The work addresses a highly relevant issue and provides a thorough comparison of the two methodologies.

It is of high scientific quality and well formulated. I believe it is ready for publishing given a few minor revisions and technical corrections.

The manuscript quantifies the difference in turbine position between mesoscale and LES in section 2.1 but then there is very little discussion on the effect of that change in positions, especially on the internal wakes. If that is not possible the change in positions should be quantified in some other way.

Please clarify l. 356ff.

Technical corrections:

l. 277: "..also cannot also ..."

The dotted black lines in figures 5, 6,7,8,9,11,12 and 14 are very hard to differentiate from solid lines, please consider a wider spacing of the dots.

The authors evaluate the capabilities of mesoscale simulations in predicting the wake effects on wind turbine power production. A comparison between mesoscale Weather Research and Forecasting simulations and large-eddy simulations for three planned wind farms under realistic atmospheric conditions is made. The WRF simulation is found to be capable of capturing the velocity deficit downstream of both single and multiple wind farms. However, the simulations under-estimate the power loss due to internal wakes, but provide accurate estimates of the combined wake losses from internal and cluster wake effects across some wind direction sectors.The findings suggest that mesoscale simulations are a viable tool for predicting wake effects for wind farms. The manuscript is generally well written and the results are of interest to the wind energy community. I recommend its publication with the following comments:

1. Although the wake effects are weaker under unstable conditions, the influences of the horizontal wind fluctuation variances may be stronger. How well does the WRF model parameterize them?

2. In figure 1, what is the relative location of panels (a) and (c) in the mesoscale simulation? Are panels (b) and (d) the full LES domain?

3. The linear regression in Fig. 11 appears to underestimate the slope of the fit (and hence perhaps the performance of the WRF model), likely due to the relatively small number of data points. Kernel methods (Kernel Smoothing, M.P. Wand 1994) in general can greatly improve the regression accuracy. I suggest that the authors give it a try.