Previsão de séries temporais de evapotranspiração de referência com redes neurais convolucionais

• Ano de defesa: 2019
• Autor(a) principal: Patrícia de Oliveira e Lucas
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais
• Palavras-chave em Português: Engenharia elétrica; Previsão; Redes neurais convolucionais; Séries temporais; Eevapotranspiração de referência; Ensemble learning
• Link de acesso: https://hdl.handle.net/1843/32870

Population growth and climate change are causing the agricultural sector to seek more accurate and efficient approaches to ensure an adequate and regular supply of food to society with less water consumption. Agriculture 4.0 comes at this context of resource scarcity as a management that seeks through technologies such as Big Data, Internet of Things (IoT), Artificial Intelligence and Robotics to provide plants and animals with exactly what they need and when they need it, increasing productivity and reducing environmental impacts. Irrigation management, an essential practice for the development of sustainable agriculture, seeks, through reference evapotranspiration forecasting, to know in advance the water needs of crops to plan and manage water resources. This dissertation is inserted in this context, aiming to investigate the use of deep learning models, especially convolutional neural networks, in the prediction of reference evapotranspiration time series (ETo). For this, three convolutional neural networks with different structures were implemented to predict a daily time series of ETo. To optimize the hyperparameters of these models a genetic algorithm was used, it sought to balance two objectives, precision and parsimony. The CNN models were validated by comparing them with known time series forecasting models such as ARIMA, WFTS and LSTM. For comparison purposes, ensemble learning with the CNN models was also implemented. The results showed that CNN models are feasible for ETo time series forecasting and that ensemble models improve predictions in terms of variance, accuracy, and computational cost over individual models.