Variability and associated uncertainty in image analysis for soiling characterization in solar energy systems

Description

The accumulation of soiling on photovoltaic modules and on the mirrors of concentrating solar power systems causes non-negligible energy losses with economic consequences. These challenges can be mitigated, or even prevented, through appropriate actions if the magnitude of soiling is known. Particle counting analysis is a common procedure to characterize soiling, as it can be easily performed on micrographs of glass coupons or solar devices that have been exposed to the environment. Particle counting does not, however, yield invariant results across institutions. The particle size distribution analysis is affected by the operator of the image analysis software and the methodology utilized. The results of a round-robin study (herein called RR2) are presented in this work to explore and elucidate the uncertainty related to particle counting and its effect on the characterization of the soiling of glass surfaces used in solar energy conversion systems. An international group of soiling experts analysed the same 8 micrographs using the same open-source ImageJ software package. The variation in the particle analyses results were investigated to identify specimen characteristics with the lowest coefficient of variation (CV) and the least uncertainty among the various operators. The mean particle diameter showed the lowest CV among the investigated characteristics, whereas the number of particles exhibited the largest CV. Additional parameters, such as the fractional area coverage by particles and parameters related to the distribution’s shape yielded intermediate CV values. The cleanliness level (L) has also been considered, based on a prior publication in which the IEST-STD-CC 1246E standard was used to describe the cumulative distribution versus the equivalent particle diameter of the deposited particles or contaminants. One prior study that serves as a background for this work is Smestad, G.P., Germer, T.A., Alrashidi, H. et al. Modelling photovoltaic soiling losses through optical characterization. Sci Rep 10, 58 (2020); https://doi.org/10.1038/s41598-019-56868-z. Together, these results can provide insights on the magnitude inter-lab variability and uncertainty for optical and microscope-based soiling monitoring and characterization. This data was utilized for the publication of the same title published in Solar Energy Materials and Solar Cells (https://doi.org/10.1016/j.solmat.2023.112437). Herein are the micrographs (Micrographs folder), the methods used (ImageJ Methods folder) and some of the results (see the Tables folder).

Steps to reproduce

A set of several glass coupons were attached to the front surface of tilted photovoltaic modules at locations in eight different countries with arid regions. These were in, or near, the Sonoran Desert in the U.S., the Atacama Desert in Chile, the Sahara Desert in Morocco and Algeria, the Great Arabian Desert in Qatar and Jordan, the Simpson Desert in Australia and the arid region of Cape Verde. The coupons were slides made of high-quality Duran soda-lime glass (76 x 26 mm according to ISO 9037-1) from the DWK Life Sciences, LLC. They had a material thickness of 1 +/- 0.05 mm and a hydrolytic class of 3. At predetermined time intervals, coupons were taken off the module and later shipped to the Fraunhofer CSP for analysis. Micrographs were taken using a Carl Zeiss Axio A1 microscope with dark field (DF) illumination. The objective lens was 20x magnification, resulting in a scale of 3.156 pixels/μm. The images had a dimension of 1388 pixels x 1040 pixels and were saved in “bmp” format. Each expert (ImageJ operator) was provided with the same set of eight micrographs. They were asked to set the aforementioned scale and generate a results table containing the areas of all the detected particles in each micrograph using ImageJ [https://doi.org/10.1038/nmeth.2089], or Fiji [https://doi.org/10.1038/nmeth.2019]. They were instructed to repeat the analysis twice: one set of results was to be produced using automatic settings for the threshold function (files denoted herein by A), while the second set was to be generated by manually setting the threshold values (files denoted herein by M). No other instructions were given. The ImageJ operators are denoted using consecutive numbers, from 1 to 11 (O1, O2, O3, …, O11). For the automatic approach to the thresholding, operators O4 through O11 set the auto threshold to the Default Dark condition. This is accomplished via the menu sequence: Image › Adjust › Threshold from the menu and then selecting Default, Dark background and “Auto.” There are several methods to set the threshold automatically. These include the aforementioned Default Dark, as well as Otsu dark, Triangle and several others. There is also a Auto Threshold available via the menu sequence: Image › Adjust › Auto Threshold [https://imagej.net/plugins/auto-threshold (accessed March 3, 2023)]. Operator O3 utilized that method. The analysis of the ImageJ results returned by each operator was carried out using a custom program written in Python 3.7. The values for various parameters for a specific micrograph were tabulated for each of the operators. There were thus 8 tables, one for each micrograph. A mean and standard deviation was obtained for each parameter. Then, a coefficient of variation (CV) was calculated as the ratio of the mean to the standard deviation. Another table was then formulated that gave the CV values for the parameters versus the eight micrographs. The Tables folder herein contains the results of the study.

Institutions

Categories

Funding

Related Links

Licence