Avaliação do modelo de previsão Prophet como ferramenta para preenchimento de falha de dados em séries climáticas

• Autor(a) principal: Lima Junior, Miguel Freitas de
• Data de Publicação: 2022
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: por
• Título da fonte: Biblioteca Digital de Monografias da UFMT
• Texto Completo: http://bdm.ufmt.br/handle/1/3094
• topic: CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; dados ausentes; preenchimento de dados; lacunas; Prophet; LSTM; INMET; séries temporais; Missing data; data filling; gaps; Prophet; LSTM; INMET; time series

In research with analysis and development of forecasting system, the main problem is the database inconsistency. Missing data is a very common problem, especially when climate databases is used. The causes can be from the different sources, like maintenance, sensor failures, discontinuity in meteorological stations and other. The Facebook's Prophet model was selected to be evaluated in climate data scenarios, despite of the Prophet model which has been standing out in time series forecasting in other contexts, especially commercial ones. For comparison with the Prophet, the Long Short Term-Memory (LSTM) model was chosen. It is a recurrent neural network quite widespread and used in different contexts, both commercial and research, especially in forecasting with historical series. This work manipulates climatic data, being them the humidity, precipitation and air temperature. This dataset was collected from the Banco de Dados Meteorológico de Ensino e Pesquisa (BDMEP). Five test cases were elaborated with different contexts of applied climatic data, using the models evaluated for the chosen seasons. Although is not possible to choose one of the studied models as the perfect one, after all the tests, we evaluated Prophet as the good choice, a valid model to deal with this type of problem, emphasizing its simple and scalable.. The compared model, LSTM, proved to be more efficient in some scenarios in filling gaps procedure, however, it is necessary to evaluate the balance between applicability and efficiency in choosing the model to be used.