Exploring artificial intelligence potential in solar energy production forecasting: Methodology based on modified PSO optimized attention augmented recurrent networks

The use of renewable power sources is vital for reducing the world's reliance on limited fossil fuels, reducing the impact on climate and mitigating the losses associated with power transmission. However, renewable sources such as solar power, often suffer from fluctuations in production due to their heavy reliance on weather conditions. This can have a significant impact on their reliability, as well as an impact on the power grid. Nevertheless, these issues could be mitigated by utilizing powerful and robust forecasting models, allowing for more efficient planning and fuller utilization of the produced power. This work explores the use of artificial intelligence (AI) in order to predict the yield of photovoltaic-generated energy. Different artificial neural network architectures are explored, including recurrent neural network (RNN), gated recurrent unit (GRU), and the long short-term memory (LSTM). Additionally, attention mechanism is integrated into the best-performing model to help further improve its performance. To ensure favorable outcomes, an adapted variant of the particle swarm optimization (PSO) is introduced to optimize hyper-parameter settings of each model. Simulations with real-world data showcased promising results while the rigorous statistical analysis confirmed that the observed improvements are statistically significant. The best-performing models were subjected to feature importance analysis to help future endeavors, as well as data collection efforts. The best performing models attained an impressive normalized mean square error (MSE) and coefficient of determination (R²) of 0.007240 and 0.894693, respectively, suggesting strong perspective for real world applications. Nonetheless, the introduction of attention mechanism did not provide further improvements to the best performing model. Lastly, this study confirmed that the modifications made to the baseline PSO strengthened the original approach, as it statistically significantly outperformed other metaheuristics.