In this study, large-eddy simulation (LES) is used to study the evolution of the wind-turbine wakes and their effects on power losses inside an idealized finite-size wind farm in the course of a full diurnal cycle. In the LES, turbulent subgrid-scale stresses are modeled using tuning-free Lagrangian scale-dependent dynamic models, while the turbine-induced forces are parameterized using a dynamic actuator disk model with rotation. The simulation results show a strong effect of atmospheric stability on the wind farm wakes and associated power losses. During the night, the relatively low turbulence intensity of the ambient ABL flow results in a relatively slow rate of entrainment of momentum into the wake and, consequently, a slow wake recovery. In contrast, during the day the positive buoyancy flux and associated turbulence production lead to a relatively high turbulence level in the background ABL flow, which enhances turbulent mixing and wake recovery. As a result, the averaged power deficit in the wind farm is found to increase with increasing thermal stability. In particular for that day, the averaged power deficit increased from 28% under the most convective condition to about 57% under the most stable condition.