Hybrid machine learning approach in forecasting renewable energy production: Extra trees, catBoost and LightGBM based stacking model

This study addresses the development and comparison of machine learning models to forecast renewable energy production in Turkey. In the study, data from 2009 to 2023 were analyzed to examine the relationships between different renewable energy sources and economic indicators. Specifically, the production of hydroelectric, geothermal, wind, and solar energy was evaluated alongside economic variables such as interest rates, exchange rates, oil prices, gross domestic product (GDP), and the industrial production index. Correlation analysis revealed strong positive correlations between total renewable energy production and other renewable energy sources. During the model development phase of the study, various machine learning algorithms were tested. These algorithms included Random Forest, Gradient Boosting, Extra Trees, LightGBM, XGBoost, and CatBoost. Based on performance evaluations, the three algorithms providing the highest prediction accuracy were identified, and a hybrid model was constructed using Extra Trees, CatBoost, and LightGBM. In this study, a novel hybrid model is introduced using a stacking regression framework with a meta-learning approach. According to the study's findings, the developed hybrid model demonstrated the highest forecasting performance. The overall success rate of the model was calculated as R2 = 82.55 %. When evaluating the error rates of the hybrid model, the mean absolute error (MAE) was found to be 989.82, while the mean absolute percentage error (MAPE) was 12.72 %. These results indicate that the hybrid model has lower error rates compared to individual machine learning algorithms. Overall, this study demonstrates that the hybrid modeling approach yields more successful results in forecasting renewable energy production than predictions made using a single algorithm.