Comparative study of daily streamflow prediction based on coupling SWAT+ with interpretable machine learning algorithms

In recent years, climate change has substantially affected the global water cycle, leading to an increase in the frequency and intensity of extreme hydrological events. Developing more accurate and efficient hydrological models is therefore essential for flood prevention, drought mitigation, and sustainable water resources management. In this study, four machine learning (ML) algorithms were coupled with SWAT+ to simulate streamflow in the Mishui River Basin (MRB). Both SWAT+–derived hydrological variables and raw meteorological observations were used as input features for the ML models, aiming to improve predictive performance. Additionally, the SHAP (SHapley Additive exPlanations) method was employed to quantify the contribution of different features to model predictions. During the validation period (2020−2023), the SWAT-Informer model exhibited the best performance, achieving R 2 and NSE values of 0.91 and 0.89, respectively. In contrast, improvements in streamflow prediction using DeepState and Bi-LSTM were less pronounced, with a notable performance decline during the testing period, likely due to the complexity of their multi-layer architectures. SHAP analysis revealed that precipitation was the most influential feature, contributing 29.1 % to the predictions. Moreover, SWAT+–derived outputs accounted for 64.9 % of the predictive power, highlighting the substantial value of SWAT+ in providing informative features for the ML algorithms. Overall, the four SWAT-ML coupled models outperformed the standalone SWAT+ model in streamflow prediction, demonstrating the considerable potential of ML techniques to enhance the performance of conceptual hydrological models such as SWAT+. Furthermore, the application of the SHAP method improved the interpretability of the models, fostering greater understanding, trust, and transparency for both researchers and decision-makers. • ML-enhanced SWAT+ improves streamflow predictions, especially for extreme events. • Informer model achieves the best performance using long-sequence attention mechanisms. • SHAP analysis reveals precipitation as the most influential feature. • SWAT+ output contributes 64.9 % to ML predictions, enhancing interpretability. • Regularization and hyperparameter tuning mitigate overfitting in DeepState and Bi-LSTM.