Photovoltaic Power Output Prediction Using PCA–KNN: A Comparative Study with SVR and MLP on Historical Operational Data
Abstract
The transition toward renewable energy integration in power systems presents challenges in managing photovoltaic (PV) power output owing to variations in operating conditions. This study evaluated the performance of K-Nearest Neighbor (KNN), Support Vector Regression (SVR), and Multi-Layer Perceptron (MLP), with and without Principal Component Analysis (PCA), for PV power output prediction. The dataset comprised 420 observations recorded at five-minute intervals over seven days of measurements from a PV system. The data were chronologically divided into 80% training and 20% testing sets, while normalization and PCA were exclusively fitted to the training data to prevent information leakage. The model performance was evaluated using the Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and coefficient of determination (R²). The results show that KNN and PCA-KNN achieved the best and identical performance, with an MSE of 265.3385, RMSE of 16.2892, and R² of 0.8421. PCA did not improve KNN accuracy but maintained its predictive performance after the dimensional transformation. PCA-MLP showed a slight improvement over MLP, whereas SVR and PCA-SVR yielded lower performances. These findings indicate that the effect of PCA depends on both the dataset characteristics and the learning algorithm. Further studies with longer observation periods, more diverse datasets, and external validation are recommended to assess the generalizability of the models.