Modelagem da temperatura do ar em transecto urbano utilizando LSTM

• Ano de defesa: 2020
• Autor(a) principal: Guarienti, César Eduardo
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Mato Grosso; Brasil; Instituto de Física (IF); UFMT CUC - Cuiabá; Programa de Pós-Graduação em Física Ambiental
• Palavras-chave em Português: Dados ambientais; Predição de temperatura; Redes neurais recorrentes; Deep learning; Séries espaço-temporais; CNPQ::CIENCIAS EXATAS E DA TERRA::FISICA; Environmental data; Temperature prediction; Recurrent neural networks; Space-time series
• Link de acesso: http://ri.ufmt.br/handle/1/3409

In cities there is a process of storing heat during the day and returning that heat to the atmosphere at night. Thus, it is necessary to understand how the temperature changes in time and space to provide information for decision making. Therefore, this research aimed to predict the temperature of the use of the Recurrent Neural Network, type LSTM (Long-Short Term Memory). Used as input data for the Neural Network, environmental and soil cover variables, collected in the night mobile project in the city of Cuiabá-MT, in the year 2011 and 2016, separated by seasons: winter, autumn, spring and summer. The preparation of data for use of the LSTM was applied to a distance calculation that considers a temporal and spatial issue. To predict air temperature, various combinations of LSTM parameters and 2011 data were tested and used for the 2016 forecast, and with these models created. To validate the tests and compare real data with data, the Coe cient of Determination (R²) and Mean Square Quadratic Error (RMSE) were used as statistics. The results obtained with an R² greater than 0.9, in most of the tests, point to the feasibility of using LSTM for forecasting time-space series. It is possible to conclude that the LSTM model was able to predict the temperature in 2016 and similar for the year 2021.