A comparative study of several machine learning based non-linear regression methods in estimating solar radiation: Case studies of the USA and Turkey regions

Abstract

In this study, the potential of six different machine learning models, gradient boosting tree (GBT), multilayer perceptron neural network (MLPNN), two types of adaptive neuro-fuzzy inference systems (ANFIS) based on fuzzy c-means clustering (ANFIS-FCM) and subtractive clustering (ANFIS-SC), multivariate adaptive regression spline (MARS), and classification and regression tree (CART) were used for forecasting solar radiation from two stations of two different locations, Turkey and USA. Wind speed, maximum air temperature, minimum air temperature and relative humidity were used as inputs to the developed models. For accurate evaluation of performance of models, four statistical indicators, root mean squared error (RMSE), coefficient of correlation (R), mean absolute error (MAE) and Nash–Sutcliffe efficiency coefficient (NS) were employed to evaluate accuracy of the developed models. Comparison of results showed that the GBT model performed better than the MLPNN, ANFIS, MARS, and CART in modeling solar radiation. The average RMSE of MLPNN, ANFIS-FCM, ANFIS-SC, MARS and CART models was decreased by 0.26%, 1.5%, 0.51%, 2.5%, and 19.34% using GBT model at Fairfield Station, 4%, 1.37%, 0.24%, 4.12%, and 24.4% at Monmouth Station, 11.99%, 48.7%, 41.6%, 8.23%, and 33.41% at Antalya Station, 11%, 54.8%, 51.9%, 19.65%, and 37.1% at Mersin Station, respectively. The overall results indicated that the GBT model could be successfully applied in forecasting solar radiation by using climatic parameters as inputs.