Treffer: Long-term comparative analysis of machine learning models: A deep dive into applications of artificial intelligence for enhancing photovoltaic performance prediction.

[Display omitted] • Achieving high prediction accuracy with over 95% R-squared scores achieved. • Comparing detailed prediction performance of various Machine Learning models. • K-Nearest Neighbors and Decision Trees showed optimal accuracy. • Support Vector Regression models were sensitive to hyperparameters. • Eight Ensemble learning methods performed comparably except Adaboost. • Inclusion of comprehensive one-year environmental data and output electrical data. This study tackles a key research gap by applying comparative analysis on several well-known classic machine learning models to predict photovoltaic performance under variable environmental conditions and dust levels. It uses a year-long dataset that includes environmental factors such as irradiance, ambient and module temperatures, wind speed, humidity, precipitation, and dust accumulation, along with electrical outputs like voltage, current, and power. Data were collected from two PV systems, one regularly cleaned and one left naturally soiled, at the same operating conditions. Six supervised machine learning models were trained: K-Nearest Neighbors (KNN), Cross-Decomposition Regression, Decision Trees, Support Vector Regression (SVR), Multi-Layer Perceptron (MLP), and Ensemble Learning. The models were trained to predict differences in voltage, current, and power between the two systems, achieving R-squared scores over 85% in most cases. KNN and Decision Trees performed best with fewer neighbors and shallow tree depths though they might suffer from computational costs in terms of large input data. SVR and MLP showed high sensitivity to hyperparameters such as activation functions and iteration limits, respectively. Although most ensemble methods performed well, Adaboost's accuracy declined as the number of estimators increased. The outcomes of this study offer practical insights for real-world applications, including predictive maintenance and energy optimization. The predictive models developed herein can support enhanced PV system design, the prediction of soiling effects using environmental data, optimized dust removal methods and cleaning schedules, and data-driven energy policy planning. [ABSTRACT FROM AUTHOR]