Modeling Daily River Discharge Using Machine Learning Ensembles in the Context of Climate Change: Application To the zhaiyk-caspian basin, Kazakhstan

Abstract The study presents a comparative assessment of eight machine learning (ML) algorithms - Random Forest (RF), Lasso Regression (LASSO), AdaBoost (ADB), Gradient Boosting Regressor (GBR), Extreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), Light Gradient Boosting Machine (LGBM), and K-Nearest Neighbors (KNN) - for modeling daily river discharge at ten hydrological stations within the Zhaiyk - Caspian water management basin. Model performance was evaluated using mean absolute error (MAE), mean squared error (MSE), and symmetric mean absolute percentage error (SMAPE). The highest predictive accuracy (MAE ≈ 0.3) was achieved by ensemble tree-based methods (Random Forest, CatBoost, Gradient Boosting, LightGBM, XGBoost), while LASSO and AdaBoost exhibited the weakest performance (MAE ≈ 22). Identifying the most significant predictors enhanced both model interpretability and forecasting quality. The findings highlight the importance of tailoring ML approaches to the specific characteristics of river basins and suggest promising prospects for their integration with physically based hydrological models to improve river discharge forecasting and strengthen water resources management under climate change conditions. Graphical Abstract The Graphical Abstract Presents an Ensemble ML Framework for Modeling Daily River Discharge Under Climate Variability in the Zhayik–Caspian basin, Kazakhstan. Input Data Include Observed Daily runoff, Air temperature, and precipitation, Applied across Eight Machine Learning Algorithms: Random Forest, CatBoost, LightGBM, Gradient Boosting, KNN, LASSO, AdaBoost, and XGBoost. The graphical summary illustrates: the study area, the structure of the modeling framework, key predictors influencing runoff, and a comparative analysis of observed and simulated discharge. The results highlight the potential of ensemble machine learning methods for improving runoff prediction in arid regions and their application in water resource planning, flood and drought risk management, and climate change adaptation strategies.