A hybrid deep learning model based on EMD algorithm for non-stationary water level prediction of estuarine systems

• Published in: Estuarine, coastal and shelf science Vol. 314; p. 109128
• Main Authors: Gao, Sheng, Feng, Xi, Xu, Hang, Wu, Yirui, Feng, Weibing
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2025
• ISSN: 0272-7714
• DOI: 10.1016/j.ecss.2025.109128

The randomness and complexity brought by multiple driving factors make it difficult to achieve fast and accurate water level forecasting. This study evaluates the application of three machine learning (ML) models (LSTM, GRU, CNN-LSTM) in non-stationary water level prediction, using the Yangtze Estuary as the test region. The statistical tool Empirical Mode Decomposition (EMD) is used for data pre-processing, and a novel integrated modeling system (EMD-ITG) is proposed. The EMD-ITG systematically combines EMD with various ML algorithms to enhance the accuracy of non-stationary water level forecasting. Results show that by using EMD-ML models, Root Mean Square Errors (RMSEs) are reduced by 5%–34% compared to conventional ML models. In tide-dominated areas (Wusong, Liyashan), GRU achieves the highest prediction accuracy, while in runoff-dominated areas (Zhenjiang), LSTM outperforms the other two. The EMD-ITG model, utilizing a high-frequency-GRU and mid-low-frequency-LSTM architecture, achieves the highest prediction accuracy at all stations, with RMSE reducing by 10–21% and Nash-Sutcliffe Efficiency (NSE) increasing by 0.2–1.2%, as compared to that of EMD-ML models. Additionally, all the EMD-ML models outperform popular harmonic tool, like NS_TIDE, with RMSE reduced by ∼30%. This study emphasizes that signal preprocessing and source interpretation are crucial for training ML models before resuming deep learning. The invented EMD-ITG model also provides a valuable reference for future hydrological forecasting. [Display omitted] Key points.1.Conventional and EMD-ML models were evaluated on nonstationary WL prediction in Yangtze Estuary.2.LSTM is superior for runoff dominated WL prediction, while GRU outperforms in tide-dominated WL.3.A hybrid EMD-ITG model was created in which decomposed WL signals go through matched ML algorithms.4.ML models are superior to NS_TIDE, while EMD-ML models outperform conventional ML models.5.EMD-ITG further reduced RMSE by 10–21% compared to EMD-ML model.