基于多平台训练的海洋平台极短期运动预报研究

在风浪流等环境条件的共同作用下,浮式海洋平台会在六自由度方向上进行摇荡运动,进而对海上作业安全构成了严峻的威胁。准确的运动极短期预报,可以作为输入条件,提高运动补偿装置的性能;另一方面也可以提供及时的实时预警信息,指导安全作业。深度学习算法是指模型通过对现有的数据进行学习,在大量的训练后使得其能够提取到数据的特征,进而能够根据输入数据对未来进行预测。通过对若干海洋平台的模型试验数据进行学习,建立了基于长短期记忆神经网络（LSTM）的深度学习模型。模型实现了对未来20~40 s内的垂荡和纵荡运动的精确预报,预报精度总体可达到80%~90%以上,并以此对模型的输入、输出窗口长度以及波浪相位差开展了敏感性研究。通过多平台混合训练得到了输入、输出窗口长度以及波浪相位差三者间的合适比例关系,并以此为基础拓展了预报时间长度,为深度神经网络模型给出了推荐的构型参考。

Research on short-term prediction of offshore platforms based on multi-platform training

Under the combined influence of environmental factors such as wind, waves and currents, floating offshore platforms undergo oscillatory motion in six degrees of freedom, posing a severe challenge to safety of offshore operations. Accurate short-term motion prediction can serve as an input condition to improve the performance of motion compensation devices. Additionally, it provides timely real-time warning information to guide safe operations. Deep learning algorithms involve models learning from existing data through extensive training to extract features and make predictions based on input data. This study leverages model trial data from various ocean platforms to establish a deep learning model based on Long Short-Term Memory (LSTM) networks. The model achieves accurate prediction of heave and surge motions within the next 20~40 s, with an overall accuracy exceeding 80%-90%.Sensitivity analysis explores the model’s input/output window lengths and wave phase differences. The appropriate proportionality relationship between multiple platform mixed training is determined, extending the forecast time length based on this foundation and providing a recommended configuration reference for deep neural network models.