Enhanced solar photovoltaic power prediction using diverse machine learning algorithms with hyperparameter optimization

Solar photovoltaic power generation accurate prediction is crucial for optimizing the efficiency and reliability of solar power plants. This research work focuses on predicting photovoltaic power using various machine learning algorithms, including ensemble of regression trees, support vector machine, Gaussian process regression, and artificial neural networks. Performance of these algorithms is further improved through hyperparameter optimization using Bayesian optimization and random search optimizers. Hourly data with a 30-min temporal resolution for an entire year is collected from a 10 MW Masdar solar photovoltaic project based in the United Arab Emirates. Photovoltaic historical power curve is generated using the System Advisor Model software, and to ensure data consistency, the collected dataset is normalized, with the interrelationships among variables computed using the Pearson relation coefficient. The results substantiate that Gaussian process regression demonstrates the best performance (lowest prediction errors) in terms of computing predicted solar photovoltaic generation power, followed by artificial neural networks, ensemble of regression trees, and the support vector machine across both optimizers. Concerning hyperparameter optimization, Bayesian optimization -based model outperformed support vector machine, Gaussian process regression, and artificial neural networks algorithms, except for the ensemble of regression trees. The proposed work contributes to the advancement of solar photovoltaic power prediction by combining the power of machine learning algorithms with hyperparameter optimization techniques. Additionally, the results emphasize the importance of hyperparameter optimization in enhancing machine learning model performance, providing valuable insights into adaptability and accuracy across varying seasonal conditions.