Which models matter: Uncertainty and sensitivity analysis for photovoltaic power systems

Predicting power for a photovoltaic system from measured irradiance requires a sequence of models, e.g.: translation of measured irradiance to the plane-of-array; estimation of cell temperature; and calculation of module electrical output. Uncertainty in predicted power arises from the aggregate uncertainty in the various models used. But which models contribute significantly, or insignificantly, to this aggregate uncertainty? We report an uncertainty and sensitivity analysis that partially addresses this question. At each step in the modeling process, we consider commonly-used models and quantify uncertainty in the model and its parameters by analyzing model residuals obtained by comparing model predictions to available measurements. We are not yet able to quantify uncertainty in all relevant modeling steps. We develop stochastic process models for these residuals and use Monte Carlo sampling to propagate uncertainty from step to step. Propagating uncertainty through all modeling steps, we obtain a sample of values for PV system output representing the aggregate uncertainty in this quantity. We use rank correlations to identify the relative contribution to the aggregate uncertainty that can be attributed to each modeling step. For the steps and models we consider, we find that uncertainty in plane-of-array transposition and in effective irradiance models dominates the uncertainty in energy production; uncertainty in cell temperature and in module DC output is significantly less influential.