A Hybrid Data-Driven Deep Learning Prediction Framework for Lake Water Level Based on Fusion of Meteorological and Hydrological Multi-source Data

Accurate prediction of lake water level is of great significance for flood prevention, reservoir scheduling, and ecological protection. However, the change in lake water level is influenced by multiple factors, and water level data as a time series also have the characteristics of complexity, which leads to difficulty in water level prediction. In view of this, a hybrid CEEMDAN–BiGRU–SVR–MWOA (CBSM) framework is proposed here for lake water level prediction based on multiple sources of environmental, hydrological and meteorological factors. Firstly, the lake water level is decomposed into modal data with different frequencies using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN). Then, the modal data are divided into internally and externally controlled items using the maximum information coefficient (MIC). Then, multivariate prediction is performed by combining the external data using bi-directional gated recurrent unit (BiGRU). The prediction results are combined using support vector regression (SVR) for prediction with higher accuracy. A modified whale optimization algorithm (MWOA) is used to optimize the parameters of CEEMDAN and the hyperparameters of BiGRU based on permutation entropy and mean square error, respectively. The proposed CBSM was applied to Dongting Lake, China, and the Nash–Sutcliffe efficiency of the prediction results reached 0.997, which was better than those of other benchmark models or frameworks, and the Diebold Mariano (DM) test further demonstrated the superiority of the proposed CBSM. Thus, this research provides a new and effective method for accurately simulating and predicting lake water levels.