Previsão de dados de níveis de água subterrânea utilizando modelos baseados em aprendizado de máquina

• Autor(a) principal: Luiz, Thiago Boeno Patricio
• Data de Publicação: 2022
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Manancial - Repositório Digital da UFSM
• dARK ID: ark:/26339/001300000n5g6
• Texto Completo: http://repositorio.ufsm.br/handle/1/24211
• topic: Hidrologia subterrânea; Níveis de água subterrânea; Lençol freático; Séries temporais; Aprendizado de máquina; Groundwater hydrology; Groundwater level; Water table; Time series; Machine learning; CNPQ::ENGENHARIAS::ENGENHARIA CIVIL

Historical records of groundwater levels in tubular wells are important for environmental studies in water resources field. Monitoring the level of groundwater sources is essential for their management as it provides the necessary support to estimate their quantity to identify possible changes in water flow. In Rio Grande do Sul state, southern Brazil, groundwater is one of the main sources of water in several locations, being mostly used for public supply and agricultural activities. The prediction of groundwater levels is a matter of interest to public and private agencies, since they provide data for the construction of groundwater scenarios, necessary for the integrated management of water resources. Methodologies structured in monitored data and using artificial intelligence in the context of machine learning were implemented to carry out the prediction of groundwater levels. The objective was to obtain models to produce the prediction of daily data of groundwater levels in monitoring wells in light of techniques based on programming and computer learning. From an architecture composed of empirical models guided by climatological data, machine learning techniques were applied to simulate the behavior of groundwater levels in different monitoring wells. Models were built from techniques of Artificial Neural Networks, Support Vector Machines, Random Forests and Gradient Boosting Machine. The results achieved evidenced the good performance in the use of methods based on machine learning, being approaches that present great advantages in view of their high accuracy and predictive attributes. The learning techniques and the structuring of models guided by climatological and spatial data showed high learning rates, which was demonstrated by the error estimators employed. Thus, this work proposes support tools for the analysis, development and implementation of methodologies for modeling time series of groundwater levels in monitoring networks.