The intersection between artificial intelligence and hydrology: scientific advances from IHCantabria

The results of two scientific articles recently published in the Journal of Hydrology highlight an innovative interdisciplinary approach that combines artificial intelligence (AI) and hydrology to address regional challenges. This is a study developed entirely at the Institute of Environmental Hydraulics of the University of Cantabria (IHCantabria), by PhD student Farzad Hosseini, under the supervision of researcher Cristina Prieto Sierra and the head of the Hydraulic Engineering Group of this institute, César Álvarez Díaz.

This study deals with the optimization of ensemble learning models, specifically long and short term memory networks (LSTM), to improve the prediction of streamflow and water levels in watersheds of the Basque Country. This region, characterized by a wet regime and fast responses to heavy rainfall, is an ideal scenario to test the capability of these advanced models.

From the first results of this study, the research team has shown that randomized search for hyperparameter optimization significantly improves the performance of LSTMs in regional hydrological environments. This breakthrough has addressed a historical challenge in hydrology: the need to develop unique models for each river basin, due to the uniqueness of its characteristics. Now, it is possible to unify predictions through an optimized regional framework, which represents a paradigm shift in the discipline.

The second article extended this knowledge by exploringensemble learning, an approach that promotes the collaborative work of multiple LSTM models (which are designed individually for each watershed). This method can contribute to improve the accuracy of predictions, as well as proving to be particularly effective in complex basins or basins influenced by external factors, such as snow or large reservoirs.

Additionally, from the field of artificial intelligence, this study has provided key information on the behavior of neural networks. The results highlight how the configuration of the hyperparameters and the architecture of the model influence its learning capability.

These advances not only have scientific but also practical implications. By improving hydrological predictions, early warning systems and water resource management strategies are optimized, which is crucial in regions such as the Basque Country, where flash floods represent a constant risk. Furthermore, the methodology developed has the potential to be replicated in other regions of the world, contributing to sustainable water management and mitigation of risks associated with climate change.

The research published in the Journal of Hydrology not only answers critical and currently relevant scientific questions, but also paves the way for future interdisciplinary collaborations. The impact of this work reinforces IHCantabria’s commitment to finding innovative solutions to solve problems related to global hydrological challenges.

Recently published articles can be consulted through the following links:

• “Hyperparameter optimization of regional hydrological LSTMs by random search”.
• “Ensemble learning of catchment-wise optimized LSTMs enhances regional rainfall-runoff modelling”.

Photo: Nervion River (stock photo)