Towards Interpreting Machine-Learning Models for Multi-Step Ahead Daily Streamflow Forecasting

Streamflow forecasting using interpretable machine learning methods (MLs) for exploring runoff processes has received a lot of attention. However, exploring multi-step ahead daily streamflow forecasting considering antecedent streamflow as an input for various interpretable MLs is very limited. Thus, three interpretable MLs for daily streamflow forecasting in the Huaihe River basin of China during 2002–2020, including eXtreme Gradient Boosting (XGBoost), long short-term memory neural network (LSTM) and convolutional neural network (CNN) with SHapley Additive exPlanations (SHAP) method, were implemented to study the role of potential controlling factors, including antecedent streamflow, soil moisture and vegetation growth, in runoff processes at lead times of 0–6 days. The forecasting performances decreased with lead times. Specifically, the LSTM model performed best at lead times of 0–3 days, followed by CNN and XGBoost. CNN was superior to LSTM and XGBoost models when the lead time was greater than 3 days. The optimal forecasting performances were 0.71–0.97, 311.45–674.27 m³/s, 0.84–0.97 and 0.75–0.97 according to Nash-Sutclife efficiency, root-mean-square error, correlation coefficient and Kling-Gupta efficiency, respectively. The interpretable results varied across different MLs and at different lead times. The antecedent streamflow consistently dominated the runoff processes, particularly in the LSTM and XGBoost models. However, the significant role of soil moisture at the depth of 28–100 cm and leaf area index for low vegetation gradually emerged with increased lead times for CNN models, even outranking the importance of antecedent streamflow. Furthermore, the interpretability demonstrated by the optimal machine learning models was validated through the infiltration model and uncertainty analysis. Overall, interpretable machine learning has great potential to enhance our understanding of basin-scale runoff processes.