The CCVCLSA model: A novel approach to medium- and long-term runoff prediction integrating multiple techniques

• Sample entropy-based secondary decomposition reduces runoff data complexity. • Integration of CNN, LSTM, and self-attention enhances feature and key information extraction. • The CCVCLSA model outperforms five other models in prediction accuracy and error control. • The CCVCLSA model effectively improves runoff prediction accuracy. Runoff sequences exhibit significant nonlinear and non-stationary characteristics, posing great challenges to hydrological prediction. This study develops the CCVCLSA runoff prediction model, integrating CEEMDAN, CPO, VMD, CNN, LSTM, and Self-Attention (SA) mechanisms, to effectively handle such complexity. After decomposing runoff sequences via CEEMDAN, Sample Entropy (SE) is used to filter high-complexity components, which are further decomposed by CPO-optimized VMD (CVMD) to convert the original sequence into smooth Intrinsic Mode Functions (IMFs) and residuals, significantly reducing non-stationarity and complexity. Subsequently, the preprocessed data are fed into a CNN-LSTM architecture integrated with SA to extract local features, capture long-term dependencies, and dynamically weight input information, enhancing the model's sensitivity to core features. Using multi-year runoff data from Hongjiadu, Xiajiang, and Manwan hydrological stations for training and testing, the CCVCLSA model is evaluated with five metrics (mean square error, MSE; mean absolute error, MAE; root mean square error, RMSE; Nash–Sutcliffe efficiency coefficient, NSEC; correlation coefficient, R) and compared with CNN, LSTM, CNN-LSTM, CNN-LSTM-SA, and CEEMDAN CNN-LSTM-SA models. Results show that CCVCLSA outperforms all other models in prediction accuracy: compared to CNN, it improves NSEC by 664.09 % and R by 45.15 % at Hongjiadu Hydropower Station, 589.88 % and 68.11 % at Xiajiang Hydrological Station, and 32.92 % and 15.31 % at Manwan Hydrological Station, respectively. These findings confirm that CCVCLSA effectively reduces the nonlinearity and non-stationarity of runoff data, with predictions closely matching actual observations.