Innovative daily runoff prediction model integrating black-winged kite algorithm and Mamba2–Transformer architecture

The accurate prediction of daily runoff is crucial for effective water resource management, flood prevention, and disaster mitigation. However, current runoff prediction models face dual challenges in extracting discriminative features from random and non-stationary data, as well as enhancing prediction performance. To overcome these challenges and improve prediction accuracy, this study proposes a hybrid feature optimization and variational prediction model (HFOVPM). The HFOVPM combines an optimized feature extraction framework with advanced techniques of temporal pattern recognition. First, the Black-winged Kite Algorithm was employed to optimize variational mode decomposition parameters for effective signal decomposition. To enhance feature representation, the resultant components were systematically organized into multivariate input vectors. Secondly, the Mamba2 architecture was used to model nonlinear dynamical interactions within the multivariate inputs. Subsequently, its learned representations were then integrated into Transformer layers to establish enhanced global temporal dependencies. Through error back-propagation optimization, the framework improved the accuracy of predictions. The model was validated using daily runoff data from three hydrological stations representing distinct eco-hydrological regimes: an alpine snowmelt-dominated basin, a subtropical rainfall-driven basin, and a mixed rain-snow basin. The key findings were as follows: (1) The HFOVPM consistently outperformed several benchmark models across all basins, achieving Kling–Gupta efficiency values of 0.94–0.99 and maintaining the mean absolute percentage error below 14.2 %. (2) The model exhibited particular strength in predicting extreme runoff events, capturing flood peaks more accurately than comparative models, which is critical for flood risk early warning. (3) Its robust performance across contrasting climates indicates a reliable generalization capability for diverse hydrological processes. In summary, the HFOVPM effectively addresses the feature extraction and temporal dependency challenges in runoff prediction, providing an accurate and transferable tool that can support hydrological forecasting and water-related decision-making in varied environmental settings. • A HFOVMP deep learning model is proposed for accurate daily runoff prediction. • The residual error BP real-time correction mechanism achieves a prediction–correction closed loop. • The hybrid architecture establishes a local–global complementary temporal representation. • K-means clustering reveals the model's adaptability to different hydrological response modes. • Three distinct types of basins are selected to demonstrate the universality of the HFOVPM.