Research On Short-term Forecasting Method Of Ship Motion Based On Deep Learning

Under the influence of environmental factors such as wind,waves and currents in the ma rine environment,the ship will continuously generate six degrees of freedom of swaying moti on during offshore operations.Ship swaying caused by the waves is an unfavorable factor affe cting the safety and efficiency of offshore operations.Accurate and reliable forecasting of shi p motion posture in the next few seconds in real time is the key to improving work safety and improving work efficiency.Under medium and high sea conditions,ship swaying motion has nonlinear characteristics.Therefore,the development of nonlinear ship motion real-time forec asting model has practical significance.The Long-Short Term Memory(LSTM)deep learning network provides a powerful tool for real-time prediction of nonlinear ship motions because of its unique advantages in processing nonlinear time series.Based on the deep learning tech nique,this paper studies the extremely short-term forecasting method of ship motion using the ship motion time history data generated by numerical simulation of GN theory.This paper first analyzes and compares the principles of different neural network models including deep feedforward network,cyclic neural network,long-term and short-term memory network and gated loop network model.The long-short-term network model is verified under Tensorflow and its upper framework Keras.The advantages and careful analysis of the calculation strategy selection in the LSTM model construction,and a set of effective network model design schemes are given,and the results are verified by calculation.In order to meet the practical engineering application,this paper focuses on the input vector sc aling problem of the LSTM network model,and analyzes and verifies the deficiencies of the t raditional method.The innovative method is based on the autocorrelation function,and the cal culation is verified.Its rationality.In terms of the impact of data preprocessing on model prediction results,this paper is bas ed on the data normalization method to adapt to the actual training needs in the LSTM networ k layer,and on the other hand the EMD deeomposition processing method to improve the data non-stationarity to model prediction.The impact of the results makes it perform better in the application environment under medium and high sea conditions.Finally,the implementation method of the mixed sea state forecasting model is given and verified by calculation.