Correcting weather predictions of sea-surface wind with deep learning

Numerical models for weather prediction are frequently utilized to drive ocean simulations due to their provision of continuous data over both time and space. However, the outputs of these models exhibit inaccuracies in certain global regions, particularly in predicting the direction of sea-surface...

Descripción completa

Detalles Bibliográficos
Autor: Ramos González, Adrián
Tipo de recurso: tesis de maestría
Fecha de publicación:2024
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/424879
Acceso en línea:https://hdl.handle.net/2117/424879
Access Level:acceso abierto
Palabra clave:Weather forecasting
Deep learning (Machine learning)
convolutional neural networks
deep learning
U-Net
marine sciences
wind
PyTorch
Previsió del temps
Aprenentatge profund
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial
Descripción
Sumario:Numerical models for weather prediction are frequently utilized to drive ocean simulations due to their provision of continuous data over both time and space. However, the outputs of these models exhibit inaccuracies in certain global regions, particularly in predicting the direction of sea-surface winds. Even the latest versions of these models demonstrate such errors, which are attributed to natural processes that remain unmodeled. The aim of this master?s thesis is to predict these errors, referred to in this project as ERA5 model biases, using the most suitable Deep Learning architecture. This work builds on previous research that employed point-by-point prediction using Machine Learning and Fully Connected Neural Networks, which showed promising results. A Neural Network, specifically the U-Net for regression, is proposed, implemented, and evaluated to predict the model differences in sea surface wind direction.