Investigating strategies to improve hydrologic model performance in a changing climate

It has been repeatedly shown that conceptual hydrologic model performance degrades under conditions that deviate from those of the calibration period. In this study, we describe three experiments that aim to understand and address this problem using the conceptual model GR4J over 164 Australian catchments. The first is an investigation of model transferability, where parameters calibrated under certain conditions are applied to simulate both similar and contrasting conditions. We find that model performance degradation is more dependent on the conditions under which the model is tested than the conditions under which it is calibrated. Because dry periods are typically more difficult to simulate than wet periods, this means that transferring dry-calibrated parameters to wet periods is more successful than transferring wet-calibrated parameters to dry periods. For both wet and dry periods the best results were obtained when climatically similar calibration periods were used, suggesting that targeted use of climatically similar calibration data could improve predictive capacity. To this end, a second experiment was designed that preferentially weights modeled series calibrated under different conditions, with series associated with more climatically similar calibration periods weighted more heavily. While this improved model performance in most cases, the success was variable across the different catchments. Given that the model weighting scheme could not easily be generalized to all catchments in the sample, a third experiment was conducted where each model parameter was defined dynamically as a function of climate. The dynamic parameters were calibrated separately for each model, so the individual sensitivities of each catchment to climate conditions could be captured. While this also gave performance improvements, especially under drier testing conditions, the results continued to vary between catchments and there was no clear pattern in the parameter variation. This suggests that nonstationarity can be captured in different parameters for models of different catchments. While both model weighting and dynamic parameters can benefit overall conceptual model performance, it seems that reliable improvements across large samples of catchments may be difficult to achieve without more physically realistic model structures.