Optimized Solar Energy Forecasting for Sustainable Development Using Machine Learning, Deep Learning, and Chaotic Models

Abstract

This study applies four forecasting approaches—Ensemble Learning (EL), Deep Learning (DL), Machine Learning (ML), and Chaotic modeling— to predict energy production from the Konya Eregli solar power plant in Turkey. Using Python, it incorporates ambient temperature and solar cell temperature as exogenous variables alongside endogenous energy data. A year’s worth of 10-min interval data is trained, with two subsequent months forecasted by each model. The False Nearest Neighbors algorithm optimizes the embedding dimension for the chaotic analysis, and an optimized Echo State Network, achieving an R-squared above 0.97, is used for accurate forecasting. Additional models include Long-Short-Term Memory and Gated Recurrent Unit (DL), eXtreme Gradient Boosting and Random Forest (EL), and Extreme Learning Machine and Feed Forward Neural Network (ML). Each model is optimized using the Tree-structured Parzen Estimator, a Bayesian optimization approach. Evaluation metrics reveal all models performed well with the integration of endogenous and exogenous variables, with LSTM achieving the best results. This research advances solar energy forecasting, supporting Sustainable Development Goals (SDGs) related to affordable and clean energy, climate action, and sustainable communities through improved renewable energy management.