- 1University of Bologna, Department of Physics and Astronomy (DIFA), Bologna, Italy (lorenzo.mentaschi@unibo.it)
- 2CMCC Foundation - Euro-Mediterranean Center on Climate Change, Italy
- 3Departamento de Ciencias y Técnicas del Agua y del Medio Ambiente, University of Cantabria, 39005 Santander, Spain
- 4Department of Civil, Chemical and Environmental Engineering, University of Genoa, 16145 Genoa, Italy
- 5Department of Marine Sciences, University of the Aegean, 81100 Mitilene, Greece
- 6HS Marine SrL, 35027 Noventa Padovana, Italy
Effective storm surge prediction is vital for safeguarding coastal communities and enhancing disaster preparedness, particularly as climate change amplifies the frequency and intensity of extreme events. Despite the growing application of Machine Learning (ML) in storm surge downscaling, systematic comparisons with high-resolution dynamical models and focused assessments of extreme events remain underexplored. This study bridges these gaps by comparing advanced dynamical modeling with ML techniques to improve storm surge forecasting in the Northern Adriatic Sea.
High-resolution simulations were conducted using the SHYFEM-MPI model, leveraging optimized physical configurations and high-quality forcing datasets. This benchmark model demonstrated strong accuracy in representing storm surge dynamics and extremes, serving as a reference for evaluating ML-based approaches. To explore ML potential, models ranging from Multivariate Linear Regression (MLR) to the more advanced Long Short-Term Memory (LSTM) networks were developed and tested. A novel validation metric, the corrected mean absolute deviation (MADc) [1], and a tailored loss function (MADc2) were employed to improve model performance, particularly for extreme event prediction.
Results highlighted that while MLR offered computational efficiency, it struggled to capture non-linear dynamics and extremes. In contrast, LSTM networks excelled at modeling temporal dependencies and non-linearities, particularly when trained using the MADc2 loss function. Training ML models on outputs from the dynamical model revealed that MADc2-based architectures aligned closely with observations, offering a cost-effective alternative to traditional downscaling when high-quality forcing data is unavailable. Moreover, direct training on observed data at key sites such as Punta della Salute and Trieste showed that ML models, including LSTM, could outperform the dynamical model on critical metrics, underscoring the value of observational data.
This study underscores the promise of ML approaches in storm surge prediction, especially when integrated with high-quality data sources. By offering accurate predictions with significantly lower computational demands, ML techniques present a compelling case as efficient alternatives to traditional numerical models. As data accessibility and computational methods continue to advance, ML approaches may redefine the future of storm surge forecasting, enabling more sustainable and cost-effective solutions for coastal resilience.
[1] Campos-Caba, R., Alessandri, J., Camus, P., Mazzino, A., Ferrari, F., Federico, I., Vousdoukas, M., Tondello, M., and Mentaschi, L. (2024). Assessing storm surge model performance: what error indicators can measure the model’s skill? Ocean Science 20, 1513-1526. https://doi.org/10.5194/os-20-1513-2024.
How to cite: Mentaschi, L., Campos-Caba, R., Alessandri, J., Camus, P., Mazzino, A., Ferrari, F., Federico, I., Vousdoukas, M., Tondello, M., and Coppini, G.: Storm surge prediction in the Northern Adriatic Sea: a comparison between Machine Learning and numerical modelling, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17094, https://doi.org/10.5194/egusphere-egu25-17094, 2025.