Desempenho de métodos de aprendizado de máquina para estimativa de temperatura de coletores solares com nanofluidos

Abstract

Research on renewable energy systems has increased significantly, because of the need to find new sources of energy that will complement or even replace fossil fuels, which are considered non-renewable and cause environmental impacts. Compared to other renewable energy sources, solar energy has been attracting interest because it is considered easily accessible and abundant. In this sense, the objective of this work is to analyze the performance of machine learning models for temperature estimation in solar collectors. The experiments were carried out at the Solar Energy and Natural Gas Laboratory (LESGN) evaluating the behavior of the temperature profiles of four concentrations of titanium dioxide nanoparticles (12.5 ppm; 25 ppm; 50 ppm; 75 ppm) in the period between October and December 2018, from 6:00 am to 6:00 pm, and silver nanoparticles (0.8125 ppm; 1.625 ppm; 3.25 ppm; 6.5 ppm) between June and October 2019, from 5:30 am to 9:30 pm, dispersed in deionized water in a solar radiation direct absorption system. The correlation between the predictors is based on the physical model variables that are shown to be most relevant. The input data for the numerical modeling are: temperature at the latest time step, ambient temperature, mean nanofluid temperature, direct solar irradiance, global solar irradiance, wind speed, and solar zenith angle, all of which were used to estimate the final temperature and temperature variation of the titanium dioxide nanofluid dispersed in deionized water using seven numerical models: Ridge Regression, LASSO, ElasticNet, kNN, XGBoost, MARS, and Ensemble Aggregate Models. Among them, two models were used as Ensemble (combined) methods: Ridge and MARS. The R programming language was applied for statistical analysis of error metrics and to rank the importance percentage of the predictors. The data was separated into a training and test set, using a 10-fold cross-validation for the training set, and the RMSE was used to choose the best performing model. At the concentrations of titanium dioxide nanoparticles, kNN showed the best RMSE performance for estimating temperature variation. The lowest RMSE was found at a concentration of 75 ppm, with a value of 0.146 °C. Thus, the computational model showed an efficiency 68.56% higher than the physical model. The best RMSE performance in the studies of silver nanoparticles was found for the concentration of 1.625 ppm on the estimation of temperature variation with the Ensemble MARS model, of about 0.136 ºC.