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

Accurate and real time flood prediction in coastal cities remains challenging due to the nonlinear and dynamic nature of the hydrological processes, coupled with limitations in long-term historical data availability. This study introduced a hybrid deep learning model, CNN-Transformer-SKANs, that combines convolutional neural networks (CNNs), Transformer layers, and Swallow Kolmogorov Arnold Networks (SKANs) to forecast the hourly water level in Venice, Italy. The model is trained on two years of high resolution meteorological and hydrological data, including variables such as wind speed, tide level, humidity, atmospheric pressure and prior water level. Comparative results with baseline models such as LSTM, CNN-LSTM, Transformer and their KANs enhanced variants demonstrate that proposed CNN-Transformer-SKANs achieves superior accuracy (Nash-Sutcliffe Efficiency, NSE ≈ 0.99; Root Mean Square Error, RMSE <0.03 m) and robustness even under reduced training data scenarios and synthetic extreme value simulations, achieving RMSE as low as 0.02-0.03 m and NSE up to 0.99, outperforming traditional LSTM and CNN-LSTM baselines (RMSE 0.04-0.07 m; NSE 0.90-0.97). The proposed architecture exhibits strong generalization performance, making it a reliable tool for an operational flood early warning system in climate sensitive regions.