Real-world challenges and opportunities in degradation rate analysis for commercial PV systems

Determining accurate annual system degradation rates from historical meteorological and production data is essential for improving the overall financial viability of photovoltaic (PV) systems. Such efforts can be quite challenging, however, as nearly every step in degradation analysis has the potential to contribute error to the resultant rates. The presence of scatter and "evental" noise (e.g., due to snow and inverter downtime events) in non-laboratory-grade data sets collected from commercial PV systems makes degradation analytics all the more challenging. A MATLAB® tool has been developed as part of an ongoing investigation into the sensitivities of annual system degradation rates to data qualification and filtration. In this study, a two year set of continuous meteorological and production data affected by each of the abovementioned issues is analyzed using this tool. An interpolative inverter efficiency function is used to convert AC to DC power. A PVUSA-related multiple regression-based power model employing module temperature data is used to derive time series data sets of DC power values normalized to an arbitrary set of conditions. A limited number of qualification and filtration configurations are simulated, including an experimental method. Linear least-squares fit and classical decomposition are used to extract degradation trends from the normalized DC power time series data sets derived for each simulated configuration. Preliminary results show how both the presence of irregular scatter and evental noise in continuous, raw data sets, as well as the type of qualification and filtration criteria employed, can influence the outputs of degradation analyses. Removal of this noisy data, however, can unfortunately result in missing weeks or, in extreme cases, months of data, thus compromising the ability to employ more advanced time series algorithms (e.g., classical decomposition) to normalized power time series data sets to yield degradation trends- extracted from their seasonality and error counterparts. A tested experimental method sought to address the issue of missing data, but the method and results remain preliminary and premature. The chosen system will be continuously investigated to see if and how additional input data affects the resultant degradation rates, as well as any preliminary observations made herein on the sensitivity of those rates to data qualification and filtration.