Scholarlink

<p>Mud-logging data, including drilling data (DD) and gas while drilling (GWD) parameters, are recorded in every oil and gas well drilled and are available for real-time assessment making them more beneficial for prompt assessment of formation assessment than the higher cost data provided by cores and wireline or measurement-while-drilling logs. The novel contribution of this research is that integrates DD and GWD data, to enhance prediction performance of formation bulk density in real time during drilling operations, It achieves this using interpretable machine learning Shapley Additive exPlanations (SHAP) analysis combined with four gradient-boosting models (CatBoost, Adaptive Boosting (AdaBoost), Light Gradient Boosting Machine (LightGBM) and Extreme Gradient Boosting (XGBoost)) applied to data collected from two productive Algerian fields: Sif Fatima undersaturated (gas-poor) oil reservoir and Oued Zine saturated (gas-rich) oil reservoir. The results show that the DD and GWD inputs could effectively predict bulk density in both gas-poor and gas-rich oil reservoirs with various levels of precision. For gas-poor oil reservoirs, DD inputs, used in isolation generate predictions with low errors, particularly with model CatBoost4 (R≈0.905, RMSE≈0.040&#xa0;g/cm<sup>3</sup>, MAE≈0.018&#xa0;g/cm<sup>3</sup>, NSE≈0.815). For the gas-rich oil reservoir, DD inputs also achieved low errors (R≈0.945, RMSE≈0.025&#xa0;g/cm<sup>3</sup>, MAE≈0.012&#xa0;g/cm<sup>3</sup>, NSE≈0.845). However, the bulk formation density predictions of these reservoirs were better using GWD data in isolation (e.g., model CatBoost5: R≈0.991, RMSE≈0.009&#xa0;g/cm<sup>3</sup>, MAE≈0.005&#xa0;g/cm<sup>3</sup>, NSE≈0.980). These results are consistent with the expectations of gas-saturated oil reservoirs, as reported for real time porosity predictions. Good prediction performance was also achieved for the gas-rich oil reservoir by integrating both GWD and DD inputs, particularly with model AdaBoost:2 (R ≈ 0.985. RMSE ≈0.012&#xa0;g/cm<sup>3</sup>, MAE≈0.009&#xa0;g/cm<sup>3</sup>, NSE≈0.962). This investigation confirms that the use of Drilling Data (DD) is more effective for predicting bulk density in gas-poor reservoirs, while predictions using Gas While Drilling (GWD) data are more accurate in gas-rich reservoirs,This study also confirms the potential for real time bulk formation density predictions using mud logging parameters, justifying the compilation of larger datasets and comparing results from a wide range of reservoir formations and fluid types.</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

Enhancing formation bulk density prediction while drilling using mud logging data and interpretable boosting machine learning

  • 期刊论文

  • 发表日期：

    2025年1月14日

查看全文
  • Ayoub Boutaghane,
  • Ouafi Ameur-Zaimeche,
  • Salim Heddam,
  • Rabah Kechiched,
  • Nasreddine Tahar-Belkacem,
  • Abdelhamid Ouladmansour,
  • Watheq J. Al-Mudhafar,
  • David A. Wood

摘要

Mud-logging data, including drilling data (DD) and gas while drilling (GWD) parameters, are recorded in every oil and gas well drilled and are available for real-time assessment making them more beneficial for prompt assessment of formation assessment than the higher cost data provided by cores and wireline or measurement-while-drilling logs. The novel contribution of this research is that integrates DD and GWD data, to enhance prediction performance of formation bulk density in real time during drilling operations, It achieves this using interpretable machine learning Shapley Additive exPlanations (SHAP) analysis combined with four gradient-boosting models (CatBoost, Adaptive Boosting (AdaBoost), Light Gradient Boosting Machine (LightGBM) and Extreme Gradient Boosting (XGBoost)) applied to data collected from two productive Algerian fields: Sif Fatima undersaturated (gas-poor) oil reservoir and Oued Zine saturated (gas-rich) oil reservoir. The results show that the DD and GWD inputs could effectively predict bulk density in both gas-poor and gas-rich oil reservoirs with various levels of precision. For gas-poor oil reservoirs, DD inputs, used in isolation generate predictions with low errors, particularly with model CatBoost4 (R≈0.905, RMSE≈0.040 g/cm3, MAE≈0.018 g/cm3, NSE≈0.815). For the gas-rich oil reservoir, DD inputs also achieved low errors (R≈0.945, RMSE≈0.025 g/cm3, MAE≈0.012 g/cm3, NSE≈0.845). However, the bulk formation density predictions of these reservoirs were better using GWD data in isolation (e.g., model CatBoost5: R≈0.991, RMSE≈0.009 g/cm3, MAE≈0.005 g/cm3, NSE≈0.980). These results are consistent with the expectations of gas-saturated oil reservoirs, as reported for real time porosity predictions. Good prediction performance was also achieved for the gas-rich oil reservoir by integrating both GWD and DD inputs, particularly with model AdaBoost:2 (R ≈ 0.985. RMSE ≈0.012 g/cm3, MAE≈0.009 g/cm3, NSE≈0.962). This investigation confirms that the use of Drilling Data (DD) is more effective for predicting bulk density in gas-poor reservoirs, while predictions using Gas While Drilling (GWD) data are more accurate in gas-rich reservoirs,This study also confirms the potential for real time bulk formation density predictions using mud logging parameters, justifying the compilation of larger datasets and comparing results from a wide range of reservoir formations and fluid types.