Optimization and comparative analysis of renewable energy generation forecasting model parameters with traditional and modern machine learning methods based on attribute selection

In this thesis, solar radiation forecasting of Adana, Osmaniye and Mersin provinces located in Çukurova in the east of the Mediterranean region, which is the most potential region of Turkey in terms of solar radiation, has been realized. In the study, solar radiation, wind speed, pressure, humidity, precipitation, all sky insolation index, temperature, maximum temperature, minimum temperature, dew point parameters covering 10 years between 01/01/2011-01/01/2021 from NASA-Power platform were used. First, attribute selection methods were applied on Adana province data to obtain attributes that will be given as input parameters to traditional and modern machine learning. Pearson Correlation Coefficient, Mutual Information, Sequential Forward Floating Selection, Lasso Regression were used in feature selection methods. The input parameters optimized in Adana province data were also used in the prediction process for Osmaniye and Mersin provinces. The forecasting models using the input parameters performed the forecasting process and the results obtained were analyzed comparatively. RMSE, MAE and R2 were used as comparison metrics. Machine learning methods such as Support Vector Regression, Extreme Gradient Boosting, Light Gradient Boosting, Artificial Neural Networks and Long Short Term Memory algorithms were used in this study.