基于WOA-BRANN的可燃气体预混燃爆试验结果预测方法  

Simulation of deflagration test of premixed combustible gas with WOA-BRANN

作　　者：王亮[1] 于佳航 张彬[2] 王博乔 夏远辰 WANG Liang;YU Jiahang;ZHANG Bin;WANG Boqiao;XIA Yuanchen(SINOPEC Qingdao Liquefied Natural Gas Co.,Ltd.,Qingdao 266400,China;Marine Engineering College,Dalian Maritime University,Dalian 116026,China)

机构地区：[1]中国石化青岛液化天然气有限责任公司,山东青岛266400 [2]大连海事大学轮机工程学院,辽宁大连116026

出　　处：《中国航海》2023年第1期55-65,共11页Navigation of China

基　　金：辽宁省自然科学基金(2020JH/10300107);国家自然科学基金(51306026)。

摘　　要：针对置障条件下可燃气体预混燃爆试验操作难度大、成本高和有效试验数据难以获取的问题,提出一种基于鲸群算法优化下的人工神经网络(Whale Optimization Algorithm Bayesian-Regularization Artificial Neural Network, WOA-BRANN)模型对试验结果进行预测,以试验获取的火焰速度和最大燃爆压力数据作为特征样本数据进行训练和测试。在较少试验数据的情况下,模型开发前需结合试验结果的机理分析并观察数据分布,将机理差异较大工况下的80%的数据划入训练集,以提高模型预测的准确性。结合敏感性分析实现模型隐含层神经元的最优配置,采用决定系数(R-Square,R2)评价指标评估模型预测性能,并建立响应面模型(Reflective Shadow Maps, RSM)和误差反向传播(Error Back Propagation, BP)模型进行对比验证。结果表明:最优结构下的贝叶斯正则化人工神经网络(Bayesian-Regularization Artificial Neural Network, BRANN)模型相比BP模型的火焰速度和压力预测数据R2值均提高了15%,相比RSM模型火焰速度和压力预测数据分别提高了34%和29%;BRANN模型相比RSM模型和BP模型的R2值震荡幅度至少降低50%,模型鲁棒性更好,验证BRANN模型在较少数据驱动情况下的适用性。针对预测结果存在的问题,在原有BRANN模型基础上做出改进,采用鲸群优化算法(Whale Optimization Algorithm, WOA)对BRANN模型的学习率进行优化,进一步提高模型的预测精度和鲁棒性。WOA-BRANN模型试验结果预测方法可为预混气体燃爆机理研究提供大量可靠试验预测结果,降低试验成本。The deflagration test of premixed combustible gas under conditions of obstacles is difficult and costly.In view of the issue,a test result prediction algorithm based on whale optimization algorithm and Bayesian regularization artificial neural network is introduced.The model is built jointly by mechanism analysis and real data training.Since the volume of available data is limited,the data for training is screened first,only the data significantly deviating from calculation results(about 80%of the total data set)is selected.The configuration of hidden layers of neuron elements are optimized based on sensitivity analysis.The coefficient of determination(R^(2))is used to estimate the performance of the prediction.A response surface model and a BP model are constructed for comparison.Experiments show that R^(2)of the flame velocity and pressure values from the optimized Bayesian regularization neural networks both improved 15%compared to those from BP network.They are 34% and 29% better compared to those from RSM model respectively.The fluctuation of R’decreased more than 50%.The introduction of the whale optimization algorithm further improves the robustness of the model.

关 键 词：燃爆试验 贝叶斯正则化 最大燃爆压力 火焰速度 预测方法 

分 类 号：U698.6[交通运输工程—港口、海岸及近海工程] TP206[交通运输工程—船舶与海洋工程]