Scholarlink

<p>Wax deposition in waxy crude oil pipelines reduces the effective flow area, increases pumping pressure, and may ultimately cause blockage or rupture. Accurate prediction of the wax deposition rate is therefore essential for safe operation and cost-effective maintenance planning. However, existing data-driven models are typically trained on very limited experimental datasets and often rely on single learning algorithms, which restricts their generalization ability in complex operating conditions. To address these challenges, we propose an integrated prediction framework that combines data augmentation using a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and a heterogeneous Stacking ensemble model. Starting from 85 experimental samples collected from a crude oil pipeline in the Qinghai oil field, WGAN-GP is trained to generate high-fidelity synthetic data under physicochemical constraints, expanding the training set to 1,060 samples. Comparative analysis against alternative augmentation methods (Gaussian noise, SMOGN, and standard GAN) demonstrates that WGAN-GP achieves superior data quality metrics (coverage ratio: 1.658; NNDR: 0.82) and yields statistically significant improvements in predictive performance (<i>p</i> &lt; 0.05), justifying its additional computational cost. A two-layer Stacking ensemble is then constructed using Random Forest, XGBoost, and Support Vector Regression as base learners and Ridge Regression as the meta-learner. On the held-out test set containing only real samples, the proposed framework achieves a coefficient of determination R² of 0.9632, a mean absolute error (MAE) of 0.4185, and a root-mean-squared error (RMSE) of 0.5272. Compared with the best single model, Support Vector Regression, MAE and RMSE are reduced by 2.58% and 2.35%, respectively, while the MAE is reduced by 10.56% relative to the Random Forest baseline. These results demonstrate that integrating WGAN-GP-based data augmentation with a heterogeneous Stacking ensemble presents a promising framework that can serve as a complementary tool for wax deposition rate prediction and pipeline risk assessment in data-scarce industrial scenarios.</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

A WGAN-GP and stacking ensemble-based framework for enhanced prediction of wax deposition rate in crude oil pipelines

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月17日

查看全文
  • Yiqiong Gao,
  • Zhengshan Luo,
  • Jianbo Guo

摘要

Wax deposition in waxy crude oil pipelines reduces the effective flow area, increases pumping pressure, and may ultimately cause blockage or rupture. Accurate prediction of the wax deposition rate is therefore essential for safe operation and cost-effective maintenance planning. However, existing data-driven models are typically trained on very limited experimental datasets and often rely on single learning algorithms, which restricts their generalization ability in complex operating conditions. To address these challenges, we propose an integrated prediction framework that combines data augmentation using a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and a heterogeneous Stacking ensemble model. Starting from 85 experimental samples collected from a crude oil pipeline in the Qinghai oil field, WGAN-GP is trained to generate high-fidelity synthetic data under physicochemical constraints, expanding the training set to 1,060 samples. Comparative analysis against alternative augmentation methods (Gaussian noise, SMOGN, and standard GAN) demonstrates that WGAN-GP achieves superior data quality metrics (coverage ratio: 1.658; NNDR: 0.82) and yields statistically significant improvements in predictive performance (p < 0.05), justifying its additional computational cost. A two-layer Stacking ensemble is then constructed using Random Forest, XGBoost, and Support Vector Regression as base learners and Ridge Regression as the meta-learner. On the held-out test set containing only real samples, the proposed framework achieves a coefficient of determination R² of 0.9632, a mean absolute error (MAE) of 0.4185, and a root-mean-squared error (RMSE) of 0.5272. Compared with the best single model, Support Vector Regression, MAE and RMSE are reduced by 2.58% and 2.35%, respectively, while the MAE is reduced by 10.56% relative to the Random Forest baseline. These results demonstrate that integrating WGAN-GP-based data augmentation with a heterogeneous Stacking ensemble presents a promising framework that can serve as a complementary tool for wax deposition rate prediction and pipeline risk assessment in data-scarce industrial scenarios.