Decision Support Systems (DSSs) that are used nowadays by water managers often predict states that do not correspond with the observed states. This is caused by changing parameters in the real systems, while the parameters used in the current DSSs are kept at a fixed level or follow a temporal pattern that does not always represents reality. Usually, these parameters are calibrated in an off-line setting, but when utilising in an on-line system there is a significant drift in performance. Therefore, there is a high need to some form of on-line parameter estimation that reduces the differences between the modelled and observed states. The objectives of this study, in order to reduce the differences between the modelled and observed states, read: (1) defining the state-of-the-art knowledge on optimisation of DSSs regarding on-line parameter updating and optimisation techniques for modelling of large-scale river networks; (2) determining whether automatic parameter updating is possible with reasonable results in a twin experiment set-up for different normative scenarios, with respect to parameter identifiability, model bias and model performance; (3) determining whether automatic parameter updating is possible with real measurement data, with respect to the same performance indicators; and (4) determining how much the performance does improve when implementing some form of parameter updating. The first objective is addressed by former studies (e.g.[2],[4],[7],[9]), which have confirmed On-line Parameter Estimation (OPE) can be applied successfully as a tool to decrease model discrepancies. Both the Doesn’t Use Derivatives (DuD) algorithm, [1], and the Shuffled Complex Evolution (SCE) algorithm, [2], have proved to be robust and effective methods for parameter estimation in multiple fields of expertise, e.g. [3],[4],[5],[6],[7]. The DuD algorithm is utilised in this study, since initial model results have illustrated that the high robustness level of the DuD algorithm. The second objective is addressed by constructively up-scaling the amount of calibration parameters by using several scenarios. The optimisation results are analysed extensively regarding the model performance in terms of robustness, effectiveness, efficiency and model bias. Prior to the OPE, an initial model analysis is performed to determine the model sensitivity to parameter perturbations and identifiability and uniqueness of the optimisation parameters. The analyses of the scenarios’ results demonstrate a high level of model performance, in terms of the performance indicators, in a twin experiment set-up. However, coincidentally the bias follows the temporal pattern in model states, which is probably a numerical error induced by the OPE tool. Nonetheless, the level of bias is sufficiently low to neglect this effect. Third objective is addressed by following the same procedure as for the second objective. However now, the observational data is assigned with white noise in order to facilitate upscaling of the twin experiment set-up to field conditions. The analyses of the results illustrate that up-scaling to field condition is very well possible, since the results show high levels of robustness, effectiveness and efficiency while suppressing the model bias. The fourth objective is addressed by implementing OPE in an existing DSS. Assignment of practical real scenarios, like river maintenance programmes, illustrates the necessity of the OPE tool to accurately estimate the correct parameter values, thereby improving the model performance of the original DSS. The transition zone between two parameter values in time, however, is not predicted, as sharp transitions cannot be predicted well as result of the used calibration window with the assumptions of this study. Moreover, local transitions in parameter values are difficult to predict by the OPE tool. Concluding, this study demonstrates that it is essential to use some form of OPE to predict the actual parameter values accurately for highly varied scenarios. This statement is grounded by the high level of performance indicators that have been observed in the results of the OPE tool. The computation time is sufficiently low that it is applicable in real-time systems. However, more research on discretisation of the transition phase, on inclusion of control actions and on other types of additive noise is required before implementing the tool in a real system. Furthermore, the added value to the model performance of using more observation locations and more parameters should be investigated.