Lischeid, G; Uhlenbrook, S: Checking a process-based catchment model by artificial neural networks, Hydrological Processes, 17, 265-277 (2003), doi:10.1002/hyp.1123 | |
Abstract: A process-based, conceptual runoff model, the tracer-aided catchment (TAC) model, is developed and applied to the 40 km2 Brugga catchment in the Black Forest Mountains of southwest Germany. The model accounts for different runoff generation processes and runoff components. A lumped artificial neural network (ANN) model was also applied using the same data set. We compared the output from the two models in an effort to improve the efficiency of the TAC model and its model structure using physical and chemical data. Both approaches yielded comparably good results for daily discharge simulation. The TAC model, which uses spatially distributed input data, was superior to the lumped ANN model in only one case. Dissolved silica concentration in catchment runoff was used to assess the runoff components mixing approach of the conceptual TAC model. The ANN simulation of silica time series was superior to that of the TAC model. However, using modelled discharge values instead of measured data markedly weakened the silica predictions and resulted in a performance similar to the TAC model. We conclude that this reflects the inevitable non-linear error propagation within the TAC model. The ANN silica model improved substantially when hourly data were used instead of daily data. According to the ANN, silica dynamics are characterized by hysteresis loops in the discharge-silica concentration relationship in the short term. In addition, this relationship depends on the antecedent moisture conditions in the medium term. About 87% of the hourly silica time series variance was explained by non-linear regression with the hydrograph only. The implications of these findings are as follows. A simulation of the silica dynamics is possible using rather simple models that take into account the dynamics described above. In this respect, the TAC approach was confirmed by the ANN. However, an appropriate silica modelling time step must be shorter than daily values to be in line with the process time scale. Last but not least, the silica time series were assumed to reflect the spatial distribution of single hydrotopes to the catchment runoff during single storms, and thus to provide an independent check for the spatially distributed model in addition to the hydrograph. This has proven wrong, as the information provided by the silica data is highly redundant to that of the hydrograph, illustrated by the fact that the silica dynamics can be well described by the empirical model without requiring any spatial information. Copyright © 2003 John Wiley & Sons, Ltd. |