Deep learning based short-term motion prediction of floating wind turbine under shutdown condition

Abstract

It has become an inevitable trend for the application of the wind power generation devices to shift from land to ocean. The safety of floating offshore wind turbine (FOWT) platforms is closely related to their motion response. The motion response of platforms will be affected by various factors, such as waves, wind, and currents. Predicting the short-term motion response of a platform with high precision and utilizing the predicted values for advanced control has always been a significant concern. At present, the development of deep learning technology has brought some potential solutions to this problem. In this paper, we proposed a method combining convolutional neural network (CNN) and gated recurrent unit (GRU) to predict the short-term motion response of the platform. Specifically, a series of numerical simulation of the 5MW OC4 semi-submersible FOWT under shutdown conditions are carried out, and the wave elevation and platform motion data are obtained based on the simulation data. The simulation data is then divided into the training set, validation set and test set. In this paper, we combine the Bayesian optimization algorithm to adaptively select the optimal hyper-parameters of the deep learning model and use the optimal hyper-parameters to establish the optimal CNN-GRU model for training. By comparing the prediction results of GRU, long short-term memory (LSTM), CNN, and CNN-LSTM models, we found that the CNN-GRU hybrid model achieves the highest prediction accuracy on different datasets. More importantly, when comparing the prediction results with the real values, it is found that the CNN-GRU hybrid model has strong non-linear expression ability for prediction duration of 2 s, 4 s, 6 s, and 8 s. The prediction accuracy is negatively correlated with the prediction duration. Finally, it is proved that the CNN-GRU model has significant advantages in terms of prediction accuracy and training time. We verify the CNN-GRU model using the dataset under combined wind-wave action, proving that the model is transferable and suitable for complex working conditions. This provides an idea for optimizing the motion response prediction model of FOWT.