Analysis And Prediction Of Semi-Submersible Platform Motion Based On Deep Learning

The Semi-submersible platform has gradually become the main force of deepwater operation due to its advantages such as stable operating state,good mobile performance and large operating depth,and has been widely applied in deepwater oil and gas development.Because of semi submersible platforms floating on the sea,under the action of wind,wave and current will produce certain movements,and large range of motion will affect normal drilling and production platform,therefore,how to accurately forecast semi-submersible platform motion response under wave action is becoming a hot spot of Marine engineering and academic research.At present,the rapid development of deep learning makes it achieve good results in the field of regression prediction,which provides a new idea for the efficient prediction of the motion response of offshore platforms under extreme environmental conditions.A method based on EMD-LSTM model is proposed to predict the short-term platform motion.This method takes the semi-submersible platform model test data as the research object.First of all,the platform motion response of preprocessing of time series and then using empirical mode decomposition(EMD)is decomposed into relatively stable component.The advantage of LSTM is to deal with complicated nonlinear time series,which uses for time series prediction,finally simulation,with traditional LSTM model and the EMD-BP model compared with,the simulation results show that the platform very short term prediction method based on EMD-LSTM model is high precision.The prediction model and method are further optimized with the addition of wave and wind load,which greatly improves the accuracy of prediction and realizes on-line continuous prediction.According to the potential flow theory,the motion response of the platform is mainly affected by the wave excitation force,wind load,current load and the restoring force provided by the mooring system.Therefore,the time history data of the incident wave is selected as the input and the platform motion response as the output.In terms of data processing,on the one hand,based on the data normalization method to adapt to the actual training requirements in the network layer;on the other hand,it uses decomposition algorithm to improve the influence of data non-stationarity on the model prediction results,so as to make it perform better in the application environment.At the same time,the deep feed forward network is used to train and simulate the test data.Based on the frequency domain analysis,the spectrum diagram of the predicted platform motion response is consistent with the actual value trend,and there is a small gap between the peak value.From the analysis of statistical value,the error of the predicted meaningful value,extreme value and actual value of the platform motion response is within 10%.