Investigating the effects of local weather, streamflow lag, and global climate information on 1 month ahead streamflow forecasting by using XGBoost and SHAP: two case studies involving the contiguous USA

The use of machine learning (ML) models for streamflow forecasting has recently proved highly successful. However, ML is typically criticized for a lack of interpretability. Here, we develop an interpretable ML model for 1-month-ahead streamflow forecasting using extreme gradient boosting (XGBoost) and Shapley additive explanations (SHAP). In addition to a performance evaluation of XGBoost compared to regression tree and random forest approaches, the effects of input variables, including local weather, streamflow lag, and global climate, on streamflow were interpreted in terms of SHAP total effect values, main effect values, interaction values, and loss values. The experimental results at two catchments in the contiguous USA are significant in four ways. First, XGBoost was superior to the other two models in terms of Nash–Sutclife efficiency, mean absolute error, root mean square error, and correlation coefficient. Second, by aggregating SHAP values, we found that the contributions of these variables to streamflow differed according to the investigated local perspectives, including streamflow at different months, low streamflow, medium streamflow, high streamflow, and peak streamflow. Third, the SHAP main effect and interaction values revealed that nonmonotonic relationships may occur between the input variables and streamflow, and the strength of variable interaction effects might be related to the variable values rather than their correlations. Fourth, variable drifts in the testing set were deduced from SHAP loss values. These findings exhibit positive significance for understanding ML for monthly streamflow forecasting.

---

Opracowanie rekordu ze środków MNiSW, umowa nr SONP/SP/546092/2022 w ramach programu "Społeczna odpowiedzialność nauki" - moduł: Popularyzacja nauki i promocja sportu (2024).