Trading Space for Time: A non-stationary approach to probabilistic flow predictions in a changing climate

Climate change will significantly affect hydrological variables in many parts of the world. Hydrological models are necessary to translate potential climate forcing into hydrological variables of interest to water management. Such models are generally parameterized through calibration on historical data or through a priori parameter estimates. The former is likely to be dependent on the climatic regime under which the historical observations were collected, while the latter is unlikely to provide reliable predictions in many cases. In both cases, the parameters are generally taken as time invariant. In this study, we try to overcome two limitations of the current approach: the absence of uncertainty estimates in predictions and the assumption that the best performing parameter sets remain unchanged under nonstationarity. We use regionalization to understand climatic and physical controls on relevant watershed response characteristics. This spatial variability enables trading-space-for-time to assess how conditioning of model parameters and hence model predictions would change under a changing climate. We show in a US wide study how such an approach can lead to drastically different results than what would be obtained from the traditional methodology in which parameters are kept constant.