Comparison of empirical daily surface incoming solar radiation models

Most environmental and agronomic models use climatic inputs such as temperature, solar radiation and rainfall. In Canada, few automatic weather stations can monitor incoming solar radiation. Hence, surface incoming solar radiation must be estimated from other meteorological data. The aim of the present research was to conduct a comparative study of three empirical models for the estimation of surface incoming solar radiation on a horizontal surface. The first two methodologies use the traditional and long-utilized linear approach based on latitude and daily temperature range. A more recent methodology uses neural networks (NN) to build a similar regression based on latitude and temperature range, but providing more flexibility in the formulation as well as a non-linear activation function. Global daily solar radiation data from 11 stations located in northeastern America was used to optimize and test the different models. Even though coefficients of the classical models were locally optimized using a standard unconstrained non-linear scheme, NNs markedly improved the estimation of incoming surface solar radiation.