Deep neural networks for spatio-temporal meteorological prediction

For centuries, Meteorological prediction has been a challenging task every civilization faced. With the development of recent technologies, it became easier to obtain accurate weather forecast. The objective of this thesis is to apply Deep Neural Networks algorithms on meteorological data to precisely predict the solar irradiance throughout the day. This is a crucial step in the optimization of renewable energy production and management. With the dawn of high performance computational hard-wares and the availability of large datasets, Recurrent Neural Networks and Convolutional Neural Networks have been built to predict the solar irradiance ratio one hour into the future. Thanks to a dataset composed by a dozen of variables recorded during five years, we built models to predict the desired value on five different stations on the La Reunion island. Compared to current state of the art algorithms, our models were able to predict accurate and truthful sequences, decreasing the error by 25% in the best case, with mae’s and rmse’s metrics reaching values as low as 0.17 and 0.19, respectively. Additionally, we experimented new types of algorithms, combining the data from multiple stations to build spatio-temporal models or combining the recurrent and the convolutional layers into one models. They gave promising results but necessitate a fine tuning of the parameters! During this work, it showed Deep Learning algorithms could improve the current solutions for meteorological predictions and that new data-driven solutions, for complex problems, will be discovered in the future as data become more and more available.