Photovoltaic performance characterization: Optimization by regression basis with application to health management

Prognostics and health management (PHM) systems in the photovoltaic (PV) marketplace are garnering increased attention and scrutiny as deployment soars and continues to accelerate. The foundation for any PHM system is the accurate prediction and subsequent comparison to measured data. For PV this is most frequently based on the output AC power. All PHM methods rely on in-field sensor readings to construct predictor variables to forecast expected power output - regardless of the solution method (regression, artificial neural networks, Bayesian, etc.) - and these observed quantities even in the ideal case constitute an incomplete basis set. This naturally gives rise to variations in derived coefficients - regardless of the model used to solve for them. These `seasonal errors´ incurred in this application are simply a statement that some dynamic features are ignored or not fully accounted for in the predictor-basis set. In this work, these variations are examined in the standard multiple linear regression framework in the “local” limit where the global variations are not known. In this paper, we examine the deviations between the observed and regression-expected AC power at a daily level as a function of the chosen polynomial basis.