Comparison of deep learning techniques for processing satellite imagery

The use of photovoltaic (PV) systems has been rapidly increasing over the years due to policies aimed at reducing CO2 emissions. However, the discontinuous nature of solar power represents a challenge for the efficient integration of PV systems into the electricity grid. Therefore, the use of solar radiation prediction models proves to be highly valuable for the smooth integration and reliable operation of the electricity grid. In the literature, several studies have been conducted using Numerical Weather Prediction (NWP) to forecast future solar radiation. However, in recent years, deep learning-based models have become dominant. Most of the models in the literature use convolutional or recurrent neural networks, such as Long Short-Term Memory (LSTM), which have shown the best performance. In recent years, attention-based models like Transformers have outperformed LSTM in many tasks. Hence, this thesis investigates the use of such models for processing satellite imagery. In particular, this thesis has the purpose of predicting Global Horizontal Irradiance (GHI) using PerceiverIO, an attention-based model introduced by DeepMind in 2021. The model takes multi-resolution satellite images from the Meteosat family of geostationary satellites managed by EUMETSAT as input. Specifically, for this work, we evaluate the model using a dataset from Carpentras, France. Additionally, through a prediction model, we calculate the values of GHI in clear sky conditions and use them as queries for the model. The forecasting horizon is set at 6 hours, and the evaluation metrics are compared for each time horizon with benchmark models, revealing improved accuracy and either outperforming or matching other state-of-the-art solutions.