SPLTRAK Abstract Submission
Spatiotemporal Modeling of  Real World Backsheets Data: Hierarchical (Multilevel) Generalized Additive Models
Raymond Wieser1, Kunal Rath1, Stephanie Moffitt2, Xiaohong Gu2, Evan Boucher3, Liang Ji3, Ruben Zabalza3, Mike Kempe4, Silvana Ayala4, Roger French1, Gregory O'brien5, Adam Hauser5, Kaushik Choudhury6, Jared Tracy6, Ken Boyce3, Laura Bruckman1
1Case Western Reserve University, cleveland, OH, United States
/2National Institute of Standards and Technology, Gaithersburg, MD, United States
/3Underwriter Laboratories, Northbrook, IL, United States
/4National Renewable Energy Laboratory, Golden, CO, United States
/5Arkema, King of Prussia, PA, United States
/6DuPont, Wilmington, DE, United States

Assessing photovoltaic module backsheet durability is critical to increasing module lifetime. 
Lab based accelerating testing has recently failed to predict large scale failures of widely adopted polymeric materials. 
Field surveyed data is critical to assess the performance of component lifetime. 
Using a documented field survey protocol, 18 field surveys were conducted. 
Each measurement is encoded with its spatial location in respect to the other modules. 
A variety of computation methods are applied to the field survey data. 
LOESS and Inference by Eye are quick identification methods for observing spatially dependent effects. 
Modules located at the edges of racking configurations were shown to to have statistically significant increases of degradation predictors. 
By combining field survey data on degradation predictors with real time satellite weather data, data-driven predictive models of backsheet degradation were trained.
Generalized Additive models have been used to model backsheet degradation, but struggled to capture the effect of individual stressors. 
To increase the usability of this framework, hierarchical models have been adapted to model the material specific contributions of degradation. 
The models have been validated using cross-validation. 
Hierarchical Models allow for group specific level smoothers which allow deviations from a globalized trend based on individualized penalties. 
It was shown that the hierarchical model was more flexible than the original GAM framework. 
Additionally, by allowing more flexibility in the model fit, individual coefficients for material specific degradation can be calculated. 
It was shown that material choice made a significant difference in the observed degradation of photovoltaic backsheets.