Estimating daily global solar radiation using deep learning

Abstract

• VAE, SAE-LSTM, SAE-GRU, and Transformer-CNN models were employed to forecast future solar radiation. • The proposed models were evaluated against benchmark approaches, including LSTM, GRU, CNN, and Transformer. • SAE-LSTM, SAE-GRU, and Transformer-CNN demonstrated enhanced predictive performance during both the training and testing phases. • The use of complex and hybrid architectures contributed to significant improvements in forecasting accuracy. Accurate prediction of daily global solar radiation (H) is essential for optimising the operational efficiency, exploitation and distribution of solar energy systems. This study investigates the potential of four Deep Learning (DL) models for five-day ahead prediction of H in Tangier, Morocco. The proposed models include a Variational Autoencoder (VAE) and three novel hybrid architectures: SAE-LSTM (Stacked AutoEncoder with Long Short-Term Memory), SAE-GRU (Stacked AutoEncoder with Gated Recurrent Units), and Transformer-CNN (a combination of Transformer and Convolutional Neural Network). These DL configurations have not been previously explored in the context of solar radiation forecasting. To evaluate their performance, they were benchmarked against four standalone DL models: LSTM, GRU, Transformer, and CNN. Among all models, the Transformer-CNN achieved the highest accuracy during the testing phase (R²=0.8530, MAE=0.5054 kWh/m²/day, RMSE=0.7623 kWh/m²/day), followed by SAE-LSTM (R²=0.8457, MAE=0.5599 kWh/m²/day, RMSE=0.7811 kWh/m²/day) and SAE-GRU (R²=0.8451, MAE=0.5701 kWh/m²/day, RMSE=0.7826 kWh/m²/day). These findings confirm that hybrid models combining SAE or CNN layers with temporal networks can significantly improve solar radiation forecasting accuracy.