Most of the research on estimating Solar Global Irradiation (SGI) is based on the development of parametric models. However, the use of methods based on the use of statistics and machine-learning theories can provide a significant improvement in reducing the prediction errors. The chapter evaluates the performance of different Soft Computing (SC) methods, such as support vector regression and artificial neural networks-multilayer perceptron, in SGI modeling against classical parametric and lineal models. SC methods demonstrate a higher generalization capacity applied to SGI modeling than classic parametric models. As a result, SC models suppose an alternative to satellite-derived models to estimate SGI in near-to-present time in areas in which no pyranometers are installed nearby.
TopIntroduction
The rising interest of high quality solar irradiation estimate is explained by its straightforward applications in solar energy, agriculture and climate change. Regarding the first, accurate measurements of solar irradiation are always required to accurately estimate electrical energy production and leveled cost of electricity on a particular plant in a certain location. This can be done by installing meteorological stations with pyranometers but today, wide areas of the planet still lack of nearby pyranometers. Subsequently, solar irradiation is usually determined by two approaches: satellite-derived databases whose spatial resolution is always insufficient to provide a high level of certainty (Perez, Seals, Stewart, Zelenka, & Estrada-Cajigal, 1994), and information derived from other meteorological variables using classical parametric models (Bristow & Campbell, 1984) or soft computing (SC) methods (Rahimikhoob, 2010; Moreno, Gilabert, & Martinez, 2011).
Other application of estimated irradiation is the correction of not-available or spurious data in pyranometer registers, which in many cases suppose a big issue to estimate the annual irradiation.
Classical and SC techniques usually require registers of temperature and rainfall, which are commonly measured in many locations. They also need additional short records of solar irradiation to calibrate or train models. As a result, these techniques can provide near-to-real time irradiation estimates that are really useful to evaluate the performance of solar energy plants of previous days. This is a crucial service that can not be properly provided by the free satellite-derived databases (Antonanzas-Torres, Martínez-de-Pisón, Antonanzas-Torres, & Lamigueiro, 2013).
The authors have previously addressed an exploratory study of 22 classical models, also proposing a new methodology to optimize the classical models to on-site characteristics (Antonanzas-Torres, Sanz-Garcia, Martinez-de-Pison, & Perpinan-Lamigueiro, 2013). This assessment was validated in La Rioja region (Northern Spain). The model defined for this region showed a low average mean absolute error (MAE) in the range of 2.195 MJ/m2day with 41.65% of the residuals along 5 years (2007-2011) within the uncertainty of pyranometers (10%). However, the estimation of the irradiation is still a science with a high percentage of uncertainties and a large amount of irrelevant variables. SC solutions based on machine learning theories are actually revealing themselves as very useful tools for this task. In fact, different soft-computing techniques have already been considered in solar irradiation estimation, i.e. artificial neural networks (ANN) (Marquez & Coimbra, 2011;Rehman & Mohandes, 2008) k-nearest-neighbors (k-NN) (Pedro & Coimbra, 2012) and support vector regression (SVR) (J. L. Chen, Li, & Wu, 2013). The article shows the advantages of the most relevant SC methods in comparison with the previously evaluated classical models when estimating solar global irradiation (SGI). To this end, the same database of air temperature and rainfall in La Rioja region (Spain) is considered in the study.
Data
Measurements of temperatures and rainfall along 5 years (2007-2011) are freely obtained from the Service of Information for Agriculture of La Rioja (SIAR, 2013) and SOS Rioja (Rioja, 2013), collating the set of 17 meteorological stations shown in Figure 1. SGI is measured with First Class pyranometers (ISO9060) in SIAR network and with Second Class pyranometers in SOS Rioja, implying 5% and 10% of daily uncertainty, respectively. The resulting database is then divided into the training-validation (years 2007-2010) and testing (year 2011) databases in a ratio 80% to 20%.
Figure 1. Location of the meteorological stations selected from La Rioja. The color band shows the elevation (m). Blue and red circles represent SIAR and SOS Rioja stations respectively