基于循环神经网络的海冰弯曲强度预测分析

海冰弯曲强度是确定锥体海洋平台、船舶和港口码头等斜面工程结构冰载荷的重要参数，是海冰力学性质的主要研究内容。本文首先对2009—2010年的渤海海冰三点弯曲试验数据进行总结，分析了温度、盐度、密度及孔隙率对海冰弯曲强度的影响。基于现场试验数据建立了海冰弯曲强度的循环神经网络（recurrent neural network，RNN）预测模型，并采用随机梯度下降法和Adam优化算法对RNN预测模型进行了优化。本预测模型将海冰温度、盐度、加载速率和密度等参数作为输入，将海冰弯曲强度作为输出。在建模过程中，数据被随机排序和归一化处理，并采用评价指标评估其模型的准确性，对RNN预测模型进行优化。将RNN模型对海冰弯曲强度的预测结果与试验经验公式结果进行了对比。预测结果表明，RNN模型对海冰弯曲强度的预测精度显著优于试验经验公式分析结果。该预测模型能有效解决海冰弯曲强度与其影响因素间的复杂的非线性关系，可为海洋工程结构的抗冰设计提供参考依据。

PREDICTION AND ANALYSIS OF FLEXURAL STRENGTH OF SEA ICE BASED ON RECURRENT NEURAL NETWORK

As a mechanical property of sea ice, the flexural strength is an important parameter for determining the ice load of inclined engineering structures, such as conical offshore platforms, ships, ports and wharves. In the paper, the three-point bending test data of Bohai Sea ice are summarized from the year of 2009 to 2010. The effects of temperature, salinity, loading rate and density on the flexural strength of sea ice are analyzed. Then, the RNN (recurrent neural network) model for predicting the flexural strength of sea ice is established based on the test data, and the model is optimized by the random gradient descent method and Adam optimization algorithm. The parameters such as temperature, salinity, loading rate and density of sea ice are taken as the input and the flexural strength is taken as the output. The data were sorted randomly and normalized, and the accuracy of the model is evaluated by using evaluation indexes. The predicted results are compared with the empirical formula from experiments. The results show that the prediction accuracy using the RNN model is significantly better than that of the empirical formula. This model can effectively solve the complex nonlinear relationship between the flexural strength of sea ice and its influencing factors, and provide a reference for the anti-ice structure design in offshore engineering.