Optimizing solar photovoltaic power forecasting via multi-architecture machine learning framework with multiple hyperparameter optimization techniques

The integration of solar photovoltaic systems into power grids has become crucial for sustainable energy development, particularly in extreme climate conditions where accurate power forecasting remains challenging. This study presents a novel multi-architecture machine learning framework with various optimizers to significantly advance PV power forecasting accuracy under the unique environmental challenges of the United Arab Emirates. This work investigates a range of machine learning techniques, including artificial neural networks with various architectures (narrow, medium, wide, bilayered, and trilayered), Gaussian process regression with diverse kernels (rational quadratic, squared exponential, Matern 5/2, and exponential), linear regression models with different configurations (standard, interactions, robust, and stepwise), regression trees of varying depths (fine, medium, and coarse), ensemble of regression trees (boosted and bagged), and support vector machines with different kernel functions (linear, quadratic, cubic, fine Gaussian, medium Gaussian, and coarse Gaussian) to forecast photovoltaic power. The system advisor model simulated a year-long photovoltaic power generation dataset, which was then normalized and analyzed using Pearson's correlation coefficient to identify key predictors. After comparing multiple configurations of each algorithm, the best-performing model was further refined using hyperparameter optimization techniques, including tree-structured Parzen Estimator, Bayesian optimization, random search with successive halving, and grey wolf optimizer. The performance of all models was evaluated based on root mean square error, mean absolute error, prediction speed, training time, and model size. Results corroborate that wide artificial neural network configuration achieved the highest accuracy among ML models (root mean square error: 6.4 kW), while TPE outperformed other optimization techniques (root mean square error: 5.2 kW). • Various machine learning configurations and hyperparameter optimization are evaluated for PV power forecasting. • Wide ANN outperforms all machine learning models and ANN configurations. • Four hyperparameter optimization techniques evaluated to optimize best-performing model. • Tree-structured Parzen estimator shows superior optimization of ANN parameters. • The analysis of trade-offs between model complexity, accuracy and computational efficiency is provided.