Physics-based deep learning architectures for water level predictions

Keywords: Deep learning, artificial intelligence, transformers, water level
Abstract Type: Virtual Paper Abstract

Abstract

Accurate water level predictions have many benefits for the management of beaches, navigation, and coastal ecosystems. Tidal predictions reflect the influence of gravitational forces on water levels but do not consider weather forcings which can be a dominant factor depending on location and conditions. Improving water level predictions with lead times of hours up to a few days would help beach managers and other coastal stakeholders with their conservation efforts. Water levels are directly related to inundation, and predictions guide preparation measures to mitigate economic losses and other coastal flooding impacts. Hydrodynamic numerical models include atmospheric forcings and improve upon tidal predictions but require substantial information such as bathymetry and wind for the full model area and are computationally expensive.

Machine learning methods have been developed and implemented over the past twenty years, but the predictions are valid only for the location where the models are trained. Previous ML approaches focus on creating a deep architecture for the neural network but may not take full advantage of the dynamic of the water levels at the location of interest. This research proposes using physics-based LSTM, RNN, and transformer architectures to accurately forecast water levels at multiple locations in the Gulf of Mexico. The initial results have shown that by incorporating transformers into the model, it is possible to improve the results of previous methods in multiple locations in the Gulf of Mexico and extend the lead time of predictions while meeting NOAA requirements for operational models.