Comparative analysis of machine learning models for discharge prediction: A case study at the Bratislava station on the Danube River

This study investigates the prediction of river flows for water resource management, flood and drought protection and hydropower planning, focusing on the limitations of traditional hydrological models due to complex data requirements. Machine learning (ML) models offer a solution by capturing non-linear relationships to improve prediction accuracy. We compare the performance of three data-driven models: Random Forest (RF), Gradient Boosting (GB), and Seasonal Autoregressive Integrated Moving Average (SARIMA), with a simple Linear Regression (LR) benchmark for monthly discharge forecasting at the Bratislava station on the Danube River. RF model performed best, with an MAE of 165.84 and an R² of 0.89. GB achieved moderate results (MAE = 335.15; R² = 0.60), while SARIMA performed below average (MAE = 550.78; R² = 0.01). Results are valid within a univariate autoregressive framework; multivariate extensions incorporating meteorological drivers are recommended for operational deployment. These results underline the potential of machine learning models, especially RF, for forecasting monthly river discharge up to 24 months in advance at the Bratislava gauging station on the Danube River. While these findings are based on a single-station, univariate analysis, they contribute to the growing body of evidence supporting the use of ensemble machine learning methods in operational hydrology for large regulated river systems. Transferability to other stations or basins would require further validation.