Operational River Discharge Forecasting with Support Vector Regression Technique Applied to Alpine Catchments: Results, Advantages, Limits and Lesson Learned

• Published in: Water resources management Vol. 32; no. 1; pp. 229 - 242
• Main Authors: De Gregorio, Ludovica, Callegari, Mattia, Mazzoli, Paolo, Bagli, Stefano, Broccoli, Davide, Pistocchi, Alberto, Notarnicola, Claudia
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 2018; Springer Nature B.V
• Subjects: Atmospheric Sciences; Basins; Benchmarks; Catchment area; Catchments; Civil Engineering; Data assimilation; Datasets; Discharge measurement; Earth and Environmental Science; Earth Sciences; Environment; Forecasting; Gaging stations; Geotechnical Engineering & Applied Earth Sciences; Ground stations; Hydrogeology; Hydrology; Hydrology/Water Resources; Lead time; Learning algorithms; Learning behaviour; Machine learning; Precipitation; R&D; Research & development; River discharge; River flow; River forecasting; River networks; Rivers; Snow cover; Stream discharge; Support vector machines; Time series; Training; Variables; Water discharge; Support vector regression; Seasonal discharge forecast; Snow cover area; Machine learning; Alpine Arc
• ISSN: 0920-4741; 1573-1650
• DOI: 10.1007/s11269-017-1806-3

The main objective of this study is to derive a flexible approach based on machine learning techniques, i.e. Support Vector Regression (SVR), for monthly river discharge forecasting with 1-month lead time. The proposed approach has been tested over 300 alpine basins, in order to explore advantages and limits in an operational perspective. The main relevant input features in the forecast performances are the snow cover areas and the discharge behavior of the previous years. Forecasts obtained by training SVR machine on single gauging stations show better performances than the average of the previous 10 years, considered as benchmark, in 94% of the cases, with a mean improvement of about 48% in root mean square error. In case of poorly gauged basins, to increase the number of training sample, multiple basins have been considered to train the SVR machine. In this case, performances are still better than the benchmark, even if worse than those of SVR machine trained on single basins, with a decrease of the performances ranging from 13% to 54%.