A multimodal runoff forecasting model based on clustering enhancement and attention mechanism

BeiPanjiang River Basin, China In this study, a novel clustered augmented attention time-series runoff prediction model, Cluster-Augmented Attention GRU (CA-AGRU), is proposed. The model adopts a multi-stage modeling strategy: firstly, the input factors for K-means clustering are screened based on SHAP value analysis, and then the K-means clustering algorithm is applied to divide the runoff process into a number of sub-processes with similar hydrological characteristics; secondly, a bidirectional GRU encoder-decoder architecture incorporating an attentional mechanism is constructed to model the sub-processes individually; and lastly, the final prediction is achieved through integrated learning. The results demonstrate that CA-AGRU significantly enhances the accuracy and robustness of runoff prediction while exhibiting superior adaptability to the nonlinear and non-stationary characteristics of runoff sequences. Compared with the other five baseline models, CA-AGRU achieves an average reduction of 13.09 % in RMSE, a 6.64 % decrease in MAE, and a 7.96 % improvement in NSE. Furthermore, this study found that the introduction of the attention mechanism and the encoder-decoder architecture enabled the model to effectively capture key hydrological period characteristics in runoff sequences. K-means clustering significantly improved the model's generalization ability, with a clustering number of 2 being more suitable for the Beipanjiang River Basin. These findings provide new insights into integrating data-driven models with physical processes in runoff forecasting. • CA-AGRU model combines K-means & attention-based GRU enc-dec for runoff prediction. • Two-cluster division (rise/decline) optimally capture watershed dynamics. • In this study, the K-means method outperformed other clustering methods. • CA-AGRU model's robustness verified via multi-metric performance assessment.