Estrutura de modelos hidrológicos e sua inter-relação com atributos físicos e assinaturas hidrológicas

• Ano de defesa: 2023
• Autor(a) principal: Gustavo de Oliveira Corrêa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ENG - DEPARTAMENTO DE ENGENHARIA SANITÁRIA E AMBIENTAL; Programa de Pós-Graduação em Saneamento, Meio Ambiente e Recursos Hídricos; UFMG
• Palavras-chave em Português: Modelagem hidrológica; FUSE; Aprendizado de máquina; LSTM; Engenharia sanitária; Recursos hídricos - Desenvolvimento; Hidrologia - Modelos; Aprendizado do computador; Hidrologia
• Link de acesso: http://hdl.handle.net/1843/60717

For some time, the hydrological literature has been seeking a relationship between hydrological signatures, physical attributes of basis, and hydrological models. Understanding this interrelation is fundamental to understanding and representing the behavior of the basins, especially given the limitations associated with available monitoring. Among the various classifications of hydrological models, one classifies them based on physical principles and data-driven approaches. Against this backdrop, there is growing discussion about comparing these approaches and how much the hydrological signatures and physical characteristics of the basins impact on the effectiveness of the models, using extensive public hydrometeorological databases. In this regard, Long Short-Term Memory (LSTM) recurrent neural networks have gained prominence in hydrological studies due to their performance compared to conceptual models. Given the myriad options for conceptual models, some studies aim, through multi-model frameworks, to identify and analyze the appropriateness of the conceptual model structure to different watersheds, assuming that hydrological models are intrinsically connected to the physical properties of watersheds given the complex interactions between climate, soils, vegetation, and topography. Something that has not yet been explored is comparing LSTM networks against the best conceptual model structure calibrated individually for each basin. This research aims to identify and analyze the interrelation between hydrological signatures, the physical characteristics of the basins and the effectiveness of different hydrological model structures, whether physical process-based or data-driven. To this end, it employs LSTM networks and, extending beyond the referenced studies, uses a Flexible Model Structure (FUSE) framework to evaluate this suitability. An extensive database of public basins was used, evaluating the multi-object calibration of conceptual models using hydrological signatures, the estimation of signatures by physical characteristics, the characterization of homogeneous groups, and the relationship between signatures, physical attributes and models, whether conceptual or data-driven. The study reiterates the conclusions of other researchers by indicating that calibration based on hydrological signatures is appropriate and extends the physical interpretability of conceptual models. The results also indicate that physical characteristics can be used to infer the behavior of hydrological signatures in basins without hydrometric monitoring and propose five groupings by similarity of hydrological metrics, with spatial consistency. The results reinforce the difficulty of establishing a relationship between hydrological signatures, physical characteristics and the structure of the conceptual model, but provide a series of association rules capable of supporting the selection of appropriate structures to represent the phenomena. Finally, the study highlights the use of hydrological signatures as input to LSTM networks, validating the hypothesis that the use of hydrological signatures as a static layer, instead of physical attributes, allows similar or better results to be achieved, with advantages already demonstrated in studies based on the analysis of models based on physical processes.