Energy production prediction with machine learning in photovoltaic systems

Although interest and investments in renewable resources such as solar and wind are increasing day by day, these facilities have the disadvantage of not being able to provide continuous energy to distribution networks. In addition, uncertainties on the consumption side can cause energy supply-demand imbalances. This situation brings forecasting studies to the fore. In this thesis study, energy production prediction was made with machine learning using two different data sets. For the first data set, a sample solar power generation plant was established and atmospheric data and instantaneous current values of the plant were collected online for eleven months using the Internet of Things Technology. The second dataset was obtained from a website that provides weather forecasts and atmospheric data worldwide. Five different datasets were created from the obtained data and trained, and energy production prediction was made using Linear Regression, Random Forest, XGBoost and LSTM machine learning algorithms. R2, RMSE and MAE performance metrics were used to evaluate the prediction performance. The best prediction result was obtained in the training with NIT dataset, RFR algorithm obtained 0.9298 and XGBoost algorithm obtained 0.9303 results from R2 metric evaluation. In the training with the second dataset, RF algorithm obtained 0.8869 and XGBoost algorithm obtained 0.8904 results from the R2 metric evaluation. These results show that machine learning algorithms can work with high accuracy in energy production forecasting and reliable forecasts can be obtained using both local atmospheric data, instantaneous current data and global weather forecast data.