Spatio-temporal wind speed forecasting with Bayesian uncertainty quantification

• Ano de defesa: 2023
• Autor(a) principal: Souza Neto, Airton Ferreira de
• Orientador(a): Mattos, César Lincoln Cavalcante
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Área do conhecimento CNPq: CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.ufc.br/handle/riufc/78733

The prediction of short and long-term wind time series has great utility for the industry, especially for wind energy generation, with various practical applications in the day-to-day operation of parks. The results are even more powerful and reliable when associated with uncertainty estimates, providing greater support for decision-making. In this work, a data-driven modeling approach based on deep neural networks is presented. The quantification of uncertainty associated with the predictive distribution can be done using a Bayesian learning approach. However, in the context of neural networks and deep learning, the conventional Bayesian approach is intractable and computationally expensive. On the other hand, there have been several recent advances in approximate Bayesian inference techniques in deep learning, particularly those that do not modify traditional training algorithms. This work proposes the use of deep neural networks for the spatio-temporal modeling of wind based on measurements collected from wind turbine data acquisition systems. It also includes predictions from widely used global climate forecasting models in the energy industry. The predictions made are accompanied by the quantification of uncertainty, extracted using approximate Bayesian inference techniques. The developed solution is evaluated using data collected from a wind farm in South of Brazil. Different combinations of models and approximations are compared based on the achieved metrics and graphs of uncertainty calibration. The conducted experiments indicate that the use of recurrent convolutional neural networks (ConvLSTM) with Deep Ensembles provides the best results for the predictive distribution, potentially assisting the operation of wind farms.