2026-05-04T20:58:19.924Z http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai

ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai:2:11663 2025-10-28T14:12:11Z 2

Attention augmented recurrent architectures for solar energy production forecasting 2025 http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai/record/2/11663 https://www.sciencedirect.com/science/article/abs/pii/S1568494625014322 V. Markovic L. Jovanovic M. Zivkovic B. Radomirovic A. Njegus M. Abdel-Salam V. Simic N. Bacanin The limited availability and non-sustainability of fossil fuels have led to the increasing interest in renewable energy alternatives. Significant obstacles must be addressed to fully integrate renewable energy into the existing power distribution grids. While reliability is a key factor in ensuring sustainable energy generation, solar power plants heavily depend on weather conditions which pose a challenge to maintain consistent and uninterrupted output without incurring substantial energy storage costs. As a result, accurate prediction of photovoltaic power production is crucial for efficient grid control and energy market operations. Traditional forecasting methods often struggle with nonlinear dependencies, while deep learning approaches are highly sensitive to hyperparameter tuning. This study proposes the application of metaheuristics optimization techniques to improve different lightweight recurrent neural network models and also considers attention mechanisms to forecast photovoltaic power generation. Additionally, an adapted metaheuristics optimizer is introduced to effectively overcome the obstacles of hyperparameters tuning. Extensive simulations are conducted using real-world dataset. The best-produced model in the simulations, which combines the gated recurrent unit with an attention mechanism, obtained a promising mean squared error score of 0.007713, indicating a significant perspective for further use in this area, with potential for deployment in resource-constrained environments like embedded and TinyML platforms. article 10.1016/j.asoc.2025.114119 114119 1568-4946 Applied Soft Computing