An EMD-LSTM-SVR model for the short-term roll and sway predictions of semi-submersible

Abstract

A reliable short-term prediction method is of significance in ensuring safety and reducing operation time. This paper proposes a hybrid empirical mode decomposition (EMD) model for the short-term prediction of motions of marine structures. The combination of long short term memory (LSTM) network and support vector regression (SVR) model offers an accurate prediction on the subcomponents of EMD process. The training and test data were provided by the model tests of a semi-submersible with the taut, catenary and tension leg mooring systems. The predicted results of present EMD-LSTM-SVR model were compared with those of the autoregressive (AR) model and the EMD-SVR model. The influence of the boundary effect, spectrum bandwidth and non-stationarity on the predicted results were investigated. The results of contrast experiments demonstrated that the proposed EMD-LSTM-SVR model obtains better predicted results than the other two models in most cases. The prediction accuracy has a negative correlation with the broad-banded spectrum and the strong non-stationarity. The EMD technique is beneficial for dealing with the broad spectral and non-stationary motion time series to reduce the prediction errors.