Lumped parameter modeling in karst hydrology has been widely developed in the last decades. Model conceptualization often leads to one unique model structure and skip the assessment of potential model structure equifinality. Some studies consider various model structures to account for this source of uncertainties in hydrological predictions but still do not assess the equifinality between these various model structures. This issue is particularly important for karst hydrology, as such hydrological systems are highly heterogenous and information about their structure is difficult to obtain. In this work, we developed a Bayesian Combined Model Averaging (BCMA), which allows estimating a model structure probability distribution associated with model parameters distribution. This approach constitutes a significant step forward compared with classical calibration, as it permits (i) providing a robust hydrological predictor which accounts for both structural and parametric uncertainties, and (ii) avoiding epistemic error related to model structure choice, which is generally influenced by the subjective conceptualization of the karst hydrological system by the modeler and (ii) improve estimation of predictive uncertainties. The proposed methodology is illustrated taking as an example the simulation of Fontaine de Vaucluse karst spring discharge (southern France).