摘要针对海洋浮式平台系泊系统在复杂多变的作业环境下易受各种线性、非线性作用力时存在的预测问题，实现对基于长短周期记忆(Long Short Term Memory,LSTM)单模型预测网络隐藏层数、迭代次数和学习速率的优化，提出搭建具有可变卷积和小波基激活函数的多层特征提取特性和变阈值残差收缩预测功能的(Residual shrinkage network,Resnet-LSTM)混合预测模型。对平台运动响应下多点系泊系统整体受力进行非线性映射分析、计算载荷系数并搭建多点系泊缆模型求解风浪流联合作用下的系泊缆张力值，使用LSTM单模型、Resnet-LSTM混合模型和改进混合模型对系泊力仿真数据集进行训练预测。结果显示：采用Resnet-LSTM混合模型预测准确度得以提高，网络预测准确度可高达0.9973，各项网络参数和预测指标得以优化，使用可变卷积改进的Resnet-LSTM预测效果优于未改进模型。证明基于Resnet-LSTM改进混合预测模型应用在多点系泊系统张力非线性时序特征预测应用方面具有提升网络性能的作用。
Abstract：To solve the tension prediction problem when the structure of the ocean floating platform mooring system was subjected to various linear and nonlinear forces in the complex and changeable operating environment, for improving the hidden layers, iteration times and learning rate parameters of the LSTM model based on long short term memory algorithm, the Resnet-LSTM model with convolution muti-layer features extraction inserted different thresholds’ residual module was established. The floating platform was regarded as an overall force loaded on the mooring lines to analysis nonlinear feature while influential factors of mooring lines tension were considered and the environment loads factors were calculated. After the muti-points mooring system was setup and solved under different ocean conditions, trained and predicted the lines tension data which came from datasets with three different models LSTM, Resnet-LSTM and improved Resnet-LSTM. Experiment comparative results showed that the Resnet-LSTM indicators of accuracy was up 0.9973, the parameters of network were optimized which proved the model based on improved Resnet-LSTM could effectively increase the prediction result of the model in nonlinear data processing of multi-points mooring lines system tension with some improvement value.
2024, Vol. 45 
