Enhancing hydrological rainfall-runoff simulation using machine learning methods

Hydrological rainfall-runoff modeling plays an important role in several water resource applications, including flood forecasting, hydroelectric power generation, and water supply planning. Moreover, accurate and reliable predictions contribute to better decision-making in these areas. While machine learning (ML) techniques have significantly advanced these models, challenges remain in improving interpretability, generalization to unseen conditions, and efficient handling of high-resolution data. This thesis advances the state-of-the-art by addressing these challenges.
Chapters 2 and 3 focus on the hybrid modeling paradigm, where data-driven techniques — more specifically, long-short term memory networks (LSTMs) — are integrated with conceptual hydrological models. Through a series of experiments, the hybrid models’ performance, interpretability, and generalization capabilities are assessed. The results demonstrate that hybrid models achieve state-of-the-art performance, comparable to stand-alone data-driven techniques, and surpassing traditional conceptual models. However, the experiments in Chapter 2 also reveal that, when coupled, the data-driven approach can compensate for structural deficiencies in the conceptual components. ... mehrThis suggests that relying solely on performance metrics for model selection in hybrid frameworks may be misleading. While hybrid models offer access to unobserved variables compared to stand-alone data-driven techniques, and provide some degree of interpretability, it is important to note that the interpretability derived from simplified basin average conceptual models is more associative than grounded in strict physical principles. Therefore, this interpretability should be taken as such.
Chapter 3 explores the generalization capabilities of hybrid models, particularly their ability to predict extreme discharges under out-of-sample conditions. Findings show that hybrid models generally perform similarly to stand-alone LSTM networks. However, stand-alone LSTMs excel in areas where the conceptual component of the hybrid model struggles with runoff generation assumptions. At the same time, hybrid models produce higher discharges for the most extreme cases of the dataset, where LSTMs are constrained by their theoretical saturation limit, defined during the training process.
While chapters 2 and 3 focus on interpretability and generalization, chapter 4 addresses the challenge of applying data-driven methods to sub-daily predictions, where computational costs remain a challenge. This issue is especially relevant, as most hydrological studies using LSTMs focus on daily-scale predictions, whereas applications like flood forecasting would benefit from higher temporal resolution to more accurately capture the dynamics of hydrographs. To overcome this limitation, chapter 4 introduces a technique that processes data at multiple temporal frequencies within a single LSTM cell. This approach enhances model generality and simplicity while maintaining state-of-the-art performance.
Overall, this thesis contributes to enhance hydrological rainfall-runoff simulation through machine learning methods by (1) evaluating the integration of machine learning into conceptual hydrological modeling and (2) advancing purely data-driven approaches.