Abstract
Modelling and prediction of mean monthly flow are of particular importance for long-term planning in hydrology and water resources management. Therefore, a simplified and robust modelling procedure, derived from clearly and concisely established methodology, can benefit both researchers and practitioners. The main objective of this study is to develop a robust yet simple model, capable of producing predictions of satisfactory accuracy on previously unseen data. Two chronological data allocation strategies (C1 and C2), differing in the proportions of training, calibration, and verification subsets, were evaluated to analyze their influence on model accuracy and reliability. Chain and ensemble modelling techniques were applied, resulting in several stacking regressors with different combinations of base models and final estimators. The best-performing ensemble (C2) consisted of a support vector machine, histogram gradient boosting regressor, elastic net, and two dummy regressors as base models, with an artificial neural network as the final estimator. Within the ensemble structure, dummy regressors and histogram gradient boosting regressor were used to extend the predictive range, while elastic net and support vector machine captured the overall flow bias and fundamental flow dynamics. The artificial neural network final estimator was used to integrate these components into the final flow prediction. Compared to C1, the C2 allocations strategy achieved improved generalization capability and narrower confidence intervals due to the larger training subset, indicating higher model reliability for long-term monthly flow forecasting. The study additionally emphasizes the importance of appropriate methodological workflow, careful dataset treatment, and comprehensive model evaluation using complementary statistical and hydrological analysis tools.

This publication has 44 references indexed in Scilit: