Enhanced tidal level forecasting: A hybrid CEEMDAN-LSTM framework with Bayesian optimization

Abstract

Accurate tidal forecasts are crucial for preventing flood disasters and saltwater intrusion caused by high tides. To this end, this study proposes a hybrid CEEMDAN–LSTM–BO framework to improve tidal forecast accuracy. First, the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is used to perform multiscale signal decomposition on the raw tide level time series; Subsequently, the decomposed components are fed into an LSTM model, and Bayesian optimization is used to determine the optimal hyperparameters for each component; finally, the optimal forecast values for each component are aggregated to obtain the final forecast results. Based on observed tide level time series data, a performance comparison and analysis were conducted for the four models: LSTM, LSTM–BO, CEEMDAN–LSTM, and CEEMDAN–LSTM–BO. The proposed CEEMDAN-LSTM-BO model achieved a coefficient of determination of 0.9971 and passed the statistical significance test; in seasonal forecasting, the root mean square error, mean absolute error, and mean absolute percentage error were 0.041 m, 0.034 m, and 1.683%, respectively, and the model maintained reliable forecasting capabilities even during storm surges. The comparison results indicate that this framework offers significant advantages: not only does it achieve outstanding forecasting accuracy, but its forecasting time also meets real-time requirements, providing critical support for the early warning of secondary disasters.