A Deep Learning technique to forecast solar radiation

Over the years, many countries have resorted to renewable energy sources (RESs), among which solar energy appears to be the most promising. Conse- quently, the photovoltaic technology (PV) is a steadily growing energy sector worldwide. However, the performance of PV systems depends heavily on the weather conditions. The chaotic nature of meteorological conditions hinders effective power management of any PV system. The scientific literature has ex- plored and continues to explore methods to predict the PV power production. This work investigates a Deep Learning (DL) approach for short-term forecast- ing of the solar radiation as Global Horizontal Irradiance (GHI), one of the most important parameters that affect the PV power production. The DL technique consists of using two models in cascade: a ”simple video prediction” (SimVP) model followed by a Multilayer Perceptron (MLP) model. The SimVP model generates a sequence of predicted images, and for each predicted image, the MLP model estimates the GHI corresponding to the center of it. The predic- tions extend up to 3-hours-ahead with a temporal resolution of 15 minutes. The dataset consists of multi-channel images from the geostationary satellites of the Meteosat series, along with GHI values under clear sky conditions computed using the Ineichen model. The data covers two years of samples (2015-2016) and include 9 sites distributed across Europe. The training and validation sets are composed of the data of 8 sites, while the test set is composed of the data of the remaining site. The work explores several sets of hyperparameters for both models, taken separately and/or together. Ultimately, the performance of the best SimVP–MLP combination is comparable to that of state-of-the-art mod- els. The results show that the best SimVP-MLP combination model achieves a normalized Mean Absolute Error (nMAE) of 34.75% and a normalized Root Mean Square Error (nRMSE) of 64.18% for 3 hours ahead predictions.