One of the main challenges presented by a limited area model ensemble prediction system (LAMEPS) concerns the limited capacity for its initial condition (IC) perturbations to correctly represent large-scale flow uncertainties due to its limited-size domain, deficiencies in formulating lateral boundary conditions and inadequate availability of observations. In addition, a mismatch between LAMEPS IC and host EPS lateral boundary perturbations can form spurious waves at the boundaries which spread through the domain, induce noise to the meteorological fields and render inoperative frequent assimilation cycles. In the present work, an ensemble Jk blending method is proposed for improving representation of large-scale uncertainties and for addressing consistent initial conditions and lateral boundary perturbations. Our approach involves employing Jk blending within a framework of 3-dimensional variational (3D-Var) ensemble data assimilation (EDA). In such a system, small-scale perturbations are generated from 3D-Var EDA, while large-scale perturbations are generated from the host ensemble via Jk blending. We hypothesize that final analyses are optimal, and contain perturbed small and large scales which are, at the same time, consistent with each other and with perturbations coming from lateral boundaries. The ensemble Jk method is implemented to the C-LAEF (Convection-permitting Limited-Area Ensemble Forecasting) system and is compared to the standard perturbedobservation EDA approach, i.e., perturbed-observation EDA without large-scale constraint. The comparison shows that the ensemble Jk method gives a more skillful and reliable EPS, especially for the upper-air variables. In addition, positive effects on the surface pressure and precipitation of large-scale perturbations are shown. The ensemble Jk method’s capacity to alleviate perturbation mismatches is also assessed. Additionally, two readily available techniques, i.e., neighborhood and lagging, to improve C-LAEF’s IC perturbation sampling of the initial uncertainties and to address the problem of relatively low model effective resolution are evaluated. Both of them show significantly positive impact on ensemble forecast quality and on detection of extreme weather events.