Predicting river discharges with deep learning models

28/11/2024

The recent evolution of large language models, such as ChatGPT, has showcased the surprising capabilities of machine learning models. These models hold significant potential for applications in the water resources field, such as predicting hydraulic variables, improving their usage and mitigating the adverse effects of floods and droughts. The present thesis aims to evaluate the effectiveness of deep learning models, particularly the Temporal Fusion Transformers (TFTs), to predict discharges in natural rivers in Portugal, and assess how they fare against classical hydrological modelling approaches. A comprehensive literature review is conducted on hydrological prediction, performance evaluation and the use of machine learning models in hydrological modelling. Precipitation, temperature, and other variables related to the studied catchments, such as drainage area, primary land use, and the Gravelius compactness index, are used to train and validate TFTs, both applied to specific gauging stations and generalised to work in the whole of mainland Portugal. Numerous simulations were performed to assess the behaviour of TFTs and their sensitivity to different hyperparameters and training data sets. The predictions achieved in this work outperformed the very popular “Hydrologiska Byråns Vattenbalansavdelning” (HBV) “classical” model, successfully predicting discharges even in “ungauged” catchments. This work establishes a framework for applying TFTs to river discharge prediction, paving the way for their use in operational forecasting contexts and resorting to larger datasets, eventually global. It is a work with profound implications that opens the way for cutting-edge research and numerous operational applications.

Autores da comunidade :

• 

  R

  Rafael Da Silva Francisco

  ist196084

Orientadores desta instituição:

• 

  J

  José Pedro Gamito de Saldanha Calado Matos

  ist151906

Mestrado em Engenharia Civil

civil-engineering - Engenharia Civil

Palavras-chave

• river discharge
• hydrological prediction
• hydrological models
• deep learning
• Temporal Fusion Transformer
• probabilistic prediction

eng - Inglês

Instituto Superior Técnico