Stacked LSTM Sequence-to-Sequence Autoencoder with Feature Selection for Daily Solar Radiation Prediction: A Review and New Modeling Results

We review the latest modeling techniques and propose new hybrid SAELSTM framework based on Deep Learning (DL) to construct prediction intervals for daily Global Solar Radiation (GSR) using Manta Ray Foraging Optimization (MRFO) feature selection select model parameters. Features are employed as potential inputs Long Short-Term Memory a seq2seq autoencoder system in final GSR prediction. Six solar energy farms Queensland, Australia considered evaluate method with predictors from Climate Models ground-based observation. Comparisons carried out among DL models (i.e., Neural Network) conventional Machine algorithms Gradient Boosting Regression, Random Forest Extremely Randomized Trees, Adaptive Regression). The hyperparameters deduced grid search, simulations demonstrate that is accurate compared other well persistence methods. obtains quality high coverage probability low interval errors. modelling results utilising an deep learning show our approach acceptable predict radiation, therefore useful monitoring systems capture stochastic variations power generation due cloud cover, aerosols, ozone changes, atmospheric attenuation factors.