Realization of solar power plant electricity production estimation with machine learning algorithms

The aim of this study is to predict electricity production in a solar power plant using machine learning algorithms (XGBoost, LightGBM, CatBoost) and compare the performances of these models. In the study, hourly and daily electricity production data were analyzed and the prediction accuracy of each model was evaluated with metrics such as MAE, MSE and R². In addition, the predictions of the models at different production levels were supported with graphs on a daily basis. The findings showed that the XGBoost algorithm had the lowest average error (MAE: 2.438) in daily predictions, but limited success (R²: 0.362) in explaining the variance in the data set. The LightGBM algorithm had a wider coverage than the other models by explaining 56% of the data variance (R²: 0.561), but the error values (MAE: 2024.321 and MSE: 8,143,625.10) were found to be relatively higher. The CatBoost algorithm, on the other hand, exhibited a moderate performance with R²: 0.527 and MAE: 2123.28 values; while it provided consistent results at high production levels, it showed higher error rates at low production levels. In general, it was determined that the models were successful at high and medium production levels, but their performance was limited at low production levels and sudden fluctuations. As a result, while the machine learning algorithms used to estimate electricity production from solar power plants showed successful results, especially during high production periods, it was observed that error rates increased at low production levels.