Prediction model for newly-added sensors to ocean buoys: Leveraging adversarial loss and deep residual LSTM architecture

Abstract

Adding new sensors to ocean buoys can extend their measurement range. However, due to the lack of historical data, the prediction model for the newly-added sensors suffers from the difficulty of training. Traditional pre-training methods simply use the same type of labelled data from other sea areas as pre-training data, failing to take into account the distributional differences between the features of data from different regions, so that the pre-training effect still has a large space to rise. Aiming at the above problems, this paper proposes a prediction model for newly-added sensors to ocean buoys based on adversarial loss and depth residual Long Short Term Memory. Firstly, this paper constructs a prediction model based on deep residual Long Short Term Memory. Then, the pre-training effect is improved by introducing adversarial loss in the loss function of the pre-training task. Finally, the model performance is validated on buoy monitoring data in the nearshore waters of Beihai City, Guangxi Province. The results showed that the pre-training effect was significantly improved with the introduction of adversarial loss compared to the traditional pre-training method.