Scholarlink

<p>Centrifugal compressors are commonly employed for pressure compensation in long-distance natural gas pipelines. Accurate prediction of compressor performance is essential for safety and economic considerations. However, due to the invariance of process requirements over time, operational data is usually insufficient to establish a comprehensive model that covers varying operating conditions. This paper presents a prediction method based on the long short-term memory (LSTM) network and transfer learning technique, integrating similarity laws to expand the feature domain and address the issue of limited data. In the case study, the constructed TransLSTM model, a fusion of transfer learning and LSTM, is compared with pure LSTM, Gaussian Process (GP), Gradient Boosting Decision Tree (GBDT), Random Forest (RF), and Support Vector Machine (SVM) models. Results indicate that TransLSTM achieves the highest prediction accuracy, with deviations from actual values of only 2.9% for power and 0.971% for compressor ratio. Verification through comparisons between the training and test sets demonstrates its excellent generalization ability, which is particularly crucial when operational data spans a limited range of working conditions. Additionally, its stability is significantly superior to that of benchmark models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Performance prediction of the centrifugal compressor using a long short-term memory-based model with transfer learning and similarity conversion

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月9日

查看全文
  • Xia Li,
  • Xiao Chen,
  • Chuanming Yang

摘要

Centrifugal compressors are commonly employed for pressure compensation in long-distance natural gas pipelines. Accurate prediction of compressor performance is essential for safety and economic considerations. However, due to the invariance of process requirements over time, operational data is usually insufficient to establish a comprehensive model that covers varying operating conditions. This paper presents a prediction method based on the long short-term memory (LSTM) network and transfer learning technique, integrating similarity laws to expand the feature domain and address the issue of limited data. In the case study, the constructed TransLSTM model, a fusion of transfer learning and LSTM, is compared with pure LSTM, Gaussian Process (GP), Gradient Boosting Decision Tree (GBDT), Random Forest (RF), and Support Vector Machine (SVM) models. Results indicate that TransLSTM achieves the highest prediction accuracy, with deviations from actual values of only 2.9% for power and 0.971% for compressor ratio. Verification through comparisons between the training and test sets demonstrates its excellent generalization ability, which is particularly crucial when operational data spans a limited range of working conditions. Additionally, its stability is significantly superior to that of benchmark models.