Results of the IX International Spectroradiometer Intercomparison and impact on precise measurements of new photovoltaic technologies

5.1 Direct normal irradiance

In this section, an analysis of direct normal irradiance (DNI) spectra is presented. Each participant followed its own measurement timing and protocol, but in general, everyone was able to measure spectra every minute or few minutes. However, not all participants were able to measure throughout the entire day: Several participants were able to measure only in several time slots, due to difficulties in reaching the observatory (special transport means allowed only few passengers per time). To select spectra set suitable for comparison, the following filtering criteria have been applied: (1) check of sky stability condition as described above; (2) for each slot with stable sky, the number of available spectra per participant has been calculated; and (3) stable slots with the highest number of acquired spectra have been considered for further analysis. It has been found that for those participants measuring DNI the days with most available measurements were June 26th and 27th; on the 26th, the comparison is done between participants Labs A, C, E, and G in the time slot 11:57–12:45 (UTC + 1); on the 27th between Labs A, B, C, D, E, F, and H in the time slot 10:45–11:10 (UTC + 1). For each time slot, the acquired spectra for each participant have been averaged to have a single spectrum to be used for wavelength and bands analysis, once having checked that all the spectra of each participant were close to each other (Figure 2), and eventual outliers and artifacts removed. Figures 2 and 3 show all the spectra acquired in the time slot 11:57–12:45 (UTC + 1), day 26th and their averages divided per participant. The total number of spectra considered for calculation in this time slot is 141.

Analyzing the graphs, the first consideration is that there are not two participants measuring in the same wavelength range, they are all four different; secondly, Lab E spectra are characterized by a drop after 1,100 nm, probably due to artefacts on the merging of raw data given by the detectors of the instrument; finally, Lab E measurements were systematically lower in the region 400–700 nm with respect to the other three participants. During this time frame, the AM varied between 1.07 and 1.11. Taking the average spectrum measured by Lab A as reference, the following graph shows the deviations of Labs C, E, and G from the reference (Figure 4). The common wavelengths for these three spectra are within 350–1,025 nm; all of them have been interpolated on an uniform grid with step 0.1 nm. A first visual comparison evidences a very good agreement between the reference and Lab C, being within ±5% on almost the entire spectrum; Lab E has better agreement after 700 nm; Lab G is generally within ±10% except the band 700–800 nm. In all cases, there are spikes in the UV part before 400 nm, but this is a common fact in almost all spectra comparisons due to the very low signals of detectors in the UV and to the calibration issue concerning the very low emissivity of halogen standard lamps in 300–400 nm region.

Of the three participants measuring in this time slot on day 26th, only Labs A, C, and G provided their measurement uncertainties; therefore, performance statistics analysis is limited to them in this time slot. Here and in all sections of this paper, the key parameter for performance statistics analysis are the E_N numbers as defined in the ISO/IEC 17043 standard¹⁵:

(1)

where x is the measured value by a participant, is the reference value, and U_C(.) is the expanded uncertainties (k = 2 in case of Gaussian distribution); this equation assumes the statistical independence between x and . The E_N numbers are calculated for all wavelengths in the common range 350–1,025 nm and are shown in Figure 5.

The classification in satisfactory or unsatisfactory depends not only on the difference between the two measured values of the same quantity but on the difference evaluated with respect to the declared uncertainties; it is possible therefore that a participant distant from the reference obtains satisfactory results while another participant closer to the reference obtains unsatisfactory E_N scores, if the latter uncertainties are too small respect to the difference. Considering this aspect, performance statistics is a useful tool for each laboratory to assess its own calculated uncertainties. In the specific case, the agreement is unsatisfactory in the UV (which is a common issue) and after 700 nm. Throughout all the analyses, the reference for EN numbers calculation is Lab A, which has been used as reference in all the previous intercomparison campaigns. The consistency between Lab A and the average of spectra of all participants is proofed in Figure 6, showing the EN numbers between Lab A and the average (taken as reference with its calculated uncertainty) of the spectra acquired in the second time frame (having more available data).

The second time frame considered above was on day June 27th between 10:45–11:10 (UTC + 1) with Labs A, B, C, D, E, F, and H measuring DNI. The AM in this frame varied between 1.07 and 1.09. The selection of spectra useful for comparison follows the same methodology already described; the suitable selection of data is shown in Figure 7.

At a first sight, some considerations can be inferred: First of all, again, all participants measured in a different wavelength range, and in particular, one participant's spectra are extended up to 4,000 nm; in this case, the instrument is not a spectroradiometer but a filter radiometer, and the part above 1,200 nm is mathematically reconstructed; this instrument therefore is suitable for outdoor use. All spectra except Lab B have their peak around 1.6 W m⁻² nm⁻¹; however, this discrepancy in the absolute values does not necessarily prelude to a difference in shape. Lab E confirms the possible artefact at around 1,100 nm, which might be due to the merging algorithm between the two sub-spectra portions acquired by the detectors of the instrument. Last, Lab D measurement is dominated by noise after 1,000 nm. These considerations shall be quantified by the wavelength-to-wavelength comparison, which is performed using the average spectrum of each participant in this time frame, having acknowledged the stability of irradiance in this period. The total number of acquired spectra for this time frame analysis is 331.

In order to compare the absolute values point-by-point, the same approach used on day 26th with different participants has been used. The common wavelength range is the same, limited within 350–1,025 nm, then interpolated at 0.1 nm step to define with sufficient precision the measured peaks. Figure 8 shows the differences for all participants using Lab A as reference.

Although a numerical quantification of the agreement between measurements will be inferred using E_N numbers, from the graph above Lab B confirms to be considerably higher than all the others, constantly above +10% respect to the reference. The reference is close to the average value of all spectra at least until 700 nm, being that the differences are distributed equally above and below the horizontal axis, whereas after 700 nm, probably the reference is slightly above the average. There is a good agreement of Lab C and Lab H to the reference and between themselves, whereas Lab F tends to deviate linearly in the NIR part of the spectrum. Lab F stated that these measurements were unintentionally taken without the internal calibration files of the spectroradiometer manufacturer due to software problems. Consequently, within these data, the nonlinearity correction of the detector as well as the correction of the wavelength dependent transmission losses is not applied. This could be the origin of the observed strong deviation in the NIR part of the spectrum. Comparing Figures 8 and 4, it seems that for DNI spectral measurement all participants are generally within ±10% respect to the reference, with only one outlier. It has to be reminded that the W-to-W comparison method is by far the most analytical, because there is no filtering action on outlier values as happens analyzing spectral band contents where integrated ratios relative to each spectrum integral are computed. Considering the expanded uncertainties provided by participants, for Labs B, C, and F, the performance statistics analysis can be performed. Looking at Figure 9, it is evident that the mutual uncertainties between Lab A and Lab B are underestimated, being Lab B completely above the “good agreement” limit of E_N = 1; regarding Lab F, the agreement with the reference is good except for some outliers in the NIR part of the spectrum. A hypothesis for this behavior can be due to dark measurements procedures and uncertainties, which play an important role in outdoor measurement campaigns where sensors internal to the instruments are continuously exposed to sunlight, and also, the instrument body is not kept at room temperature like in indoor operations. Therefore, the two Labs might have used different routines for dark measurements, whereas the mutual uncertainties seem to be quite well in agreement in this case. For Lab C, the agreement is generally good except for outliers in the NIR and UV regions.

5.2 Global normal spectral irradiance analysis

Three laboratories measured global normal irradiance (GNI): Lab A (as reference), Lab D, and Lab I, using three systems with very different characteristics: Lab A used a system without collimating tubes composed by three detectors, Lab D an instrument with rotating shadow band, able to measure contemporarily GNI, DNI, and diffuse spectral irradiances; Lab I used a three-channels instrument, equipped with integrating sphere and optical fiber. The instruments of the three participants were mounted on different trackers. The experimental approach to measure GNI was the same as DNI measurements and the data analysis techniques as well. In this case, the best period in terms of sky stability and number of available data was on day 27th between 13:40 and 14:10, with AM comprised between 1.20 and 1.28. In Figure 10, the spectra selected for comparison are shown, for a total of 89 spectra, together with the averages per participant. Once again, the irradiance stability measured by the reference pyrheliometer is confirmed by the acquired spectra, which are very close to each other; however, looking at the bottom right plot in Figure 10, Lab I peak point around 1.4 W m⁻² nm⁻¹ seems too low. All cosine diffusers connected to optical fibers were mounted onto the same tracker, equipped with a mounting structure for the alignment, so it is improbable this difference to be due to misalignment; one of the possibilities could be related to the spectroradiometer calibration. Looking at the W-to-W comparison of the average spectra shown in Figure 11, Lab D appears well in-line with Lab A (being the difference within ±10% throughout all the range except for UV before 350 nm and NIR near 1,100 nm), whereas Lab I measured around 20%–30% lower. This considerable difference however might not be dramatic in terms of bands (spectrum shape matters) and consequently also its impact on spectral mismatch factor (MF) calculation. Lab I has previously reported that the in-house calibration was performed with a deuterium lamp with primary calibration (traceable NPL) and an halogen lamp with unreliable calibration, provided by the lamp manufacturer. Lab I has later on performed interlaboratory comparisons and found a systematic deviation in the VIS/NIR range, which is consistent with the deviations observed in this section.

In terms of E_N numbers, only Lab I provided measurement uncertainties, and therefore, the following graph shows only the performance statistics between the reference (Lab A) and Lab I. However, excluding Lab D, this operation can be extended up to 1,700 nm. From the result shown in Figure 12, it is evident that are in agreement only in the UV range, where Lab I has calibrated deuterium reference; in 400–1,100 nm, the measurement is affected by a systematic deviation arising from the biased in-house calibration.

5.3 Global horizontal spectral irradiance

For this analysis, the best time slot was during day 27th from 11:55 to 13:30 (UTC + 1). Figure 13 shows all the acquired spectra in this time frame, for a total of 196.

The graph of the average spectra (Figure 13) evidences a slight change in shape near 500 nm, where Lab A crosses Lab J spectrum; this effect is better appreciated in the point-by-point analysis below (Figure 14).

The agreement between the spectra is generally within 5% excluding the portion below 400 nm and around 1,400 nm due to the very low values there. Compared with previous absolute values comparisons, this agreement can be considered good. To assess the agreement of the two average spectra with the two mutual uncertainties, the E_N numbers have been calculated; they are shown in Figure 15; it is evident that the uncertainties have been correctly estimated in the 300–400 nm part, whereas near 700 nm, they have been underestimated. The E_N < −1 after 1,600 nm are attributable mainly to noise.

5.4 Spectral contents in bands

For most of the calibration involving a PV device, only the knowledge of the spectrum shape of the light impinging on the reference and on the device under calibration is needed. One notable exception is the calibration of reference cell using the integrated spectral irradiance method. Spectroradiometers in photovoltaics are mainly used for classifying light sources (normally solar simulators) according to IEC 60904-9 or calculating the spectral mismatch factor, which will be described in the next section. In this section, the relative spectral content of the spectra in several defined bands of interest is analyzed: Two different sets of bands have been chosen for the analysis: The first is defined in the present IEC 60904-9 Ed.2 (see Table 4) and the second in the future IEC 60904-9 Ed.3, which at this moment is under approval process at the International Electrochemical Committee. Both of them are (will be) used to classify solar simulators accordingly to their match with the AM1.5G standard spectrum defined in IEC 60904-3, but the second one will extend the range from 400–1,100 nm to 300–1,200 nm; this last is a more appropriate range for the calibration of recent PV device technologies, exhibiting a wider range of spectral responsivities.

The principle used to define the new bands is the energy-equivalent content of all bands respect to the AM1.5G spectrum (each band contains about 16.6% of the spectral content of the 300–1,200 nm region); the challenge for the spectroradiometry community working in PV concerns the measurement range: Not all the spectroradiometers participating to the ISRC were able to measure on the entire 300–2,500 nm range.¹⁶ The relative spectral content is calculated by integration using

(2)

with the integration extrema (a_i, b_i) specified in the above table, and E(λ) the measured spectrum. From Equation 2, it can be seen that multiplying the spectrum by a constant factor does not influence the relative spectral content, and therefore the correctness of the spectrum in absolute terms is not necessary, where effects distorting the shape if the spectrum are taken into account. The measurement uncertainties used previously (Table 3) cannot be directly transferred to the new quantities , because of the effects due to integration. Several works have been done to study the error propagation between measured quantities and functionals having them as variables; apart for very basic cases, an analytical expression of the derived uncertainty is not possible, and numerical methods involving Monte Carlo simulations are more and more frequently used.

5.5 Direct normal spectral irradiance

The same spectra sets used for the point-by-point absolute analysis will be analyzed in this section to assess the congruency between participants in terms of relative spectral content. Therefore, 472 spectra will be used in this section. For analysis in bands, the average spectrum for each participant in each time frame is considered. Figure 16 shows the calculated relative contents for both the two bands sets specified in the two versions of IEC 60904-9. Here, Lab A is not acting as reference, and results of the other participants are not evaluated against it. In Figure 16, each group of columns represents the content for a specific band; numerically, above each group of columns, the average and max-min deviation of the corresponding values are indicated. In this case, given the very low number of points, the max-min deviation has been preferred to the standard deviation. Not all the participants comparing in the W-to-W analysis (see Figures 3 and 7) could be considered here: Looking at Table 5, only those participants complying with the different range requirements are considered here. Looking at the results, the deviations for each band are all within ±1%, although some of the spectra had considerable differences W-to-W, in the order of ±10% or even more. This uncertainty reduction is partly due to the effect of integration, which smooths errors due to spikes, and partly to the normalization (denominator of Equation 2). The spectral content looks very different using Ed.2 bands respect to Ed.3 bands: In the former, the energy content decreases going to the right part, whereas in the latter, it remains almost constant (the bands of Ed.3 were thought to have a constant contribution for the AM1.5G spectrum; here, the spectra were DNI close to AM 1.1). The agreements shown in Figure 16 look quite comfortable in view of using these instruments on solar simulators, but only a few of them are able to fully cover the 300–1,200 nm range (typically, this problem concerns single-detector instruments). In particular, amongst the three participants on day 26th measuring DNI, the comparison is possible only using bands of IEC 60904-9 Ed.2, whereas only Lab A is able to cover the Ed.3 range, and therefore this comparison is absent from Figure 16.

5.6 Global normal and global horizontal spectral irradiance

The same approach is followed to analyze the relative spectral content of GNI spectra acquired on day 27th (Figures 10 and 13). The comparison according to Ed. 2 has been possible amongst all the three Labs measuring GNI, but only two of them covered the full range 300–1,200 nm required by Ed.3 (Figure 17), whereas for GHI measurements, both Lab A and Lab J cover the 300–1,200 range. Once again, the deviations are in the same order of those for DNI. Remarkably, the range where the largest deviation is observed (0.9–1.0%) is 300–470 nm, where most of the calibration lamps have poor irradiance content, and therefore, calibrations have larger uncertainties. The GHI spectra show larger deviations in relative spectral content, more evident below 400 nm but low in the last band up to 1,200 nm (Figure 18).

6.1 Spectral corrections from DNI spectra

Photovoltaic devices can be calibrated outdoor under natural sunlight both measuring the irradiance in GNI or DNI; in any case, the reference spectrum is the AM1.5G. Here, all the suitable DNI spectra, covering at least the range 300–1,200 nm, measured during day 27th have been used to calculate the mismatch factors for the different technologies against a c-Si reference cell. This is a useful exercise to assess the variation of M_F during a single day for the different technologies and spectra (Figure 20). The crystalline silicon device has been considered to act as “known-case,” to check the validity of the algorithm to compute the others; as expected, all participants' mismatch factors are within ±0.13% (half max-min in the central part of the smile) so very close to each other. The differences amongst laboratories considering other new technologies are higher; for Perovskite, the difference is in the order of ±1.6%, for DSSC ±1.9%, and for OPV devices even ±2.2%. Looking at the M_F, the entity of the correction to the measured IV curve can be appreciated: This would be less than 0.5% for c-Si for all participants, from −2% to +5% for a DSSC device, from −1% to +4% for perovskite, and from −4% to +5% for OPV, depending on the specific laboratory. Given that the IV correction equation is linear respect to MF, these differences would be directly translated into STC performance discrepancies. However, in a real outdoor calibration campaign, most of them would be disregarded after having filtered them for irradiance stability and outliers detection. However, looking at all the subplots of Figure 20, there is a clear tendency of participants to be positioned in the same order respect to the others for all technologies, with Lab F tendentially higher than the other three.

6.2 Spectral corrections from GNI spectra

The same approach has been followed to compare the impacts on spectral mismatch factor of GNI spectra, measured on day 27th. Figure 21 is the analog of Figure 20 for GNI spectra. In this case, the three participants are not always in the same relative position on the graphs; in the case of c-Si, the larger deviation observed between Lab A and Lab D (than the deviation between Lab A and Lab I) can be explained by the fact that Lab D instrument does not reach 1,200 nm, resulting in larger spectral mismatch. The differences amongst laboratories considering other new technologies are in this case comparable to c-Si; for Perovskite, the difference is in the order of ±0.6%, for DSSC ±0.5%, and for OPV devices ±1.1%, being the only technology leading to a difference between laboratories above 1%. These corrections would lead to corrections in the order of half a percent, except a 1% correction for OPV. Having AM1.5G as reference spectrum, a GNI measurement leads to a spectral correction of minor entity, because GNI spectra are closer to the AM1.5G than DNI spectra.