Water level forecasting in coastal cities using a hybrid deep learning approach

Area of Science:

• Hydrology and Water Resource Management
• Artificial Intelligence and Machine Learning
• Climate Change Adaptation

Background:

• Coastal cities face significant challenges in real-time flood prediction due to complex hydrological dynamics and scarce historical data.
• Accurate forecasting is crucial for developing effective flood early warning systems, especially in climate-sensitive regions.

Purpose of the Study:

• To introduce and evaluate a novel hybrid deep learning model for accurate hourly water level forecasting in Venice, Italy.
• To assess the model's performance against established baseline methods under various data conditions.

Main Methods:

• Development of a hybrid deep learning model, CNN-Transformer-SKANs, integrating Convolutional Neural Networks (CNNs), Transformer layers, and Swallow Kolmogorov Arnold Networks (SKANs).
• Training the model on two years of high-resolution meteorological and hydrological data, including wind speed, tide level, humidity, atmospheric pressure, and prior water levels.
• Comparative analysis against baseline models like LSTM, CNN-LSTM, and Transformer variants.

Main Results:

• The CNN-Transformer-SKANs model achieved superior accuracy, with a Nash-Sutcliffe Efficiency (NSE) of approximately 0.99 and Root Mean Square Error (RMSE) below 0.03 m.
• The model demonstrated robustness under reduced training data scenarios and synthetic extreme value simulations, maintaining high accuracy (RMSE 0.02-0.03 m, NSE up to 0.99).
• Outperformed traditional LSTM and CNN-LSTM baselines, which showed RMSEs of 0.04-0.07 m and NSEs of 0.90-0.97.

Conclusions:

• The proposed CNN-Transformer-SKANs model offers a reliable and highly accurate solution for operational flood early warning systems in coastal urban environments.
• The model's strong generalization performance makes it suitable for climate-sensitive regions requiring advanced flood prediction capabilities.
• Hybrid deep learning architectures show significant promise for addressing complex hydrological forecasting challenges.