Desenvolvimento de um modelo para a estimação da carga de radiação solar com base em variáveis climáticas

Abstract

Data related to solar radiation (SR) incidence in a given region play a crucial role in the design, modeling, and operation of solar energy conversion systems. Also, this type of information helps in future energy investment policies governments. It contributes to numerous other areas of study and several applications, such as creating thermal chimneys, thermal comfort analysis in buildings, crop growth models, and others. However, these data are not measured for all regions due to the lack of appropriate equipment for the measurement of SR in meteorological stations and the costs involved in the activity. Although different alternative methods for obtaining SR have been proposed and used in the literature, lately, most research has focused its efforts on exploring and using machine learning algorithms to address the problem. Following this line, the present work evaluates applying a linear regression with a regularization factor of type L2 combined with an expansion strategy of the input data to estimate the intensity of incident daily solar radiation. The models’ adjustment and evaluation are performed on eight different datasets containing climatic variables easily accessible by the weather stations. Each set refers to the other region distributed throughout Burkina Faso, a country located in Sub-Saharan Africa. For the selection of method parameters, the exhaustive search algorithm was applied to find the set of hyperparameters that reinforce the models’ predictive capabilities. The evaluation and comparison between the models are performed according to root mean square error, mean absolute error, coefficient of determination and variance accounted for. The results obtained validate the work’s strategy, showing the positive impact on the model generated by the inclusion of the regularization term. The experiments suggest a better performance compared to the values found in the literature produced by models with greater complexity, showing a reduction in the error values up to 50% for some stations.