Using artificial neural networks to improve hydrological streamflow predictions from PCR-GLOBWB

Abstract

This project aimed to improve the streamflow discharge simulation performance of the PCR-GLOBWB hydrology model in the Rhine basin by modelling its residuals using artificial neural networks. Two architectures were used, a more generic fully connected network and a temporal convolutional network, as well as a multiple linear regression as a baseline. The predictors included a bunch of PCR-GLOBWB output variables (e.g. runoff components, groundwater recharge, snow, groundwater stores, etc.) and meteorological input variables (precipitation, temperature and reference potential evaporation), which were fed to the models either directly or by adding lagged versions of them of up to 60 days. The results showed increased performances to the original PCR-GLOBWB simulations, but no significant differences were found between the different machine learning models.