Optimizing Solar Energy Prediction: Insights from Comparative Analysis of Models, Optimizers, and Performance Evaluation

Abstract

This research aimed to enhance solar energy prediction accuracy through a comparison of models and optimization techniques. Long Short Term Memory and Recurrent Neural Network models were analyzed using nine features extracted from Typical Meteorological Year 3 format data obtained from National Solar Radiation Database, revealing distinctive patterns in seasonal variations and temporal availability of solar energy. Integration of Recurrent Neural Networks and Long Short Term Memory models significantly improved performance, addressing limitations in capturing long-term dependencies and uncertainties. Long Short Term Memory consistently outperformed other metrics by 2-3%, with larger hidden layer sizes enhancing predictive accuracy. The importance of selecting the appropriate optimizer considering accuracy, computational resources, and training time constraints was emphasized. Project-specific analysis underscored the significance of tailoring solar dimensions to location-specific yield data, informing cost optimization strategies for sustainable development.