Waterschap Brabantse Delta (WBD) has the intention to implement measures that enhance the baseflow. Baseflow consists of the groundwater flow and a small part of the interflow. During dry periods, streams are dependent on the baseflow. Enhancing the baseflow has a proper effect on the ecologically relevant quality of waters, and is therefore wanted for WBD according to the regulations of the Water Framework Directive. It is needed to quantitatively examine these measures that have been implemented in subcatchments of WBD. Therefore, sufficient data of good quality is needed. Especially, the stream discharge itself is important to collect. In this research, the stream discharge is obtained by a physically based model (the GR4J rainfall-runoff model). Moreover, a new method is applied: relating groundwater heads to stream discharge by applying machine learning algorithms. These two different methods are used for subcatchment Chaamse Beken, for which it is wanted to simulate stream discharge between 2003-2019 (flow measuring weir has been removed in 2003).
Four different machine learning algorithms are used: decision tree regression (DTR), random forest regression (RFR), gradient boosting regression (GBR) and support vector regression (SVR). The training set of these models is set from 1985-1999, whereas the test set is from 1999-2003. Moreover, different input variables and combinations of these variables are chosen for the models: shallow wells (screen-1 wells), deeper wells (screen-2 wells), precipitation and potential evaporation. The model performance is evaluated with the metrics Nash-Sutcliffe Efficiency (NSE), mean absolute error (MAE), fourth root mean quadrupled error (R4MS4E) and mean squared logarithmic error (MSLE). The first two are considered for overall model performance, whereas the latter two are for high flow and low flow model performance.
The best overall and low model performance is obtained by using the algorithm SVR and using inputs groundwater heads of shallow wells, precipitation and potential evaporation (a NSE of 0.75). In order to examine if this machine learning model can be used in the future for stream discharge simulation, the SVR model is compared with an existing conceptual hydrological model GR4J. The GR4J model has a NSE value of 0.80 and can be rated as good. It has a larger NSE value than the SVR model and performs better than the SVR machine learning model.
It is important to stress that for the GR4J model the memory (or state) of the system is included. This inclusion of the memory of the system is not the case for the machine learning algorithms. Furthermore, an important difference is the fact that groundwater heads play a significant role in the simulation of the stream discharge by using machine learning algorithms. These groundwater heads are not directly used in the GR4J model. Lastly, it is stressed that for building the GR4J model physical understanding of the hydrological system is needed, whereas for machine learning this is not the case.
Overall, it can be concluded that GR4J is still favoured above SVR, but SVR shows promising results for further research in simulating stream discharge.