Developing a novel hybrid deep learning model for extratropical storm surge forecasting: A case in the Bohai Sea

Abstract

Storm surges pose persistent threats to coastal communities, endangering both human lives and infrastructure. While numerical models remain computationally intensive, artificial intelligence (AI) approaches have emerged as efficient alternatives for storm surge forecasting through their superior accuracy and computational efficiency. However, most existing site-specific forecasting models rely on single-point wind and pressure measurements, neglecting the role of regional wind fields that limit the precision of extratropical storm surge forecasts. To address this gap, we developed a novel end-to-end multi-station forecasting framework designed to establish mapping relationships between wind-pressure fields and observational stations. We employ a 3D UNet for spatiotemporal feature extraction from atmospheric fields, followed by Multilayer perceptrons (MLPs) to project these features onto multiple monitoring sites, with integrated Long Short-Term Memory (LSTM) networks for temporal sequence modeling. Validation experiments in the Bohai Sea demonstrate the model's dual capability in multiscale feature abstraction and temporal dynamics capturing, enabling comprehensive storm surge process forecasting. The proposed model achieves significant reductions across multiple error metrics in 48- and 72-h prediction tasks compared to baseline models. This study provides theoretical and practical insights for advancing multi-step storm surge forecasting systems and hybrid models for coastal disaster prevention, particularly for extratropical storm surge.