Wind Turbine Noise Reduction from Seismological Data

Over the last years, installations of wind turbines (WTs) increased worldwide. Owing to
negative effects on humans, WTs are often installed in areas with low population density.
Because of low anthropogenic noise, these areas are also well suited for sites of
seismological stations. As a consequence, WTs are often installed in the same areas as
seismological stations. By comparing the noise in recorded data before and after
installation of WTs, seismologists noticed a substantial worsening of station quality leading
to conflicts between the operators of WTs and earthquake services.

In this study, we compare different techniques to reduce or eliminate the disturbing signal
from WTs at seismological stations. For this purpose, we selected a seismological station
that shows a significant correlation between the power spectral density and the hourly
windspeed measurements. Usually, spectral filtering is used to suppress noise in seismic
data processing. However, this approach is not effective when noise and signal have
overlapping frequency bands which is the case for WT noise. As a first method, we applied
the continuous wavelet transform (CWT) on our data to obtain a time-scale representation.
From this representation, we estimated a noise threshold function (Langston & Mousavi,
2019) either from noise before the theoretical P-arrival (pre-noise) or using a noise signal
from the past with similar ground velocity conditions at the surrounding WTs. Therefore, we
installed low cost seismometers at the surrounding WTs to find similar signals at each WT.
From these similar signals, we obtain a noise model at the seismological station, which is
used to estimate the threshold function. As a second method, we used a denoising
autoencoder (DAE) that learns mapping functions to distinguish between noise and signal
(Zhu et al., 2019).

In our tests, the threshold function performs well when the event is visible in the raw or
spectral filtered data, but it fails when WT noise dominates and the event is hidden. In
these cases, the DAE removes the WT noise from the data. However, the DAE must be
trained with typical noise samples and high signal-to-noise ratio events to distinguish
between signal and interfering noise. Using the threshold function and pre-noise can be
applied immediately on real-time data and has a low computational cost. Using a noise
model from our prerecorded database at the seismological station does not improve the
result and it is more time consuming to find similar ground velocity conditions at the
surrounding WTs.