Multimodal deep learning for solar radiation forecasting

The efficient utilization of solar radiation is imperative for sustainable energy generation and the reduction of greenhouse gas emissions. Precise forecasting of solar radiation is of fundamental importance in optimizing photovoltaic energy systems and other renewable applications. However, its stochastic nature and variability make predicting energy generation challenging. This research presents a robust methodology for developing multimodal solar radiation prediction models using two types of neural networks based on real data from the CESA-I installation. These multimodal neural networks use both images and weather data as inputs. First, it presents a preprocessing of meteorological data and images to facilitate neural networks in extracting features. Then, the inclusion of state-of-the-art hybrid neural networks such as Convolutional Neural Networks (CNN) and Long Short Term Memory cells (LSTM) and some results are introduced. These neural network models are validated by statistical indicators and compared with other models that used only meteorological features or images. The results demonstrate the strong potential of these techniques for solar radiation prediction, achieving an immediate forecast between 10 min and 1 h with an error as low as 2.5 % in the best case. This represents a significant step forward in the pursuit of sustainable and energy-efficient solutions. • Solar radiation predictions improve energy management, achieving 3 % NRMSE in 10 min. • Sky images capture the properties of clouds and therefore help in better prediction. • Deep learning hybrid models outperform unimodal ones, reaching 10 % NRMSE at 1 h.