Evaluation of Reflector Configurations for Enhanced Photovoltaic Power Prediction Using Machine Learning

The power output of photovoltaic (PV) systems fluctuates significantly due to variations in solar irradiance and weather conditions, creating challenges for reliable energy management and system optimization. Although performance can be improved through design features such as solar trackers, reflectors, and adjustable tilt mechanisms, these modifications also introduce added complexity and uncertainty in predicting PV output. Recent approaches using standalone artificial intelligence (AI) models, including machine learning (ML) and deep learning (DL) techniques, have shown promise for PV output forecasting. While these methods show considerable promise, they often struggle to generalize across diverse operating conditions and may fail to capture both linear and nonlinear dependencies in the data, resulting in inconsistent performance. In this study, a novel experimental setup that investigates the influence of tilt angle and innovative reflector configurations, using two and four reflectors, on PV module output is proposed. A total of 452 data points were collected under controlled variations of tilt (5°–90°) and reflector angles (90°–130°), capturing key parameters including solar radiation, surface temperature, voltage, current, and power. We performed a comparative analysis of six artificial intelligence (AI) models: three ensemble methods, including Averaging Ensemble, Boosting Ensemble, and Bagging Ensemble; and three standalone models, namely, Decision Tree, Support Vector Machine, and Multi-Layer Perceptron. The models were evaluated using standard statistical metrics: Coefficient of Determination (R²), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Results show that ensemble methods, especially the Averaging Ensemble, consistently outperformed standalone models, achieving an R² of 0.9974 (training) and 0.9822 (testing), with RMSE values of 0.0054 and 0.0066, respectively. It also reduced MAPE by 21.22% (training) and 20.45% (testing) compared to the second-best model, and over 60% compared to the least accurate model. These outcomes highlight the capability of ensemble methods to capture the complex nonlinear dynamics of PV output and emphasize their potential for real-time solar energy optimization in systems with adaptive reflector configurations.