Scholarlink

<p>Given the increasing demand for efficient technologies in heavy-oil recovery and upgrading, methods that can reduce viscosity and improve oil mobility are of considerable interest. Heavy oils are difficult to produce and process because of their high viscosity and density, as well as their elevated contents of asphaltenes, heteroatoms, and heavy metals. Although thermal and chemical recovery methods can reduce viscosity during production, the oil may regain viscosity after reaching the surface, often requiring diluents or additional upgrading steps. In this context, nanocatalyst-assisted in-situ upgrading has attracted attention as a route to simultaneously improve oil mobility and quality. In this study, a data-driven framework was developed to predict viscosity reduction during NiO nanoparticle-assisted in-situ upgrading. Predictive performance was evaluated using fivefold out-of-fold cross-validation (OOF-CV) to obtain a more reliable estimate of generalization. Conventional models (MLP, RBF, and ANFIS) were compared with modern tree-based ensembles, including Random Forest and Gradient Boosted Decision Trees (GBDT). GBDT delivered the best cross-validated performance, with a pooled OOF <InlineEquation ID="IEq1"><EquationSource Format="TEX">\({R}^{2}\)</EquationSource></InlineEquation> of 0.925 and a normalized RMSE of 0.1628, demonstrating superior predictive capability over the other examined models. For design-of-experiments (DOE)-based analysis of main effects and two-factor interactions, an MLP surrogate was retained to enable efficient response evaluation across the design space; therefore, DOE findings are interpreted as qualitative, surrogate-dependent trends rather than the most accurate pointwise predictions. Permutation-based importance analysis identified upgrading time as the most consistently influential variable, whereas the secondary driver depended on model family, with temperature contributing more strongly in tree-based ensembles and acidity/catalyst-related descriptors being more prominent in neural/surrogate models. Pressure showed a comparatively smaller contribution within the investigated range. Independent laboratory tests supported the overall predictive trends; however, deviations under some conditions indicate inter-source heterogeneity in the compiled literature data and limited representation of extreme operating regimes. Overall, the proposed framework can serve as a screening and decision-support tool for prioritizing operating conditions and guiding future experimental design in heavy-oil in-situ upgrading studies. The findings of this study can support faster screening of operating conditions, improve understanding of viscosity-reduction behavior, and assist the design of data-driven decision tools for catalytic upgrading of heavy crude oils.</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

Comparative analysis of artificial intelligence models for predicting oil viscosity in the in-situ catalytic oil upgrading process: a study of influential parameters

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月10日

查看全文
  • Alireza Ghodrati Dizaj,
  • Hamed Namdar,
  • Arezou Jafari

摘要

Given the increasing demand for efficient technologies in heavy-oil recovery and upgrading, methods that can reduce viscosity and improve oil mobility are of considerable interest. Heavy oils are difficult to produce and process because of their high viscosity and density, as well as their elevated contents of asphaltenes, heteroatoms, and heavy metals. Although thermal and chemical recovery methods can reduce viscosity during production, the oil may regain viscosity after reaching the surface, often requiring diluents or additional upgrading steps. In this context, nanocatalyst-assisted in-situ upgrading has attracted attention as a route to simultaneously improve oil mobility and quality. In this study, a data-driven framework was developed to predict viscosity reduction during NiO nanoparticle-assisted in-situ upgrading. Predictive performance was evaluated using fivefold out-of-fold cross-validation (OOF-CV) to obtain a more reliable estimate of generalization. Conventional models (MLP, RBF, and ANFIS) were compared with modern tree-based ensembles, including Random Forest and Gradient Boosted Decision Trees (GBDT). GBDT delivered the best cross-validated performance, with a pooled OOF \({R}^{2}\) of 0.925 and a normalized RMSE of 0.1628, demonstrating superior predictive capability over the other examined models. For design-of-experiments (DOE)-based analysis of main effects and two-factor interactions, an MLP surrogate was retained to enable efficient response evaluation across the design space; therefore, DOE findings are interpreted as qualitative, surrogate-dependent trends rather than the most accurate pointwise predictions. Permutation-based importance analysis identified upgrading time as the most consistently influential variable, whereas the secondary driver depended on model family, with temperature contributing more strongly in tree-based ensembles and acidity/catalyst-related descriptors being more prominent in neural/surrogate models. Pressure showed a comparatively smaller contribution within the investigated range. Independent laboratory tests supported the overall predictive trends; however, deviations under some conditions indicate inter-source heterogeneity in the compiled literature data and limited representation of extreme operating regimes. Overall, the proposed framework can serve as a screening and decision-support tool for prioritizing operating conditions and guiding future experimental design in heavy-oil in-situ upgrading studies. The findings of this study can support faster screening of operating conditions, improve understanding of viscosity-reduction behavior, and assist the design of data-driven decision tools for catalytic upgrading of heavy crude oils.