基于RAdam Bi-LSTM的LNG动力船舶上甲板储罐泄漏后果预测方法
Consequence prediction of deck tank leakage on LNG powered ship using RAdam Bi-LSTM
-
摘要: 针对液化天然气(Liquid Natural Gas, LNG)动力船舶上甲板储罐泄漏后果预测难度大、预测时间慢和预测成本高等问题,提出一种基于修正自适应矩估计(Rectified Adaptive Moment Estimation, RAdam)优化算法的双向长短期记忆模型循环神经网络(Bi-Directional Long Short-Term Memory, Bi-LSTM)对泄漏后果进行预测。利用FLACS软件对LNG动力船舶上甲板储罐泄漏过程进行数值模拟,并将数值模拟结果作为神经网络的数据集,使用决定系数(R-Square,R2)作为评价预测性能指标。为提高Bi-LSTM网络模型的预测精度和适应性,分别对其激活函数修正线性单元(Rectified Linear Unit, Relu)、Sigmoid、Tanh与优化器RAdam、自适应矩估计(Adaptive Moment Estimation, Adam)和随机梯度下降(Stochastic Gradient Descent, SGD)进行对比分析计算,发现基于Relu激活函数的RAdam Bi-LSTM网络模型的R2均值可达到0.97。为验证Bi-LSTM网络模型的优越性,对循环神经网络(Recurrent Neural Networks, RNN)、长短期记忆模型循环神经网络(Long Short-Term Memory, LSTM)和Bi-LSTM的预测结果进行对比,发现Bi-LSTM网络模型的R2较其他两个模型分别提高4.5%和1.5%,确定使用Bi-LSTM作为所提出的预测方法的网络模型。因此,基于Relu激活函数的RAdam Bi-LSTM网络模型为所提出预测模型中的最优模型,可作为LNG储罐泄漏后果的快速预测方法,以解决事故后果预测速度的问题。Abstract: To Predict the consequences of deck storage tank leakage on a LNG powered ship is difficult, time consuming and costly. The RAdam-LSTM is introduced to handle the problem. A set of data for network training is produced by means of software FLACS. The predictions are evaluated with their determination coefficient. The parameters with different Bi-LSTM networks are calculated and compared, including the following parameters: rectified linear unit(Relu), sigmoid, tanh, Adam, and SGD. Based on the calculation and comparison, the most accurate and suitable Bi-LSTM network is selected. The advantage of the Relu activated Bi-LSTM model is verified through comparing it to RNN, ordinary LSTM and ordinary Bi-LSTM. Experiment shows that the R2 achieved by the Bi-LSTM network is 4.5% and 1.5% higher than that from the other two models, respectively.