Refining the working hypotheses of parameter identification in hydrological modelling: the concept of stochastic calibration

The aim of this study is the introduction of a promising strategy for hydrological calibration, which utilizes synthetic data as drivers to identify model parameters and evaluates the adjusted model structure against the full historical sample. These synthetic time series incorporate the marginal properties and dependence structure of the observed data across multiple time scales. One of the main advantages of this methodology against conventional split-sample approaches is the estimation of more robust and “stable” parameter sets. This is due to the model being trained over a much longer dataset and extended hydroclimatic conditions. Another important advantage of this methodology is that the defined modelling scheme is evaluated against the full set of the observed data, hence the validation data set is also extended. In order to prove that the proposed calibration framework is independent of the chosen model, five lumped hydrological models of varying complexity were used for the testing of the calibration scheme. Initially, a proof-of-concept was employed on a representative catchment and across two time scales, monthly and daily, by using for each scale analysis two hydrological models with alternative parameterization. The proposed calibration technique performed equally well as the classical split-sample scheme at the monthly time scale, whereas it demonstrated slightly lower performance at the daily scale. After proving the functionality of the stochastic calibration for this case study, this strategy was tested against a large set of catchments of the MOPEX database, at the monthly scale, to further reinforce the validity of the recommended methodology. From this large scale analysis it is deduced that the stochastic calibration outperformed the split-sample approach in more than half of the examined cases, regardless of the chosen hydrological model. In addition, stochastic calibration proved to be independent of the model structure’s complexity.