摘要
为提升压缩机状态监测精度与运维智能化水平,以某天然气管道PCL803型离心压缩机为例,提出基于贝叶斯优化反向误差传播神经网络(BOA-BPNN)的离心压缩机性能曲线预测方法,并考察所建模型对多种运行工况和不同运行时间下性能曲线实时预测的准确性。该方法以压缩机实际运行数据与出厂性能曲线作为循环训练基础,通过贝叶斯优化算法(BOA)对BPNN进行超参数优化搜索。研究结果表明:在BOA-BPNN模型、PSO-BPNN模型和相似换算3种方法中,BOA-BPNN模型的预测精度最高,预测多变能头与多变效率曲线的平均相对误差分别为0.242%、0.025%,均方根误差分别为0.056、0.028,决定系数分别为0.996、0.997;在不同工况下,BOA-BPNN模型预测结果的误差保持相对稳定,预测多变能头与多变效率曲线的平均相对误差均未发生明显变化,预测能头曲线的平均相对误差均为0.270%,预测效率曲线的平均相对误差分别为0.112%、0.082%;BOA-BPNN模型可以准确预测不同压缩机的实时性能曲线,预测结果的平均相对误差均维持在1%以下,均方根误差维持在0.07以下,决定系数维持在0.99以上。研究结果有助于压缩机运行方案的制定,也可为压缩机性能评估提供参考。
To enhance the status monitoring accuracy and operation and maintenance intelligence level of compressors,taking PCL803 centrifugal compressor in a natural gas pipeline as an example,a centrifugal compressor performance curve prediction method based on Bayesian optimization algorithm-backpropagation neural network(BOA-BPNN)was proposed,and the accuracy of the model in real-time prediction of performance curves under various operating conditions and different operating times was investigated.In this method,the actual operation data and the factory performance curve of the compressor are used as the basis for cyclic training,and BOA is used to search for optimal hyperparameters of BPNN.The research results show that among the three methods of BOA-BPNN model,PSO-BPNN model and similarity conversion,the BOA-BPNN model has the highest prediction accuracy,with average relative errors of 0.242% and 0.025%,root mean square errors of 0.056 and 0.028 and determination coefficients of 0.996 and 0.997 for predicting polytropic energy head and polytropic efficiency curves,respectively.Under different operating conditions,the prediction errors of the BOA-BPNN model remain relatively stable,the average relative errors for predicting polytropic energy head and polytropic efficiency curves show no significant change,the average relative error for predicting polytropic energy head curve is 0.270%,while the average relative errors for predicting polytropic efficiency curves are 0.112% and 0.082%,respectively.The BOA-BPNN model can accurately predict real-time performance curves of different compressors,and the average relative error,root mean square error and determination coefficient of the prediction results is maintained below 1%,below 0.07 and above 0.99,respectively.The study results are helpful for the formulation of compressor operation scheme,and provide reference for the evaluation of compressor performance.
作者
朱汪友
刘家豪
杨君明
刘震
侯磊
Zhu Wangyou;Liu Jiahao;Yang Junming;Liu Zhen;Hou Lei(College of Mechanical and Transportation Engineering,China University of Petroleum(Beijing);PipeChina Beijing Pipeline Company)
出处
《石油机械》
北大核心
2025年第12期1-9,共9页
China Petroleum Machinery
基金
国家管网集团北方管道有限责任公司与中国石油大学(北京)合作项目“油气站场转动设备健康状态评价与预测技术研究”(GWHT20230020030)。
关键词
离心压缩机
性能曲线
贝叶斯优化
误差传播
神经网络
多变能头
超参数
centrifugal compressor
performance curve
Bayesian optimization
error propagation
neural network
polytropic energy head
hyperparameter
