Scholarlink

<p>Asphaltene precipitation and deposition pose significant challenges in the oil industry, leading to well blockage, formation damage, impairment of process equipment, and reduced reservoir permeability, which ultimately affect oil production and economic efficiency. Traditional laboratory-based methods for detecting asphaltene precipitation are often costly and time-consuming, emphasizing the need for rapid and reliable predictive techniques. This study develops predictive models for asphaltene precipitation using three white-box machine learning algorithms: Group Method of Data Handling (GMDH), Gene Expression Programming (GEP), and Genetic Programming (GP). The models were trained and validated based on a dataset of 308 laboratory measurements from 25 types of crude oils, considering relevant parameters influencing model performance. Their predictive capability was benchmarked against a thermodynamically consistent cubic-plus-association (CPA) equation of state developed in this work. Statistical evaluations confirmed the reliability of the developed models. Among the correlations, GMDH demonstrated the best performance, achieving the correlation coefficient (R<sup>2</sup>) of 0.858 and a root-mean-square error (RMSE) of 0.171. GMDH delivers quick forecasts for new cases without requiring further tuning, while still accurately predicting asphaltene precipitation. Sensitivity analysis revealed that pressure (P), oil °API gravity (°API), and bubble point pressure (P<sub>b</sub>) exhibit the strongest negative correlations with asphaltene precipitation, indicating that the highest risk occurs during P drops near the P<sub>b</sub> and in heavier crude oils with lower °API values. Finally, Leverage-based outlier analysis confirmed the reliability of the dataset, with 97.080% of data points within the valid region, reinforcing the credibility of both experimental measurements and predictive 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

White-box modeling of asphaltene precipitation during natural depletion of oil reservoirs

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月28日

查看全文
  • Sara Sahebalzamani,
  • Mohsen Mohammadi,
  • Ghazal Piroozi,
  • Fahimeh Hadavimoghaddam,
  • Abdolhossein Hemmati-Sarapardeh

摘要

Asphaltene precipitation and deposition pose significant challenges in the oil industry, leading to well blockage, formation damage, impairment of process equipment, and reduced reservoir permeability, which ultimately affect oil production and economic efficiency. Traditional laboratory-based methods for detecting asphaltene precipitation are often costly and time-consuming, emphasizing the need for rapid and reliable predictive techniques. This study develops predictive models for asphaltene precipitation using three white-box machine learning algorithms: Group Method of Data Handling (GMDH), Gene Expression Programming (GEP), and Genetic Programming (GP). The models were trained and validated based on a dataset of 308 laboratory measurements from 25 types of crude oils, considering relevant parameters influencing model performance. Their predictive capability was benchmarked against a thermodynamically consistent cubic-plus-association (CPA) equation of state developed in this work. Statistical evaluations confirmed the reliability of the developed models. Among the correlations, GMDH demonstrated the best performance, achieving the correlation coefficient (R2) of 0.858 and a root-mean-square error (RMSE) of 0.171. GMDH delivers quick forecasts for new cases without requiring further tuning, while still accurately predicting asphaltene precipitation. Sensitivity analysis revealed that pressure (P), oil °API gravity (°API), and bubble point pressure (Pb) exhibit the strongest negative correlations with asphaltene precipitation, indicating that the highest risk occurs during P drops near the Pb and in heavier crude oils with lower °API values. Finally, Leverage-based outlier analysis confirmed the reliability of the dataset, with 97.080% of data points within the valid region, reinforcing the credibility of both experimental measurements and predictive models.