Deep-learning prediction of high-frequency sea levels in the Adriatic Sea

Intense high-frequency sea-level oscillations (HFOs) in the Mediterranean Sea, sometimes leading to destructive meteotsunamis, are generated by specific meteorological conditions that are spatially limited (from several tens to a few hundred kilometers). Although the physical mechanisms driving extreme HFOs are well understood, existing forecasting systems based on hydrodynamic models remain unreliable and computationally demanding.

To address these limitations, we developed deep-learning models (CNNs and ViTs) to predict HFOs using data from the Adriatic tide-gauge station Bakar, which provides a long record (2003–2025) but is not particularly prone to meteotsunamis. Models trained at Bakar can, however, be transferred to meteotsunami-prone Adriatic locations with shorter data records (Stari Grad, Vela Luka, Mali Lošinj, Sobra, etc.). Models were trained using measured 1-minute sea levels together with two sources of simulated atmospheric data (2D and 3D fields): hourly ERA5 data at 30 km resolution and 3-hourly CERRA data at 5.5 km resolution.

We will present model architectures and predictions of HFO amplitudes as a function of (i) forecasting horizon (up to several days, using different input windows of 6 h and 24 h), (ii) atmospheric data source (ERA5 vs. CERRA), and (iii) different combinations of training, validation, and testing periods. The main findings are as follows: (i) daily HFO amplitudes remain reasonably predictable over multi-day horizons, with comparable results from CNN and ViT approaches; (ii) forecast skill is higher for low-amplitude HFOs (up to ~12 cm); (iii) higher-amplitude events (10–40 cm) are generally underestimated; (iv) higher-resolution atmospheric forcing (CERRA) does not improve forecast skill, suggesting that meteotsunami-triggering atmospheric disturbances are not better represented at higher resolution; and (v) the choice of training, validation, and testing intervals has little effect on forecasting of small-amplitude events but affects forecasts of larger-amplitude HFOs.