@misc{9247666,
  abstract     = {{Extreme meteorological events are projected to increase in both frequency and intensity as climate change progresses, placing greater demands on accurate hydrological modelling for flood risk assessment and infrastructure design. In Sweden, the process-based S-HYPE model is widely used to simulate streamflow across approximately 26340 catchments, but like many regional rainfall-runoff models it systematically underestimates peak flows during extreme events, which poses a structural risk when designing culverts, retention basins, and urban drainage systems in ungauged catchments.

This study develops and evaluates a regression-based post-processing framework to quantify and correct this bias for small-to-medium Swedish watercourses with 100-year return period flows below 1000 $m^3/s$. Daily observed and S-HYPE-simulated streamflow data from 2010 to 2025 were compiled from an nationwide network of active gauging stations. Annual maximum daily flows were extracted and subjected to flood frequency analysis. The Generalized Extreme Value distribution fitted via L-moments (GEV-Lmom) was identified as the best framework, outperforming the Gumbel, Log-Normal, and Log-Pearson type III distributions across Anderson–Darling and Probability Plot Correlation Coefficient goodness-of-fit tests.

Baseline comparison confirmed that unadjusted S-HYPE outputs underestimate peak flows by a mean bias of $-7.3\%$, despite maintaining strong spatial--temporal correlation ($R^2 = 0.964$). To correct this systematic underestimation while preserving the model's ranking performance, three correction configurations, Full, Reduced, and Flow--only, were evaluated using Multiple Linear Regression (MLR) and Random Forest (RF) algorithms across return periods from 2 to 100 years. Feature selection via permutation testing isolated five statistically significant physiographic predictors for the Reduced configuration: modelled flow, catchment area, upstream area, regulation fraction, and agricultural land--use fraction. 

Both MLR and RF successfully reduced the baseline bias, reducing absolute peak flow errors to below $+1.4\%$ across all configurations. The RF model consistently achieved higher predictive accuracy ($R^2 = 0.990$) with tighter error distributions across all return periods, reflecting its capacity to capture non-linear catchment interactions. The log-linear MLR formulation, while slightly less accurate, produced an explicit and transferable scaling equation well suited for operational application at ungauged sites. The framework demonstrates that integrating statistical post-processing with physical catchment descriptors provides a scalable tool for improving extreme peak flow estimation across ungauged Swedish basins and while some uncertainties remain a future study could include a Peaks Over Threshold and a hold-out validation to improve on the results.}},
  author       = {{Johnsson, Jacob}},
  issn         = {{1101-9824}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{TVVR 5000}},
  title        = {{Adjustment of peak flows in water courses without gauging from S-HYPE data. A comparison between observed and modelled flow in Sweden}},
  year         = {{2026}},
}