Abstract
Improvements in solar radiation forecasting models have several positive effects on the solar energy industry. The existing improved models exhibit shortcomings, including loss of diversity, premature convergence, and convergence to local optima. These can undermine solar energy production. To address these shortcomings, this study proposed a novel hybrid adaptive variational mode decomposition with an enhanced particle swarm optimization for long short-term memory networks (AVMD-EPSO-LSTM). The AVMD strategy eliminates the high-frequency components from the solar radiation dataset using Harris Hawk Optimization (HHO) and VMD. The EPSO comprises a bi-diversity and a tournament elite selection method. The bi-diversity involves affinity propagation clustering (APC), which groups the swarm into subswarms, and a ring topology that searches for the best neighbor in each subswarm. The tournament elite selection finds the best solution from the archived best neighbors. The EPSO algorithm obtains the best LSTM’s hyperparameters that revamp the forecasting model's capability. Several performance metrics and models are used to evaluate and benchmark the proposed model, respectively. The experimental outcomes revealed that the AVMD-EPSO-LSTM model demonstrates exceptional performance in handling loss of diversity, premature convergence, and local optima. Furthermore, the Wilcoxon signed-rank test improvement analysis revealed the model's reliability. Consequently, the AVMD-EPSO-LSTM model can serve as a valuable operational tool for solar energy stakeholders, ensuring the sustainable development of renewable energy.
[Display omitted]