Poyang Lake, the middle Yangtze River basin. Lake water levels are essential indicators for flood control, drought risk management, and water resource allocation. Due to dynamic, variable river-lake connectivity and intense anthropogenic perturbations, floodplain lake water levels are more challenging to predict accurately. However, they can be treated as multi-frequency signals whose predictability can be enhanced by separating them into simpler, more regular components using signal decomposition. This study proposed a data-driven framework for predicting floodplain lake water levels using machine learning, guided by signal decomposition, and tested it at Poyang Lake. The framework seamlessly integrated signal decomposition (Variational Mode Decomposition, VMD), machine learning (LightGBM or LSTM), and a hyperparameter tuner (Optuna) to form VMD-LightGBM-Optuna and VMD-LSTM-Optuna prediction models, respectively. In addition, the SHAP model was used to improve the framework’s interpretability regarding input features. The framework, which achieved high prediction accuracy at Poyang Lake, has been proven to be a reliable tool for predicting floodplain lake water levels. Its accuracy during the 2022 extreme drought of Poyang Lake provides vital implications for the different influences of weak representativeness of hydrological training data on the performance of LightGBM and LSTM. The feature rankings from SHAP analysis show that discharge from the Yangtze and Ganjiang Rivers rank as the top two, consistent with the hydrological background of Poyang Lake. • Signal analysis boosts machine learning's ability to predict lake water levels. • A framework for accurately predicting floodplain lake water levels was proposed. • Generalization of hydrological training data affects data-driven model performance. • The input feature’s rank conforms to the Poyang Lake’s hydrological background.
Gan et al. (Thu,) studied this question.