Streamflow Variability and Predictive Modeling in the Carpathian Basin: Assessing the Performance of Machine Learning Algorithms

The Carpathian Basin, is a climatically sensitive region influenced by Atlantic, continental and Mediterranean climates. Understanding the river dynamics in this region is crucial for sustainable water management given the diverse climatic and hydrological conditions. Despite extensive research, few studies have thoroughly compared the performance of advanced machine learning models for predicting river discharge in this region.

In this study we show that Random Forest (RF), LightGBM (LGBM), Support Vector Regression (SVR), Temporal Convolutional Networks (TCN) and XGBoost can improve streamflow prediction by utilizing their ability to capture nonlinear and temporal relationships in hydrological data. Using daily discharge data for 1961-2020 period from Danube, Sava, Tisa and Drava Rivers, we tested these models at six stations and analyzed their effectiveness using metrics such as RMSE, MAE and R². The Augmented Dickey-Fuller test confirmed the stationarity across all stations and thus confirmed the robustness of our prediction framework.

The RF model performed consistently better than the other models, achieving the lowest RMSE (e.g. 31.739 m³/s at Bezdan station and 19.582 m³/s in Donji Miholjac station) and the highest R² values (e.g. 0.999 at Szolnok and Bezdan station). In contrast, the SVR showed the weakest performance with significantly higher RMSE values and lower R² values at all stations. XGBoost and LGBM also performed strongly, but fell slightly short of the prediction accuracy of RF. These results emphasize the robustness of RF in capturing complex, nonlinear hydrological dynamics and its resistance to overfitting.

Our results suggest that Random Forest is the most reliable model for predicting discharge in the Carpathian Basin, providing high accuracy and robustness for different rivers. These results have significant implications for improving predictive hydrological models that enable more effective water resource management and adaptive strategies in climatically sensitive regions.