Noise emission prediction technology for converter stations based on digital intelligent networks

Traditional noise prediction techniques are mostly based on simplified physical models, which make it difficult to characterise the multi-source coupling and time-varying nonlinear noise emission characteristics of converter stations. This paper proposes a noise emission prediction technique for converter stations based on digital intelligent networks, and designs a hybrid deep learning model that combines graph neural networks (GNN) and long short-term memory (LSTM) to explicitly simulate the spatial relationships and temporal dynamics between devices. The experimental results show that a mean absolute error (MAE) of 1.9 dB was achieved in LAeq (equivalent continuous a-weighted sound pressure level) prediction, which is 27% and 21% lower than traditional support vector regression (SVR) and single LSTM models, respectively. In addition, even in complex scenarios such as load fluctuations and data loss of up to 10%, the fluctuation of model error is still less than 0.3 dB, indicating excellent stability of the model.