**5.1 Computed spectra**

As important as the percentage of accurate predictions is to state that not only the levels in scale A are predicted in a reasonably adjusted way, but particularly that


**Table 4.**

*Quality of results according to distance to the wind turbine.*


**Table 5.**

*Quality of results according to wind velocity at the hub height.*

*Prediction of Environmental Sound Pressure Levels from Wind Farms: A Simple but Accurate… DOI: http://dx.doi.org/10.5772/intechopen.103159*


#### **Table 6.**

*Quality of results according to distance to the wind turbine and wind velocity at the hub height.*

the spectra obtained with the proposed model are also rightly adjusted to the measured ones [10].

**Figure 5** shows some results for short and long distances. The blue bar is the measured sound pressure level; the pink bar is the computed sound pressure level using ISO 9613-2 with attenuation only due to geometric divergence and atmospheric absorption; and the dark red bar is the result of our prediction proposal. As it can be seen, our model achieves a good performance as a prediction tool.

#### **5.2 Comparison with a nonnegative matrix factorization (NMF) estimation**

It is not usual to find references that use a variable depletion law according to the frequency.

We compared our attenuation curves with those presented in a paper published in 2021 [18]. The authors estimate the sound pressure level related to wind turbines with a nonnegative matrix factorization (NMF), a machine learning technique.

They present some attenuation filters in third-octave bands from 31.3 to 2000 Hz, for attenuation only and for attenuation considering three kinds of residual noise designed by the authors with basis on real noise samples. The filters were published in graphic format for three distances: 500, 1000, and 1500 m. We read the graphs and compared the attenuations proposed in [18] with the attenuation achieved for our prediction model in the same frequencies range.

The comparison was done using the Wilcoxon's test for differences between pairs. H0 was the equivalence of the compared curves; accepting H0 at 95% confidence means that our attenuation curves are equivalent to those from [18]. Test results are summarized in **Table 7**. We conclude that each one of our attenuation curves reasonably fit at least one case of the filters suggested by [18], the filter with residual noise 1 being the most similar to our proposal.
