Solar irradiation forecasting by the application of five machine learning algorithms

• Autor(a) principal: Marinho, Felipe Pinto
• Data de Publicação: 2019
• Outros Autores: Brasil, Juliana Silva, Amorim Neto, Juarez Pompeu de, Rocha, Paulo Alexandre Costa, Silva, Maria Eugênia Vieira da, Lima, Ricardo José Pontes
• Tipo de documento: Artigo de conferência
• Idioma: por
• Título da fonte: Repositório Institucional da Universidade Federal do Ceará (UFC)
• Texto Completo: http://www.repositorio.ufc.br/handle/riufc/61784
• topic: Solar irradiation forecast; Machine learning; Global irradiation; Minimal learning machine; Renewable energy

In this work, the forecast of global solar irradiation for a one-day ahead forecast horizon was carried out using some machine learning models, namely: Minimal Learning Machine, Support Vector Machine, Random Forests, K- Nearest Neighbors and a regression tree with the application of a Bagging procedure. The Minimal Learning Machine algorithm is a relatively recent method based on the distance calculation between vectors and used for supervised learning purposes in both classification and regression problems. In addition, we used a data set with the presence of attributes (predictors) formed by exogenous variables (insolation, air temperature, precipitation, etc.), endogenous variables (solar irradiation historical data) and temporal variables (year, month and day of measurement) totalizing 44 attributes and 3254 observations. The root mean squared error and forecast skill obtained by applying the Minimal Learning Machine in the validation set were respectively 40.882 W/m² and 7.637 %, and the arithmetic mean of the root mean squared error in conjunction with the arithmetic mean of the forecast skill obtained by the use of the other models for the same validation set were 40.752 W/m² and 7.93 %. In this way, it can be drawn by the evaluation of the results that the Minimal Learning Machine presents a performance comparable to the classic machine learning methods. Furthermore, it presents the advantage in the training stage of using only a single adjustment parameter.