Statistical and domain analytics for informed study protocols

Abstract

To optimize and extend the lifetime of photovoltaic (PV) modules, a better understanding of the modes and rates of their degradation is necessary. Lifetime and degradation science (L&DS) is used to better understand degradation modes, mechanisms and rates of materials, components and systems in order to predict lifetime of PV modules. Statistical analytic methods were used to investigate the relationships between various subsystem characteristics related to suspected degradation pathways, as well as their impact on changes in module performance. A PV module lifetime and degradation science (PVM L&DS) model developed in this way is an essential component to predict lifetime and mitigate degradation of PV modules. Previously published accelerated testing data from Underwriter Labs, featuring measurements taken on 18 modules with fluoropolymer, polyester and EVA (FPE) backsheets, were used to develop the analytical methodology. To populate this dataset, three performance characteristics for each module were tracked over a maximum of 4000 hours while the modules were exposed to stressful conditions. Two of the eighteen modules' performance characteristics were measured with no exposure to stress, and then dissassembled immediately to provide baseline measurements. Eight of the sixteen remaining modules were exposed to 85% relative humidity at 85°C (Damp Heat, DH) and the final eight were exposed to 80W/m2 of ultraviolet light at 280-400nm wavelengths and 60°C (UV). Four of the sixteen modules being exposed (two from DH conditions and two from UV conditions) were removed at each 1000 hour time point and disassembled to provide observations for eleven component level experiments, six directly related to degradation mechanisms and five to material performance characteristics. The resulting dataset comprised of coincident observations of 15 variables (time, three system-level performance variables, and eleven component-level variables) was statistically analyzed using the developed methodology. Limitations in the quantity of coincident observations constrained the statistical study to require the use of domain knowledge to pre-select a subset of variables for analysis, which introduced undesirable bias and prevented the full development of a prognostic model from this dataset alone. The results and lessons learned help guide the experimental design for better structuring further accelerated and real-world experiments, providing necessary insight in order to sample data effectively and efficiently, obtain maximum information for identifying statistically significant relationships between variables, and develop a PVM L&DS model construction methodology to determine degradation modes and pathways present in modules and their effects on module performance over lifetime.