Desenvolvimento de um modelo de predição de temperatura e umidade de uma estufa baseado em aprendizado de máquina

• Ano de defesa: 2020
• Autor(a) principal: Luísa Castello Branco de Sá
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ENG - DEPARTAMENTO DE ENGENHARIA MECÂNICA; Programa de Pós-Graduação em Engenharia Mecanica; UFMG
• Palavras-chave em Português: Estufa; Máquina de vetores de regressão; Perceptron multocamadas; LSTM; Engenharia mecânica; Estufa (Plantas); Redes Neurais (Computação)
• Link de acesso: http://hdl.handle.net/1843/51290

Greenhouses production depends on controlling its environmental conditions so that they are suitable for vegetable growth. In this context, greenhouses temperature and humidity forecast models have been developed to support the development of robust climate control strategies based on feed-forward control. Therefore, the present studied aimed to develop a climate forecast model for a greenhouse located in Belo Horizonte, which produces hydroponic lettuce. The historical temperature, humidity and luminosity data were acquired with three meteorological stations inside the greenhouse and one located in its external environment. Then, three machine learning techniques were evaluated for the climate forecast model, one of them based on support vector regression (SVR) and two based on artificial neural networks (MLP and LSTM). Random searches were performed to determine the best configurations for each model, and their average performance was assessed through a backtesting approach. The MLP architecture was chosen for the final model, due to its higher stability during the training stage and to lower RMSE and MAPE scores compared to the other techniques. The final model went through a new random search and backtesting evaluation, but now being trained with data from the external meteorological station, which is suitable for feed-forward control. The MLP model was able to predict temperature and humidity patterns inside the greenhouse satisfactorily, capturing in advance climate disturbance that would later affect its internal state. Further tuning of the model with temperature and humidity data from subsequent periods may increase its prediction ability, contributing to the development of the aimed feed-forward control strategy.