Wind turbine noise (WTN) assessment is a critical issue for environmental impact studies and regulatory compliance, particularly for residents near wind farms. This paper investigates the use of convolutional recurrent neural networks (CRNNs) to extract the wind turbine noise from the total noise. A dedicated learning dataset is constructed by combining measured background noise in the environment with simulated WTN signals, yielding representative total acoustic noise time series. The simulations rely on a physics-based framework using an aeroacoustics source and an outdoor sound propagation model to build a site-specific time series of noise levels in 1/3-octave bands, accounting for wind speed, wind direction, turbine operating modes, and day/night atmospheric conditions. The dataset covers multiple wind farms and receiver locations, with wind directions selected based on long-term meteorological statistics. The proposed CRNN-based approach is trained to capture local patterns and the temporal dynamics of WTN, with the objective of accurately estimate overall sound pressure level (OASPL) of WTN in various acoustic conditions.
Rkhiss et al. (Wed,) studied this question.