Current lightning predictions are uncertain because they either rely on empirical diagnostic relationships based on the present climate or use coarse-scale climate scenario simulations in which deep convection is parameterized. Previous studies demonstrated that simulations with convection-permitting resolutions (km-scale) improve lightning predictions compared to coarser-grid simulations using convection parameterization for different geographical locations but not over the boreal zone. In this study, lightning simulations with the NASA Unified-Weather Research and Forecasting (NU-WRF) model are evaluated over a 955x540 km2 domain including the Great Slave Lake in Canada for six lightning seasons. The simulations are performed at convection-parameterized (9 km) and convection-permitting (3 km) resolution using the Goddard 4ICE and the Thompson microphysics (MP) schemes. Four lightning indices are evaluated against observations from the Canadian Lightning Detection Network (CLDN), in terms of spatiotemporal frequency distribution, spatial pattern, daily climatology, and an event-based overall skill assessment. Concerning the model configuration, regardless of the spatial resolution, the Thompson scheme is superior to the Goddard 4ICE scheme in predicting the daily climatology but worse in predicting the spatial patterns of lightning occurrence. Several evaluation metrics indicate the benefit of working at a convection-permitting resolution. The relative performance of the different lightning indices depends on the evaluation criteria. Finally, this study demonstrates issues of the models to reproduce the observed spatial pattern of lightning well, which might be related to an insufficient representation of land surface heterogeneity in the study area.