV modules in outdoor operation is in- (t) = τ + Δτ sin(ωt) (54)
T
fluenced by multiple environmental stresses, models based on
multiple stresses are viable for outdoor service lifetime predic- 1 1 1 
τ= − . (55)
tion. 2 Tmin Tmax
1) Model of Gaines: Gaines et al. [37] proposed a model for Tmin and Tmax are the minimum and maximum temperatures
power output degradation based on multiple accelerated envi- associated with the temperature cycles.
ronmental stresses. The model suggested is Recommendation: The model of Gaines presents the previous
Pmax 1
approach on multiple stress modeling, however, the user should
= [1 − RD t] β (47) take caution that this model was developed and applied on PV
Pmax(0)
modules that had a different construction from today’s modules.
RD = AfT fR H fM fG fω (48) Therefore, its application might need some modification to fit
the current PV module construction types.
where RD is the degradation rate and the factors fT , fR H , fM , 2) Model of Subramaniyan: Another model to calculate the
fG , and fω are associated with a decrease in power output due degradation rate because of combined environmental stresses
to effects of temperature T , relative humidity RH, mechanical has been proposed recently by Subramaniyan et al. [59]. The
stresses M (due to temperature differences), gaseous concen- model takes into account the effect of both static and cyclic
tration G, and the frequency of the temperature excursion. The temperature, UV radiation and relative humidity as
mathematical form of each factor is formulated to represent the  
underlying physical phenomena. An Arrhenius form is used for β1
Rate(T, ΔT, UV, RH) = β0 exp −
  kB Tmax
B
fT = exp − . (49) × (ΔTdaily )β 2 × (U Vdaily )β 3 × (RHdaily )β 4 (56)
T
where Rate(T, ΔT, U V, RH) is the reaction rate, Tmax the daily
B denotes a constant parameter and T denotes temperature. The
maximum temperature of the module [K], ΔTdaily the daily
effect of relative humidity fRH is represented by
cyclic temperature of the module [K], UVdaily the daily daytime
  C − DT average irradiance [W/m2 ], RHdaily the daily average relative
1 1
fRH = 1 + (RH)0 exp C0 − . (50) humidity [%], and k is the Boltzmann constant. The model
TC T0
parameters β0 , which is the frequency factor [s−1 ], β1 , the acti-
The second term in the bracket corrects the relative humid- vation energy [eV], β2 , the effect of cyclic temperature, β3 , the
ity as a function of temperature, given a specified relative hu- effect of UV radiation and β4 , the effect of RH, can be estimated
midity at T0 . C and D are constant parameters. The mechani- from measured data through data fitting techniques. In the scope
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE JOURNAL OF PHOTOVOLTAICS

of this paper, the model is presented but has not been applied or more advanced models. In case of ARIMA, the time-series has
tested in any way due to a lack of combined stress data. A more to be stationary. The model parameters need to be chosen based
detailed description and application of the model can be found on the time-series behavior. That means that PV-related param-
in [59]. eters, like the varying temperature dependency of different PV
technologies, or the prevailing weather conditions have to be
taken into account.
VI. DISCUSSION AND CONCLUSION Unfortunately, it is not possible to determine if a (or which)
In this paper, several PV system performance loss method- degradation mode occurs on the basis of calculated PLRs. Apart
ologies are reviewed. Hereby, statistical and analytical models from degradation modes, other factors such as shading or soiling
are taken into account. might be a reason for a reduced performance. Because of that,
First, a discussion about statistical models to determine the it is important to not only study the data of a PV system, but
PLR of PV systems from available outdoor data is presented. also to undertake regularly visual and electrical inspections and
The performance loss trend is retrieved by applying filters, connect the findings with the calculated PLR. An idea of how
performance metrics, and statistical models on data sets. By to isolate the occurrence of degradation modes within a module
performing detailed in-depth performance studies, it might be under surveillance might be the application of the presented
possible to gain a greater understanding about the root causes models on the short-circuit current, the open-circuit voltage and
of the decrease in the power output of PV systems over time. the fill factor. Hence, it might be possible, together with the
Especially when considering current and voltage behavior, spe- inclusion of results from the studies of accelerated tests, to find
cific degradation modes could be identified and at a later stage patterns in the trends of these values.
verified by visual inspection techniques. In the second part of this work, several analytical models
As a first measure, appropriate filtering techniques have to for specific degradation modes, which trigger the aforemen-
be applied on the data set in question. The choice of the filter tioned performance losses to a large extent, are further studied.
will strongly depend on the performance metric and/or statis- The models proposed for corrosion and PID were implemented
tical model and, in case of an inappropriate filter window, will to simulate the maximum power degradation in three climatic
falsify the final outcome. On the other hand, pretreatment of zones. A strong influence of the climate is evident in all the sim-
the data set is necessary to eliminate outliers, noise, and mini- ulation results where, as expected, a more severe degradation is
mize seasonal oscillation. Before deciding which performance predicted for arid climates. The key observations in this study
metric or statistical model to use, the PV system technology, the are as follows.
length of the observed period, climatic conditions, and mounting 1) Although these models give a preliminary approximation
system (rack, tracker) should be taken into consideration. Pre- of the time-evolution of power performance, they do not
vailing seasonality, temperature/irradiance dependency of the provide any information on the physical processes taking
I–V curve parameters and noncorrelated outliers (data errors, place within the module.
shading effects, etc.) will increase the uncertainty of the re- 2) The models are developed based on numerous assump-
sults and influence the final PLR. The final aim is to receive tions and simplifications, moreover the hypotheses of a
a clear performance trend. In Section IV-C, the discussed sta- particular degradation mode are modeled depending on
tistical models have been applied on monitored field data of environmental stress factors and do not take into consid-
one monocrystalline silicon and one amorphous silicon system eration the influence of material parameters.
to retrieve long-term performance trends. Thereby, the applica- 3) None of the models is universal, that is, they can well
tion of SLR resulted in performance ratings with the greatest describe the degradation of a specific type of a PV tech-
uncertainties in comparison. The usage of CSD produced per- nology and fail on the other. Therefore, it is necessary to
formance rates with low uncertainties but due to the elimination be certain that a chosen degradation model is valid for a
of the first and last months of monitored data through the cen- specific application.
tered moving average, this technique is not recommended for 4) The analytical models are developed and validated based
data sets, which just consider short time periods. It was seen on indoor data from accelerated tests. Although some au-
that CSD overestimated the performance loss of both systems. thors went on to simulate outdoor conditions based on
The remaining techniques, namely HW exponential smoothing, indoor observations, a big challenge remains of how to
ARIMA, and STL, are performed on a similar high level of ac- interpret the results for multiple environmental stresses
curacy and the results are almost identical. HW experiences a using indoor data.
slightly higher uncertainty when applied to the amorphous sili- According to these observations, we recommend further
con PV system. It seems that ARIMA and STL are better suited developments for models that take into account both material
for noncrystalline PV systems due to their favorable treatment and multiple environmental stress factors. The development of
of the temperature behavior of the system in question. These such models need to be related to indoor as well as outdoor
three techniques exclude the seasonal part in time-series of PV observations.
performance metrics, which is an important modeling step in or-
der to receive a clear performance trend. Nevertheless, it has to
be stressed that the application of statistical models, especially ACKNOWLEDGMENT
ARIMA, has to be performed with great care and that it is not a This project has received funding from the European Union’s
rudimentary exercise to retrieve accurate model parameters for Horizon 2020 research and innovation programme in the frame-
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LINDIG et al.: REVIEW OF STATISTICAL AND ANALYTICAL DEGRADATION MODELS FOR PHOTOVOLTAIC MODULES AND SYSTEMS 13

work of the project “SolarTrain” under the Marie Sklodowska- [24] C. Sax, Seasonal: R Interface to X-13-ARIMA-SEATS, R package version
Curie GA No 721452. 1.6.1. 2017. [Online]. Available: https://CRAN.R-project.org/package=
seasonal
[25] A. Phinikarides, G. Makrides, and G. E. Georghiou, “Comparison of
REFERENCES analysis method for the calculation of degradation rates of different pho-
tovoltaic technologies,” in Proc. 27th EU PVSEC Proc., 2013, pp. 3211–
[1] C. S. Solanki, Solar Photovoltaics - Fundamentals, Technologies and 3215.
Applications vol. 2. New Delhi, India: PHI Learning Private Limited, [26] R. B. Cleveland, W. S. Cleveland, J. E. McRae, and I. Terpenning, “STL:
2012. A seasonal-trend decomposition procedure based on loess,” J. Official
[2] I. M. Peters, H. Liu, T. Reindl, and T. Buonassisi, “Global prediction Statist., vol. 6, no. 1, pp. 3–33, 1990.
of photovoltaic field performance differences using open-source satellite [27] Evaluation of Measurement Data—Guide to the Expression of Uncertainty
data,” Joule, vol. 2, no. 2, pp. 307–322, Feb. 2018. in Measurement,”International Bureau of Weights and Measures, BIPM-
[3] M. Kontges et al., Assessment of Photovoltaic Module Failures in the JCGM-100:2008, Sevres, FR, Guide, 2008.
Field. Report IEA PVPS T13-09:2017, Subtask 3, 2017. [28] E. Hasselbrink et al., “Validation of the PVLife model using 3 million
[4] D. Moser et al., “Identification of technical risks in the photovoltaic value module-years of live site data,” in Proc. IEEE 39th Photovolt. Spec. Conf.,
chain and quantification of the economic impact,” Prog. Photovolt.: Res. Jun. 2013, pp. 0007–0012.
Appl., vol. 25, no. 7, pp. 592–604, 2017. [29] D. C. Jordan, C. Deline, S. R. Kurtz, G. M. Kimball, and M. Anderson,
[5] International Electrotechnical Vocabulary. Chapter 191: Dependability “Robust PV degradation methodology and application,” IEEE J. Photo-
and Quality of Service, IEC60050-191, International Electrotechnical volt., vol. 8, no. 2, pp. 525–531, Mar. 2018.
Comission, Geneva, CH, Standard, 1990. [30] A. Ndiaye et al., “Degradations of silicon photovoltaic modules: A liter-
[6] D. C. Jordan and S. R. Kurtz, “Photovoltaic degradation rates - An an- ature review,” Solar Energy, vol. 96, pp. 140–151, Oct. 2013.
alytical Review: Photovoltaic degradation rates,” Prog. Photovolt.: Res. [31] N. Park, C. Han, and D. Kim, “Effect of moisture condensation on long-
Appl., vol. 21, no. 1, pp. 12–29, Jan. 2013. term reliability of crystalline silicon photovoltaic modules,” Microelec-
[7] A. Phinikarides, N. Kindyni, G. Makrides, and G. E. Georghiou, “Review tron. Rel., vol. 53, no. 12, pp. 1922–1926, Dec. 2013.
of photovoltaic degradation rate methodologies,” Renew. Sustain. Energy [32] M. D. Kempe et al., “Acetic acid production and glass transition concerns
Reviews, vol. 40, pp. 143–152, Dec. 2014. with ethylene-vinyl acetate used in photovoltaic devices,” Solar Energy
[8] D. Moser, M. Pichler, and M. Nikolaeva, “Filtering procedures for reliable Mater. Solar Cells, vol. 91, no. 4, pp. 315–329, Feb. 2007.
outdoor temperature coefficients in different photovoltaic technologies | [33] K. Whitfield, A. Salomon, S. Yang, and I. Suez, “Damp heat versus field
journal of solar energy engineering | ASME DC,” J. Solar Energy Eng., reliability for crystalline silicon,” in Proc. 38th IEEE Photovolt. Spec.
vol. 136, no. 2, pp. 021006-1–021006-10, 2013. Conf., Jun. 2012, pp. 1864–1870.
[9] C. M. Whitaker et al., “Application and validation of a new PV per- [34] A. Masuda, N. Uchiyama, and Y. Hara, “Degradation by acetic acid for
formance characterization method,” in Proc. Conf. Record 26th IEEE crystalline Si photovoltaic modules,” Jpn. J. Appl. Phys., vol. 54, no. 4S,
Photovolt. Spec. Conf., Sep. 1997, pp. 1253–1256. pp. 04DR04-1–04DR04-6, Mar. 2015.
[10] G. H. Yordanov, “Relative efficiency revealed: Equations for k1-k6 of the [35] R. Pan, J. Kuitche, and G. Tamizhmani, “Degradation analysis of solar
PVGIS model,” in Proc. IEEE 40th Photovolt. Specialist Conf., Jun. 2014, photovoltaic modules: Influence of environmental factor,” in Proc. Annu.
pp. 1393–1398. Rel. Maintainability Symp., Jan. 2011, pp. 1–5.
[11] D. L. King, J. A. Kratochvil, and W. E. Boyson, “Photovoltaic array [36] B. Braisaz, C. Duchayne, M. Van Iseghem, and K. Radouane, “PV aging
performance model,” Tech. Rep. SAND2004-3535, 919131, Sandia Nat. model applied to several meteorological conditions,” in Proc. 29th PVSEC
Lab., Albuquerque, NM, USA, Aug. 2004. Proc., Amsterdam, The Netherlands, Sep. 2014, pp. 2303–2309.
[12] E. Skoplaki and J. A. Palyvos, “On the temperature dependence [37] G. B. Gaines et al., “Development of an accelerated test design for
of photovoltaic module electrical performance: A review of effi- predicting the service life of the solar array at mead, nebraska,” Quarterly
ciency/power correlations,” Solar Energy, vol. 83, no. 5, pp. 614–624, May Techn. Rep., NASA, USA, Feb. 1979.
2009. [38] L. A. Escobar and W. Q. Meeker, “A review of accelerated test models,”
[13] D. C. Jordan, J. H. Wohlgemuth, and S. R. Kurtz, “Technology and climate Statist. Sci., vol. 21, no. 4, pp. 552–577, Nov. 2006.
trends in PV module degradation,” in Proc. 27th EU PVSEC Proc., 2012, [39] R. Laronde, A. Charki, D. Bigaud, and P. Excoffier, “Reliability evaluation
pp. 3118–3124. of a photovoltaic module using accelerated degradation model,” in Proc.
[14] T. Huld et al., “A power-rating model for crystalline silicon PV mod- SPIE, 2011, vol. 8112.
ules,” Solar Energy Mater. Solar Cells, vol. 95, no. 12, pp. 3359–3369, [40] M. Koehl, M. Heck, and S. Wiesmeier, “Modelling of conditions for
Dec. 2011. accelerated lifetime testing of Humidity impact on PV-modules based on
[15] C. Jennings, “PV module performance at PG&E,” in Proc. Conf. Rec. 20th monitoring of climatic data,” Solar Energy Mater. Solar Cells, vol. 99,
IEEE Photovolt. Spec. Conf., Sep. 1988, pp. 1225–1229. pp. 282–291, Apr. 2012.
[16] C. Whitaker, T. Townsend, and H. Wenger, “Effects of irradiance and [41] J. Berghold et al., “Potential-induced degradation (PID) and its correlation
other factors on PV temperature coefficients,” in Proc. Conf. Rec. 22nd with experience in the field,” Photovolt. Int., vol. 19, pp. 85–92, 2013.
IEEE Photovolt. Spec. Conf., Oct. 1991, pp. 608–613. [42] M. Kontges et al., “Review of failures of photovoltaic modules,” Report
[17] Photovoltaic System Performance Monitoring—Guidelines for Measure- IEA PVPS T13-01:2014, Subtask 3.2, 2014.
ment, Data Exchange and Analysis, IEC61724:1998, International Elec- [43] P. Hacke et al., “Elucidating PID degradation mechanisms and in-situ dark
trotechnical Comission, Geneva, CH, Standard, 1998. I-V monitoring for modeling degradation rate in CdTe thin-film modules,”
[18] G. Belluardo et al., “Novel method for the improvement in the evaluation IEEE J. Photovolt., vol. 6, no. 6, pp. 1635–1640, Nov. 2016.
of outdoor performance loss rate in different PV technologies and com- [44] P. Hacke et al., “Accelerated testing and modeling of potential-induced
parison with two other methods,” Solar Energy, vol. 117, pp. 139–152, degradation as a function of temperature and relative humidity,” IEEE J.
Jul. 2015. Photovolt., vol. 5, no. 6, pp. 1549–1553, Nov. 2015.
[19] D. C. Jordan and S. R. Kurtz, “Analytical improvements in PV degradation [45] V. Naumann et al., “Potential-induced degradation at interdigitated back
rate determination,” in Proc. 35th IEEE Photovolt. Spec. Conf., Jun. 2010, contact solar cells,” Energy Procedia, vol. 55, pp. 498–503, Jan. 2014.
pp. 688–693. [46] A. Halm et al., “Potential-induced degradation for encapsulated n-type
[20] C. C. Holt, “Forecasting seasonals and trends by exponentially weighted IBC solar cells with front floating emitter,” Energy Procedia, vol. 77,
moving averages,” Int. J. Forecasting, vol. 20, no. 1, pp. 5–10, Jan. pp. 356–363, Aug. 2015.
2004. [47] R. Swanson et al., “The surface polarization effect in high-efficient silicon
[21] R. J. Hyndman and G. Athanasopoulos, Forecasting: Principles and Prac- solar cells,” in Proc. 15th Int. PVSEC Proc., Jan. 2005.
tice. Melbourne, Australia: OTexts, 2013. [48] D. Lausch et al., “Sodium outdiffusion from stacking faults as root cause
[22] A. Phinikarides, G. Makrides, N. Kindyni, A. Kyprianou, and G. E. for the recovery process of potential-induced degradation (PID),” Energy
Georghiou, “ARIMA modeling of the performance of different pho- Procedia, vol. 55, pp. 486–493, Jan. 2014.
tovoltaic technologies,” in Proc. 39th IEEE Photovolt. Spec. Conf., [49] E. Annigoni et al., “Modeling potential-induced degradation (PID) in
Jun. 2013, pp. 0797–0801. crystalline silicon solar cells: From accelerated-aging laboratory testing
[23] S. G. Makridakis and S. C. Wheelwright, Forecasting, 3rd ed. New York, to outdoor prediction,” in Proc. 32nd EU PVSEC Proc., Munich, Germany,
NY, USA: Wiley, Jan. 1998. Jun. 2016, pp. 1558–1563.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE JOURNAL OF PHOTOVOLTAICS

[50] J. Hattendorf et al., “Potential induced degradation in mono-crystalline David Moser received the B.Sc. and M.Sc. degrees
silicon based modules: An acceleration model,” in Proc. 27th Eur. PVSEC from the University of Trento, Italy, in 2003 and 2006,
Proc., Oct. 2012, pp. 3405–3410. respectively, the Ph.D. degree in physics from the Sal-
[51] C. Taubitz, M. Krber, M. Schtze, and M. B. Koentopp, “Potential induced ford University, Salford, U.K., in 2010.
degradation: Model calculations and correlation between laboratory tests He coordinates the activities of the research group
and outdoor occurrence,” in Proc. 29th Eur. PVSEC Proc., Nov. 2014, Photovoltaic Systems, Institute for Renewable En-
pp. 2490–2494. ergy, EURAC, Bolzano, Italy. His work focuses on
[52] L. Dunn, M. Gostein, and B. Stueve, “Literature review of the effects of characterizing indoor and outdoor behavior, perfor-
UV exposure on PV modules,” in NREL PV Module Reliab. Workshop, mance, and reliability of photovoltaic (PV) modules
Feb. 2013. and systems, building integration of PV systems, and
[53] M. A. Munoz, M. C. Alonso-Garcia, and Vela, “Early degradation of monitoring of outdoor PV plants. He is also active in
silicon PV modules and guaranty conditions,” Solar Energy, vol. 85, no. 9, PV potential studies on a regional scale and member of the Board of Directors
pp. 2264–2274, 2011. of The Association of European Renewable Energy Research Centres.
[54] M. Koehl, D. Philipp, N. Lenck, and M. Zundel, “Development and ap-
plication of a UV light source for PV-module testing,” in Proc. SPIE - Int
Soc. for Opt. Eng. 7412, Aug. 2009, pp. 7412021–7412027.
[55] X. Gu, Y. Lyu, L.-C. Yu, C.-C. Lin, and D. Stanley, “Effect of inten-
sity and wavelength of spectral UV light on discoloration of laminated
Glass/EVA/PPE PV module,” in 3rd Atlas-NIST Workshop PV Mater.
Durability, Dec. 2015.
[56] D. C. Jordan, T. J. Silverman, J. H. Wohlgemuth, S. R. Kurtz, and K. T.
VanSant, “Photovoltaic failure and degradation modes,” Progress Photo-
voltaics: Res. Appl., vol. 25, pp. 318–326, 2017.
[57] N. Bosco, T. J. Silverman, and S. Kurtz, “Climate specific thermomechan- Karl-Anders Weiß received the Diploma degree
ical fatigue of flat plate photovoltaic module solder joints,” Microelectron. in physics and economics and the Ph.D. degree in
Rel., vol. 62, pp. 124–129, Jul. 2016. physics from the University of Ulm, Germany, in
[58] M. D. Kempe, “Evaluation of the uncertainty in accelerated stress testing,” 2005 and 2014, respectively.
in Proc. IEEE 40th Photovolt. Spec. Conf., Jun. 2014, pp. 2170–2175. He is Head of the Service Life Analysis Group,
[59] A. B. Subramaniyan, R. Pan, J. Kuitche, and G. TamizhMani, “Quantifica- Fraunhofer Institute for Solar Energy Systems,
tion of environmental effects on PV module degradation: A physics-based Freiburg, Germany. His areas of interest include
data-driven modeling method,” IEEE J. Photovolt., vol. 8, no. 5, pp. 1289– degradation of materials in solar applications, accel-
1296, Sep. 2018. erated testing of components and materials of solar
systems, numerical simulations, methods to analyze
degradation of polymers, nondestructive analytical
Sascha Lindig received the B.Eng. degree in pho- methods climatic loads and service life prediction.
tovoltaic and semiconductor technologies from the
University of Applied Science Jena, Jena, Germany,
in 2011, and the M.Eng. degree in environmental and
energy engineering from the Leipzig University of
Applied Sciences, Leipzig, Germany, in 2014. He is
currently working toward the Ph.D. degree in elec-
trical engineering with the University of Ljubljana,
Ljubljana, Slovenia, in collaboration with EURAC
Research, Bolzano, Italy.
He is working in the frame of the Marie
Sklodowska Curie SOLAR-TRAIN project on statistical performance loss mod- Marko Topic received the Ph.D. degree from the
els of PV systems, degradation patterns and solar economics. University of Ljublana, Slovenia, in 1996.
He is the Head of Laboratory of Photovoltaics
and Optoelectronics, University of Ljubljana, Ljubl-
jana, Slovenia, acts as the Chairman of the European
Ismail Kaaya was born in Uganda. He received the
Technology and Innovation Platform Photovoltaics
B.S. degree in applied physics from the International
(ETIP-PV) since 2016 and previously as the Chair-
University of Africa, Khartoum, Sudan, in 2013, the
man of European Photovoltaic Technology Platform
Postgraduate Diploma degree in condensed matter (since 2014). He is currently a full Professor with the
physics from the Abdus-Salam International Center
Faculty of Electrical Engineering at the University
for Theoretical Physics, Trieste, Italy, in 2015, and the
of Ljubljana, and he has a very broad research ex-
M.S. degree in renewable energy science and technol-
perience in photovoltaics, thin-film semiconductor materials, electron devices,
ogy from Ecole Polytechnique, Palaiseau, France, in optoelectronics, electronic circuits, and reliability engineering.
2016. He is currently working toward the Ph.D. de-
Dr. Topic is a member of the Slovenian Academy of Engineering and has
gree in photovoltaics with Fraunhofer Institute for
received several prestigious awards, including the “Zoisova nagrada” in 2008
Solar Energy Systems, Freiburg, Germany.
(the highest award of the Republic of Slovenia for Scientific and Research
His research interests includes advanced characterization of hydrogenated
Achievements).
amorphous silicon, modeling spatial inhomogenities/(microscopic defects) in
solar cells, and development of service life models for PV modules.