Several forecasting models for solar irradiance (the resource) (Mellit, 2008; Mellit & Pavan, 2010; Marquez & Coimbra, 2011; Elizondo et al., 1994; Mohandes et al., 1998; Hammer et al., 1999; Sfetsos & Coonick, 2000; Paoli et al., 2010; Lara-Fanego et al., 2011) and for solar-power output (Picault et al., 2010; Bacher et al., 2009; Chen et al., 2011; Chow et al., 2011; Martin et al., 2010) have been developed in the past few years.
Stochastic-learning methods based on artificial neural networks (ANNs), fuzzy logic (FL), and hybrids (GA/ANN, ANN-FL) are well suited to modeling the stochastic nature of the underlying physical processes that determine solar irradiance at the ground level (and thus the power output of PV installations) because of their robust nature and their ability to compensate for systematic errors and even more complex learnable deviations. Other regression methods often employed to describe complex nonlinear atmospheric phenomena include autoregressive moving averages (ARMA) as well as nonstationary variations such as autoregressive integrated moving averages (ARIMA) (Gordon, 2009).
In this section, we describe several forecasting methodologies that can be used to produce a forecast with no exogenous variables. The goal is to obtain a model in the generic form
y (t + Th) = f (y(t), y(t – Ді), ., y(t – иД?))
where the diacritic л is used to identify a forecast variable, and Th is the forecasting horizon. The time dependent variable y(t) is usually known only as a discrete variable or time series. For the univariate test, a forecasting model can be a function of any “current” or “past” values of the time series, but no other time series (e. g., temperature, relative humidity, cloud cover.) is used.