Evaluating and analyzing the performance of PV power output forecasting using different models of machine-learning techniques considering prediction accuracy

Solar energy as a clean, renewable, and sustainable energy source has considerable potential to meet global energy needs. However, the intermittent and uncertain character of the solar energy source makes the power balance management a very challenging task. To overcome these shortcomings, providing accurate information about future energy production enables better planning, scheduling, and ensures effective strategies to meet energy demands. The present paper aims to assess the performance of PV power output forecasting in PV systems using various machine learning models, such as artificial neural networks (ANN), linear regression (LR), random forests (RF), and Support Vector Machines (SVM). These learning algorithms are trained on two different datasets with different time steps: in the first one, a historical weather forecast with a one hour time step, and in the second one, a dataset of on-site measurements with a 5-minute time step. To provide a reliable estimation of prediction accuracy for different learning algorithms, a k-fold cross-validation (CV) is applied. Through a comparison analysis, an assessment of the accuracy of these algorithms based on various metrics such as RMSE, MAE, and MRE is performed, providing a detailed evaluation of their performance. Results obtained from this study demonstrate that the random forest algorithm (RF) outperformed other algorithms in predicting PV output, achieving the smallest prediction error, where the best values for RMSE, MRE, MAE, and R² for the weather dataset were 0.856 W, 0.256%, 0.364 W, and 0.99999, respectively, while thevalues for RMSE, MRE, MAE, and R² for the on-site measurements dataset were 8.525 W, 11.163%, 3.922 W, and 0.99922, respectively.