Transmission Line Audible Noise Prediction Based on CNN-BiLSTM-Attention Method

Abstract

High-voltage transmission lines’ audible noise parameters are impacted by a variety of multidimensional elements. in order to better the accuracy of audible noise prediction and effectively utilize the time-series properties in the observed audible noise data. In this paper, we propose a combined model of a convolutional neural network (CNN) and a bidirectional long and short-term memory network (BiLSTM)-based attention mechanism based on feature filtering for transmission line audible noise prediction. Firstly, using the transmission line real-world audible noise data as the dataset, the multidimensional factor time series parameters are optimally filtered, and high correlation feature vectors are extracted by using CNN. Secondly, the extracted feature vectors are fed into the BiLSTM for training and prediction, and the prediction performance is further improved by introducing an attention mechanism at the BiLSTM end so that the model focuses on learning more important data features. Finally, the prediction analysis using actual recorded audible noise data from a 500 kV AC transmission line in Sichuan Province demonstrates that the combined CNN-BiLSTM-Attention model suggested in this paper has a higher prediction accuracy than the BiLSTM, CNN-BiLSTM, and BiLSTM-Attention models.