Why does a conceptual hydrological model fail to correctly predict discharge changes in response to climate change?

Several studies have shown that hydrological models do not perform well when applied to periods with climate conditions that differ from those during model calibration. This has important implications for the application of these models in climate change impact studies. The causes of the low transferability to changed climate conditions have, however, only been investigated in a few studies. Here we revisit a study in Austria that demonstrated the inability of a conceptual semi-distributed HBV-type model to simulate the observed discharge response to increases in precipitation and air temperature. The aim of the paper is to shed light on the reasons for these model problems. We set up hypotheses for the possible causes of the mismatch between the observed and simulated changes in discharge and evaluate these using simulations with modifications of the model. In the baseline model, trends of simulated and observed discharge over 1978-2013 differ, on average over all 156 catchments, by 95 +/- 50 mm yr(-1) per 35 years. Accounting for variations in vegetation dynamics, as derived from a satellite-based vegetation index, in the calculation of reference evaporation explains 36 +/- 9 mm yr(-1) per 35 years of the differences between the trends in simulated and observed discharge. Inhomogeneities in the precipitation data, caused by a variable number of stations, explain 39 +/- 26 mm yr(-1) per 35 years of this difference. Extending the calibration period from 5 to 25 years, including annually aggregated discharge data or snow cover data in the objective function, or estimating evaporation with the Penman-Monteith instead of the Blaney-Criddle approach has little influence on the simulated discharge trends (5 mm yr(-1) per 35 years or less). The precipitation data problem highlights the importance of using precipitation data based on a stationary input station network when studying hydrologic changes. The model structure problem with respect to vegetation dynamics is likely relevant for a wide spectrum of regions in a transient climate and has important implications for climate change impact studies.