Author(s): Florentin Hofmeister; Alice Spadina; Gabriele Chiogna

Linked Author(s): Florentin Hofmeister

Keywords: Hydrological modeling; Support Vector Machine; Support Vector Regression; Computational time; Alpine hydrology

Abstract: Hydrological modeling of Alpine catchments is particularly challenging due to the high variability of hydrological processes in space and time. Although physically-based and fully-distributed hydrological models, such as WaSiM, are able to simulate these small-scale variabilities, the computational time for running a model on hourly time step and 25 m grid resolution in mesoscale catchments (10-100 km²) is significantly high. This becomes particularly relevant when large time periods (>30 years) are to be simulated for climate change studies. Therefore, we applied Support Vector Regression (SVR) to reproduce the results of a high-resolution WaSiM model (25 m grid, hourly time step) using as an input a coarse spatial (100 m grid) and temporal (daily) resolution of the model and hourly meteorological time series. For solving the SVR functions, we used two different approaches. The first approach fits an exact SVR model by applying a kernel function while the second approach fits a Gaussian kernel regression model using random feature expansion, which results in an approximation of the kernel function. As a result, the computational time was reduced by 93% for the model setup with hourly time step and 25 m grid resolution. The quality of the SVR results was quantified through different indicators: Root Mean Squared Error (RMSE), Standard deviation Ratio of RMSE (RSR), relative Error, Nash-Sutcliffe Efficiency (NSE) and logarithmic NSE (logNSE). Additionally, the SVR results were compared with the flow duration curve. All indicators show an excellent performance (NSE=0.89) of the SVR in reproducing WaSiM results. We tested the robustness of the SVR also considering different data, such as meteorological inputs from different stations and simulated discharges of subcatchments. Except for the cases of small subcatchments with little glacier contribution, very good performance levels were achieved.

DOI: https://doi.org/10.3850/IAHR-39WC252171192022902

Year: 2022