Redes neuronales recurrentes para la predicción de energía eólica y fotovoltaica

Abstract

Renewable energies are the order of the day, due to their competitiveness and necessity compared to fossil fuels, which cause climate change. Green energies ended 2020 with a 44% share of total generation, and today they are around 51% according to REE (Red Eléctrica de España). The increased is installed capacity, mainly wind and solar energy, meant that almost half of the total energy generated in Spain in 2020 was renewable, with 110,556 GWh, 12.9% more than previous year. It is on these two energies that this TFM focuses, with the aim of predicting the behaviour of short-term production over time. This aspet is of vital importance for electric companies, which rely on this type os estimates for the control of each of their plants or the management of the electricity market. For the development of this project, we have chosen to study the following recurrent neural networks: SimpleRNN from keras, Gated recurrent Unit and Long Short-Term Memory. In addition, the results will be compar3ed with the Ridge Regression and Multilayer Perceptron Regression regression models. Itis important that these models are mainly fed by the time series of production in each of the different experiments. In this proyect we will first carry out an analysis of the behaviour of the models in peninsular Spain for the two types of energy and then we will carry out two experiments located in more specific areas: in the wind field, the Sotavento win d farm will be studied and n the photovoltaic field, the behavior on the island of Mallorca will be analyzed. For the latter two problems, in addition to the corresponding production series, a time series of meteorological variables provided by the ECMWF (European Centre for Medium-Range Weather Forecasts) will be used.