Quantifying the influence of encapsulant and backsheet composition on PV‐power and electrical degradation

Although the technical and economic properties of the standard polymer photovoltaic (PV) materials (ethylene‐vinyl acetate (EVA) encapsulant and fluorine‐containing polyethylene terephthalate (PET) backsheet) meet the basic technical requirements, more sustainable polyolefin‐based encapsulants and backsheets have been developed. These new polyolefin materials have to prove their performance compared to the established standard materials in terms of the electrical performance of the modules and in terms of reliability. The long‐term stability of the new materials is tested and evaluated using accelerated aging tests and degradation modelling. Based on experimental results, the influence of the type of encapsulant and backsheet (i) on the electrical output power of PV test modules and (ii) on the aging‐related electrical and material degradation under accelerated stress tests was estimated using statistical modelling approaches. First results showing significant effects for encapsulant, backsheet and the combination of both on the initial power output are presented. In general, modules with polypropylene‐based backsheets have a higher initial power (PMPP) than those with PET‐based backsheets, with the combination of thermoplastic polyolefin (TPO) encapsulation material and polyolefin backsheet being superior to the other material combinations. A comparison of the material‐dependent degradation rates obtained from the mixed‐effects models clearly shows that the degradation rate upon damp heat exposure for modules with EVA is significantly larger than that using polyolefin encapsulants. The derived relations aim to provide valuable input for innovative material developments as well as predictive maintenance specifications.

of copper or silver wires/ribbons, 3 which are encapsulated/protected by a polymer. 4,5 The conductive wires/ribbons are either soldered or fixed to the cell using electrically conductive adhesives. 6 The polymeric encapsulant is covered on the front with a glass pane and on the rear with either another glass pane or a polymeric backsheet.
In most cases, panels have frames that provide structural support, aid in module fixation and seal modules edges. This module architecture based on a layered encapsulation structure was developed and proven to protect the solar cells and their connecting wires from the stresses caused by the sometimes harsh environmental conditions in which the modules are operated. [7][8][9] Furthermore, this layered encapsulation structure provides mechanical stability and high functionality combined with optimised power output and electrical safety. 1,2,10 In the recent years, PV technology saw many changes, 1,10 amongst others, the rise of new cell technologies, the trend towards bigger wafer and module sizes, half cells, new interconnection approaches like shingling or multi-wires and new polyolefin-based encapsulant and backsheet films. [11][12][13][14][15][16] The main motivation and benefits for new material developments have been mainly cost reduction but also performance increase, sustainability and legal regulations as well as new technological requirements. 1 In particular, the role of polymers in PV modules has been underestimated for a long time, as they do not play an active role in power generation itself. However, the choice of polymers has a significant influence on the properties of the PV modules not only for efficiency but more importantly for reliability, as most PV module degradation modes are directly linked to polymer degradation and material interactions with polymer components. 2,17,18 New materials have to prove their performance compared to the established standard materials in terms of the electrical performance of modules as well as long-term stability and reliability. 19 Since PV modules are currently expected to have a service life of more than 25 years, the long-term stability of the new materials must be tested and evaluated using accelerated aging tests and degradation modelling. This task becomes even more challenging when a service life of 40 years or more is targeted. 20,21 These long lifetimes can currently only be postulated. The development of long lifetimes is complicated further by a precedented speed of change in materials being incorporated into PV modules. 1,20 With current production capacities of around 175GW, 22 many gigawatts of modules with new technologies and materials can be produced and installed without having sufficient experience about long-term reliability. In the worst case, this has led to unexpected degradation mechanisms several years after field deployment, which were not predicted in laboratory accelerated testing, such as potential-induced degradation 23 (PID) or backsheet cracking. 20,24 Modelling approaches, both for prediction and inference, could be the foundation for developing devices with such lifetimes. 20,25 Springer et al. elaborate that current tools fall short in accurately assessing degradation mechanisms and failure modes over such extended periods. They propose a unifying modelling framework to enable a holistic assessment of PV module reliability and to accelerate the learning cycle for new developments. 25 Jordan et al. point out the need to detect and understand failures and underperformance more quickly requiring more investment in the degradation science and physics-of-failure. However, this requires libraries of degradation mechanisms, material property models and stressor models and not just field observable failure modes. 20 Typical degradation models used for PV modules and PV systems are using electrical performance data. On system level, where the degradation mode is unknown and therefore physics-or chemistry-based analytical models cannot be used, various statistical models are applied, as described by Lindig et al. 26 If monitoring data are available at module level, also analytical models can be applied to model power degradation for specific degradation modes like corrosion or PID. 18,26 However, none of the presented approaches have been used to investigate and especially quantify the influence of different material compositions on efficiency and reliability of PV modules. In recent years, more focus was given to the modelling of material degradation in PV modules, either trying to link material degradation to power loss or to better understand the complex degradation mechanisms and material interactions. 18 Network structural equation modelling was used to assess statistically significant relationships between applied stressors, mechanistic variables and performance level responses. [27][28][29] Using network structural equation modelling (netSEM), Gok et al. 30,31 investigated the degradation mechanisms of PET films used for backsheets and were able to construct a set of degradation pathway network models. Nalin Venkat et al. 32 used netSEM for modelling the degradation of glass/backsheet and double-glass PV modules using different encapsulant films. This method allows for a combination of electrical performance data with material testing results and therefore can be very helpful in understanding degradation pathways and the influence of material degradation on power output. However, netSEM needs time series as input data and therefore cannot be used for quantifying the influence of certain material combinations on initial power output of a PV module.
Ascencio-Vasquez et al. 33 evaluated three polymer degradation mechanisms (hydrolysis degradation, thermomechanical degradation and photo-degradation) and the total degradation rate of PV modules due to the combination of temperature, humidity and ultraviolet irradiation for different climatic conditions. 34,35 However, this approach did not account for changes in microclimates based on different encapsulant-backsheet combinations.  36 The characteristics of the test modules were measured before, during and after the accelerated aging tests. As part of the project, the measurement results were evaluated by comparing the aging-related changes in the electricaland material-related properties and deriving basic relationships. 11,12,16 These test results are now further evaluated using modern techniques from statistics and data science within the follow-up Austrian R&D project ADVANCE!. 37 ADVANCE! explores the potential of innovative and complex statistical and machine learning data processing methods for digital analysis and improved modelling of the time-and stress-dependent performance (degradation and reliability) of PV modules. 38 The main objective of this paper is to better understand and quantify the effects of the encapsulation and backsheet type on the electrical performance and the degradation behaviour of the test modules. To achieve the objectives, two different statistical learning approaches have been adapted and implemented: (i) A two-way analysis of variance (ANOVA) is performed in order to identify significant differences between the initial power of the modules with different material combinations.
(ii) Mixed effect models are used to determine the influence of encapsulant and backsheet composition on electrical degradation.
These correlations are aimed to provide valuable impact for innovative material developments and for predictive maintenance specifications.

| EXPERIMENTAL
The starting point is a comprehensive database that was generated in the Austrian flagship project INFINITY (2015-2018). 36 It consists of extensive measurement and characterisation data from more than a hundred sample modules. After a light stabilisation step, these modules were characterised in their original state and then subjected to precisely defined accelerated aging scenarios. Several intermediate characterisation steps were performed during and after the aging tests. These existing data time series of multiple characterisation methods are used to derive internal causal relationships. The focus is on the relationships between the aging of materials and material composites and the electrical module performance.  Table 1). Table 2 and Table 3 summarise the main characteristics of the selected encapsulants and backsheets. The encapsulants differ mainly in their laminating behaviour (crosslinking vs. thermoplastic) and in their chemical degradation mechanism (e.g. tendency to form volatile acidic degradation products: EVA [39][40][41][42] or not: polyolefins).

| PV module composition
The two selected backsheets differ not only in material composition but also in the production process (laminated vs. co-extruded films). The laminated backsheet (PPF) consists of two layers of polyethylene terephthalate (PET), with the core layer being coated with an additional 5-μm thin fluorinated layer. The co-extruded backsheet (CPO) consists of three differently modified polypropylene (PP) layers.
Apart from the reflectivity, the biggest difference between these two backsheets can be found in the permeation behaviour, the water vapour transmission rate WVTR, the oxygen and acetic acid transmission rate (AATR). Whereas the PET backsheet (PPF) exhibits good barrier properties towards oxygen, water vapour and acetic acid, the co- oxygen and acetic acid. This enables a fast escape of the corrosive acetic acid from the PV module and at the same time prevents a high ingress of water vapour from the environment into the module. 43,44 Detailed information on sample production, the composition and properties of the encapsulants 11 and backsheets 12,16 can be found in three previously published papers by the same group of authors.
Three samples each were used for three different accelerated aging tests, 11,16 and the fourth set of three modules was retained as a reference. This paper focuses on the results of the damp heat test, originally performed at 85 C and 85% R.H. for 1000, 2000 and 3000 hrs. In order to obtain better models for the performance degradation, still available samples were further exposed to damp heat con-

| Data engineering
As described above, the measurement/characterisation results and all related sample history information are the most important inputs for data science analysis. The collection of all this information in a database is necessary in order to be able to use established analysis methods.

| Input information characteristics
All measurement parameters and data output formats were defined in advance of the measurements for the various characterisation methods. This was necessary because measurements were carried out by the project partners at different locations with different devices, which led to different, system-specific file formats.

| Data allocation
Since new measurement information is available after each aging cycle, the database must be generated regularly (previously 27 times) and made available for ongoing analysis. As described above, in addition to the measured numerical values, scalar descriptors were also derived from (i) image information (EL, UV fluorescence imaging) and (ii) spectral information (IR and UV fluorescence spectroscopy). This information generated in post-processing then also had to be integrated into the database and was used for the detailed analysis of aging indicators. The generated database is provided on a projectspecific server, metadata is managed with GitLab. 45 The work presented here focuses only on the influence of aging on the electrical power at P MPP , further publications on the descriptors generated from images 46 and the spectral descriptors are in progress.  The initial power of the modules is displayed in the boxplots in F I G U R E 2 Initial power at STC for the different combinations of backsheet and encapsulant material.

| Database scheme
A two-way ANOVA is performed in order to identify significant differences between the initial P MPP of the modules with different material combinations, also referred to as factors. ANOVA assesses whether the variation in the observed data points can be accounted to the given factors (material combinations). This translates to comparing the means in the different groups and testing them for equality.
It is achieved by calculating the ratios of the within-and betweengroup variabilities, the so-called F-statistics. 51 Given equal group means, the F-statistic follows an F-distribution with two parameters.
Those parameters are the degrees of freedom derived from the within-and the between-group variability.  If ANOVA returns significances for the factors and/or their interactions, this means that at least two of the group means differ from each other. In this case, a posthoc analysis is carried out to identify the factors and factor combinations that cause these differences. The so-called contrasts describe the differences in means between two chosen material combinations. In the presence of significant interactions, the effects of the two materials cannot be treated independently. In this case, only the interaction effects can be interpreted but not the baseline effects of the factors.
The results from the posthoc analysis are displayed in Table 5.

| Modelling of influence of encapsulant and backsheet composition on electrical degradation (mixed-effects models)
The second goal of the analysis is to quantify the degradation rates of the modules with the different material combinations. The modules were exposed to up to 6000 h of damp heat conditions and characterised every 1000 h of exposure. For the analysis, P MPP at STC is normalised to start at one for all modules by dividing by each module's initial P MPP . Thus, the differing initial powers do not influence the estimated degradation rates. The resulting P MPP values are displayed in An artefact in all curves is an initial power increase due to an increase in transmittance of the encapsulant upon artificial aging. 9 In order to separate this increase from the degradation rate, an F I G U R E 3 P MPP degradation of the modules with different material combinations after up to 6000-h exposure to damp heat conditions (normalised values of P MPP ). The observations marked by 'x' are used for model validation after model fitting.
additional function called Ramp is introduced to extend the regression model. It increases up to 500 h and is then fixed at 1 for the rest of the exposure time: where 1 condition f gis the so-called indicator function. It evaluates to one if the condition in brackets is fulfilled and zero else. We include the ramp as a fixed effect. As the power increase varies for each module, we additionally model it as a random effect on module level.
Throughout our analyses, the overall power increase is estimated to be roughly 2.5% of the initial module power. 9,38

| Encapsulant-specific degradation rates
Let pmpp i,j,t denote the normalised P MPP of module i with backsheet material j at time t. The model equation is then given by The fixed effects are denoted by β 1 to β 4 . While β 1 separates the initial power increase from the actual power degradation, β 2 , β 3 and β 4 describe the encapsulant-specific degradation rates in squared exposure time (see Figure 4). The backsheet effect is modelled as a random effect, denoted by b j . Thus, depending on the backsheet material, the degradation rates obtained from the fixed effects are corrected either up-or downwards. Furthermore, the random effects a i describe the module-specific power increases, which vary from module to module (and material to material). The intercept, corresponding to the initial module power, is fixed at 1. The model output is depicted in Figure 5. The estimated degradation rates support the trend observed in Figure 3.
The coefficient estimates and their confidence intervals show that the degradation rate for modules with EVA is significantly larger than that using TPO and POE encapsulants. The degradation rates for TPO and POE variants do not differ significantly from each other.
Additional experiments could help to further differentiate the degradation rates between the two polyolefin encapsulants.
The realisations of the random effects disclose some further information on the structure of the data (see Figure 5)

| Backsheet-specific degradation rates
In analogy to the previous subsection, let pmpp i,k,t denote the normalised P MPP of module i with encapsulant material k at time t. The model is described by the following equation: The fixed effects are denoted by β 1 to β 3 . Again, β 1 separates the initial power increase from the actual power degradation, and β 2 and F I G U R E 4 Results of model fitting for encapsulant-specific degradation rates. Estimated degradation rates (left) and estimated coefficients ß 1 -ß 4 with 95% confidence intervals (right). The model output is displayed in Figure 6.

| Model validation
In order to assess the model's predictive ability, we compare the root mean squared errors (RMSE) for the model-fitting period (measurements up to 5000-h exposure time) and for the predictions for the measurements after 6000 h of exposure. The resulting RMSE, the observation variance and their quotient ('relative RMSE') are displayed in Table 6.
The relative RMSE is comparable for the model-fitting period and the predicted values. Hence, the predictive performance of the model is good. The larger absolute RMSE can be explained by the increasing variation of the measurements with increasing exposure time.
F I G U R E 5 Realisations of the random (module-and material-specific) effects for the encapsulant model.

F I G U R E 6
Results of model fitting for backsheet-specific degradation rates. Estimated degradation rates (left) and estimated coefficients with 95% confidence intervals.

| Discussion
The modelling results (3.1) on the influence of encapsulant and backsheet composition on module power before exposure (analysis of variance) clearly show (see Figure 3, Table 4 and Table 5) that there are significant effects for encapsulant, backsheet and the combination of both. The combination of TPO encapsulant with CPO backsheet has a significantly higher initial power than all other material combinations (see Figure 2).
Generally, modules with CPO backsheets tend to exhibit a larger initial P MPP than those with PPF backsheets. The main reason for this effect is seen in the improved reflectance of the CPO (80.4%) compared to the PPF backsheet (68.4%; see Table 3).    Figure 8, left) this cell corrosion effect is visible as dark areas. The integrated area of the dark corrosion areas is directly proportional to an increase in the serial resistance R s and a loss in the P MPP.

9,63
Surprisingly, the EVA encapsulated cells in the test modules with the CPO backsheet (see Figure 8, right) did not show the effect that pronounced after the same exposure time (4000-h DH). The decisive difference between these two backsheets is their deviating permeation properties in respect to acetic acid, AATR. 16,39 As described already in the experimental part (table 3), the co-extruded PP backsheet (CPO) exhibits a more than 10 times higher AATR than PPF.
Thus, the formed acetic acid is not accumulated in the encapsulant but can escape from the PV module with a high rate.
However, when comparing the EL images of two EVA-PPF modules given in Figure 8, a clear difference is obvious. This effect is also F I G U R E 7 UV transmission spectra of the three encapsulants EVA, TPO and POE; the normalised spectral response (SR) of the silicon solar cell multiplied with the AM1.5 standard solar spectrum is printed as orange dotted curve.
T A B L E 6 Absolute and relative root mean squared errors of model fit by exposure time.  tended to exhibit a larger initial P MPP than those with PPF backsheets, assumedly due to its higher reflectance values.
The second goal of the analysis was to quantify the degradation rates of the modules with the different material combinations. The modules with EVA encapsulant show a stronger degradation than those with TPO and POE for both backsheet materials. In order to quantify the degradation rates, linear mixed-effects models were used. The degradation rate for modules with EVA was found to be significantly larger than that using TPO and POE encapsulants. The deg-