Predicting high-frequency sea-level oscillations in Bakar and Ploče using deep learning

Meteotsunamis are rare events of extreme high-frequency sea level oscillations (HFOs) which occasionally occur in specific Mediterranean ports and harbors, where they have the potential to cause substantial damage. These events are generated by spatially limited atmospheric disturbances (ranging from tens to hundreds of kilometers in scale) and arise only when properties of these disturbances (rate of air pressure change, period, speed and direction) are such that a resonant transfer energy between atmospheric disturbance and long ocean waves occurs, with the properties of the long ocean waves governed by the bathymetric characteristics of the shelf in front of bays/harbors. This interplay explains why several Croatian harbors (e.g., Vela Luka, Stari Grad, Široka, Vrboska) are particularly vulnerable to this phenomenon. Due to the complex interplay of forcing factors, meteotsunamis remain difficult to predict with hydrodynamic models.

We will present the first application of deep learning techniques to forecast height of HFOs at two Adriatic tide-gauge stations: Bakar and Ploče. Although these locations are not particularly prone to meteotsunamis, their long-term high-frequency sea-level records (available from 2003 onward) make them suitable for model development. The pretrained models created for these sites can be adapted and fine-tuned for use at other Adriatic locations, including meteotsunami-prone locations, with shorter data histories. We trained deep convolutional neural networks on measured sea-level data and 3D atmospheric fields (ERA5 reanalysis). A range of experiments tested different network architectures, input configurations, and prediction targets to determine the optimal setup.

The key findings are as follows: (i) the model can provide reasonable forecasts of daily maximum HFO amplitudes up to three days ahead; (ii) predictions are more reliable for low-amplitude HFOs; (iii) high-amplitude events tend to be underestimated; (iv) expanding the input dataset (e.g., extending the temporal window or including additional sea-level components) does not enhance prediction quality; (v) for 1-minute sea-level predictions, while for daily forecast horizons the model significantly underestimates the amplitude, much better performance is achieved over shorter forecast horizons (e.g., next six hours).