bachelorThesis
Towards the improvement of machine learning peak runoff forecasting by exploiting ground- and satellite-based precipitation data: A feature engineering approach
Fecha
2023-05-10Autor
Muñoz Pauta, Paul Andrés
Institución
Resumen
Peak runoff forecasting in complex mountain systems poses significant challenges in hydrology
due to limitations in traditional physically-based models and data scarcity. However, the
integration of machine learning (ML) techniques offers a promising solution by balancing
computational efficiency and enabling the incorporation of satellite precipitation products (SPPs).
However, debates have emerged regarding the effectiveness of ML in hydrology, as its black-box
nature lacks explicit representation of hydrological processes, hindering performance
improvement and result reproducibility. To address these concerns, recent studies emphasize the
inclusion of FE strategies to incorporate physical knowledge into ML models, enabling a better
understanding of the system and improved forecasting accuracy. This doctoral research aims to
enhance the effectiveness of ML in peak runoff forecasting by integrating hydrological concepts
through FE techniques, utilizing both ground-based and satellite-based precipitation data. For
this, we explore ML techniques and strategies to enhance accuracy in complex macro- and mesoscale
hydrological systems.
Additionally, we propose a FE strategy for a proper utilization of SPP information which is crucial for overcoming spatial and temporal data scarcity.
The integration of advanced ML techniques and FE represents a significant advancement in hydrology,
particularly for complex mountain systems with limited or inexistent monitoring networks.
The findings of this study will provide valuable insights for decision-makers and hydrologists, facilitating effective mitigation of the impacts of peak runoffs. Moreover, the developed methodologies can be adapted
to other macro- and meso-scale systems, with necessary adjustments based on available data
and system-specific characteristics, thus benefiting the broader scientific community.