Seasonal and intraday effects on spectral mismatch corrections for photovoltaic performance modelling in the United Kingdom

Abstract

Modelling photovoltaic (PV) performance is essential for improving system design and operation. Current models to account for the spectral influence on PV performance (spectral correction functions, SCFs) are typically developed and validated on annual or multi-year timescales. Through an empirical analysis of short-term (monthly and intraday) meteorological and PV performance data, this work shows that there is significant variation in the accuracy of different methods to characterise the prevailing spectral irradiance conditions that are adopted by published SCFs. Compared with the use of weather– and system-specific models, a one-size-fits-all approach to model selection may result in an order of magnitude increase in the model residual sum of squares (RSS). One of the reasons for these inaccuracies includes the fact that model performance depends on the prevailing weather conditions. A model that performs well under clearsky conditions can suffer from reduced accuracy in “dynamic sky” conditions, as characterised by fast-changing partial cloud cover. Four SCFs are studied in this paper, namely an air mass model, $f\left(A{M}_a\right)$, average photon energy model, $f\left(\phi \right)$, air mass and clearsky index model, $f(A{M}_a,k_c)$, and an average photon energy and spectral band depth model, $f(\phi ,\varepsilon )$. The two single-variable models (air mass spectral correction, $f\left(A{M}_a\right)$, and the average photon energy spectral correction, $f\left(\phi \right)$) are shown to be unreliable across the seasons, with reduced performance in summer and under dynamic sky conditions. Furthermore, they exhibit systematic time-of-day errors, even under clear skies, resulting in part from the non-bijective relationships between the spectrum and the independent variables ($A{M}_a$ and $\phi$). On the other hand, the multivariable approaches (air mass and clearsky index, $f(A{M}_a,k_c)$, and average photon energy and the depth of a water absorption band, $f(\phi ,\varepsilon )$) offer increased accuracy by mitigating the bijectivity issue. These improvements are reflected by substantial increases (decreases) in goodness of fit metrics such as $R_{adj}^2$ and RSS for the three PV devices studied in this paper. However, the choice of multivariable SCF is device specific. $f(\phi ,\varepsilon )$ is found to model the spectral effect on the cadmium telluride and amorphous silicon PV devices most effectively, while $f(A{M}_a,k_c)$ offers the best approach for the multicrystalline device.