Development of empirical solar radiation models with genetic algorithm and extended validation procedure

Abstract

ABSTRACT The strong dependence of the accuracy of empirical solar radiation models on natural and climatic conditions causes their diversity and ambiguity of choice. The paper proposes an approach to choosing two most accurate models from a variety of empirical models designed to estimate global and diffuse solar radiation over the entire territory based on data from several locations. The approach focuses on the improvement of the accuracy and adaptability of models to a variety of natural and climatic conditions. To improve the adaptability of models, their coefficients are determined using a genetic algorithm. In contrast to the least-squares method, the objective function is aimed at minimizing the maximum mean absolute error (MAE) in estimating solar radiation at locations. A new procedure for extended validation of empirical models is proposed to improve their accuracy. In addition to the standard validation based on estimates of global and diffuse radiation, the accuracy of direct solar radiation estimates obtained as the difference between models’ estimates is checked. A numerical example is given for the Baikal natural territory. Based on the proposed approach, 18 empirical models are calibrated and the 2 most accurate models are selected. The maximum MAE in estimates of global, diffuse, and direct radiation for nine locations of the territory was 0.51, 0.57, and 0.6 MJ/m2day, respectively. The least-squares method of calibration yielded higher estimates of the maximum MAE: 0.64, 0.64, 0.88 MJ/m2day. The estimates of the developed models are more accurate compared to those provided by ERA5 and NASA POWER databases.