Neural Network Early-Warning Prototype for Coastal Wave Forecasting in the Southern Tyrrhenian Sea (Italy)

Effective short-term wave forecasting is essential for coastal preparedness, particularly in regions with limited observations and exposure to hazardous sea states. This work presents a lightweight neural-network early-warning prototype for the Southern Tyrrhenian Sea (Italy), developed to support local authorities. The system produces 3-24 h forecasts of significant wave height (𝐻􀯦) and integrates an operationally structured workflow intended for real-time decision support. The forecasting model is based on a Temporal Convolutional Network trained on ERA5 wave and atmospheric reanalysis data covering the period 1985–2023. It uses a 48-hour window of wave, wind, and pressure parameters, complemented by short-term change indicators, to generate 𝐻􀯦 predictions at five forecast horizons. Training is conducted in five chronological folds with warm starts, ensuring temporal consistency across decades. Calibration layers and weighted loss functions are applied at each horizon to improve performance under severe conditions. Evaluation includes long-term performance assessment and analysis of storm events, demonstrating reliable forecasts up to approximately 12 hours and useful predictive capacity at 24 hours. The model is embedded within a near-operational pipeline aligned with regional civilprotection procedures, serving as a preparatory framework for retraining and deployment using local real-time data. Forecasts are classified into Green–Red alert levels, and the update frequency increases from 3-hour to 1-hour cycles under severe conditions, with optional automated notifications in operational mode. This prototype is transferable to other coastal areas through retraining on locally available buoy or model data, highlighting the potential of compact neural-network models to strengthen coastal hazard resilience.