Hybrid framework for robust runoff forecasting via decomposition and machine learning

Accurate runoff forecasting is essential for flood control and water resource management, yet strong nonlinearity and seasonal non-stationarity often limit traditional models. This study proposes a hybrid decomposition-integration-optimization framework (TVFEMD-LSSVM-BiLSTM-FSS-IBKA), integrating time-variant filter empirical mode decomposition (TVFEMD), least squares support vector machine (LSSVM), bidirectional long short-term memory (BiLSTM), flood season segmentation (FSS), and an improved black kite algorithm (IBKA). TVFEMD mitigates noise and non-stationarity, LSSVM and BiLSTM capture nonlinear and temporal features, FSS distinguishes seasonal hydrological regimes, and IBKA adaptively optimizes ensemble weights. Using daily runoff data from Shebu and Dongjiang stations, the proposed model reduces RMSE by 42.5% and 39.8%, respectively, compared with the best benchmark, with KGE improvements up to 27.2%. The framework effectively captures flood peaks and maintains stability during non-flood periods, avoiding prediction lag and underestimation. Overall, the proposed model demonstrates strong accuracy, robustness, and applicability for runoff forecasting under complex hydrological conditions.