Improvement of Discharge Prediction in Ungauged Swedish Basins: A Data-Driven Approach

Sammanfattning

Discharge prediction in ungauged basins remains a significant challenge in hydrology. In Sweden, the process-based S-HYPE model is used for nationwide simulations, yet there is potential for improving local accuracy through data-driven approaches, as these simulations do not always align with observed measurements. This study aims to develop a transparent hybrid workflow to enhance discharge predic-tion in ungauged basins. The analysis includes 223 gauged basins across Sweden, which are grouped into five clusters based on 12 hydrological signatures. A Random Forest regression model was developed and implemented as a post-processing tool for each cluster to correct systematic errors in S-HYPE simulations. Selected gauged basins are treated as ungauged basins (so-called pseudo-ungauged) to assess the model’s ability to predict discharge in un-gauged basins. This hybrid approach demonstrates good performance in both temporal and spatial gen-eralization with median NSE values of 0.85 and 0.84, respectively. To account for uncertainty in region-alization, a Monte Carlo simulation was used to assess the impact of basin misclassification on predic-tion accuracy. The results indicate that the hybrid workflow is robust to classification uncertainties, maintaining stable performance across most scenarios with small standard deviations. Overall, the find-ings suggest that integrating physical model outputs with machine learning can significantly improve spatial generalization over standalone data-driven models. However, analysis of Flow High Volume (FHV) reveals a systematic negative bias for the top 2% of discharges. Furthermore, a frequency analysis using time series of annual maxima shows that the hybrid approach improves median errors and error distributions in three out of five identified clusters compared to S-HYPE. The study highlights the po-tential of hybrid modeling to refine national-scale hydrological simulations while identifying specific limitations in capturing high-flow extremes.