Wang Zisong, Cheng Zhiliang, Ding Xiujian, Xia Lu
School of Civil Engineering and Transportation, Weifang University, Weifang, Shandong, China.
School of Petroleum Engineering, China University of Petroleum (East China), Qingdao, Shandong, China.
PLoS One. 2024 Dec 5;19(12):e0314108. doi: 10.1371/journal.pone.0314108. eCollection 2024.
The process of extracting oil and gas via borehole drilling is largely dependent on subsurface structures, and thus, well log analysis is a major concern for economic feasibility. Well logs are essential for understanding the geology below the earth's surface, which allows for the estimation of the available hydrocarbon resources. The incompleteness of these logs, on the other hand, is a major hindrance to downstream analysis success. This study, however, addresses the above challenges and presents a deep Long-Short Term Memory (LSTM) model specialized using a new hyperparameter tuning algorithm. There is an evidence gap that we try to fill: well log prediction using LSTM has not been extensively documented, particularly on reconstruction of missing data. In order to remedy this, we develop a new algorithm entitled Elite Preservation Strategy Chimp Optimization Algorithm (EPSCHOA), which will improve the tuning of LSTM hyperparameters. EPSCHOA enhances prediction performance by preserving the diversity of the strongest candidates and transforming the most effective predictor resources into less effective ones. A comparative analysis of the LSTM-EPSCHOA model was carried out with both LSTM and E-LSTM models, including their various extensions, LSTM-CHOA, LSTM-HGSA, LSTM-IMPA, LSTM-SEB-CHOA, and LSTM-GOLCHOA, even as common forecasting models using Artificial Neural Network (ANN), Adaptive Neuro-Fuzzy Inference System (ANFIS), Gradient Boosting (GB), and AutoRegressive Integrated Moving Average (ARIMA). The results of the performance tests demonstrate that the LSTM-EPSCHOA model outperforms in all aspects, as evidenced by its R2 values of.98, RMSE of 0.022, and MAPE of 0.701% during training, and R2 values of 0.96, RMSE of 0.025, and MAPE of 0.698% during testing. These are considerably superior to other measures used compared to what was achieved using explicit modeling using LSTM, which stood at R2 of 0.59, RMSE of 0.101, and MAPE of 2.588%. The LSTM-EPSCHOA proved to give models faster rates of convergence and lower error measurements than usual models, which clearly demonstrated its efficiency in solving the problem of inadequate well-log data. The new approach is regarded as having many useful potentials to boost well-log interpretations in the oil sector.
通过钻孔开采石油和天然气的过程在很大程度上取决于地下结构，因此，测井分析是经济可行性的主要关注点。测井对于了解地球表面以下的地质情况至关重要，这有助于估算可用的碳氢化合物资源。另一方面，这些测井数据的不完整性是下游分析成功的主要障碍。然而，本研究解决了上述挑战，并提出了一种使用新的超参数调整算法专门设计的深度长短期记忆（LSTM）模型。我们试图填补一个证据空白：使用LSTM进行测井预测尚未得到广泛记录，特别是在缺失数据的重建方面。为了弥补这一点，我们开发了一种名为精英保留策略黑猩猩优化算法（EPSCHOA）的新算法，它将改进LSTM超参数的调整。EPSCHOA通过保留最强候选者的多样性并将最有效的预测资源转化为较无效的资源来提高预测性能。对LSTM-EPSCHOA模型与LSTM和E-LSTM模型及其各种扩展模型LSTM-CHOA、LSTM-HGSA、LSTM-IMPA、LSTM-SEB-CHOA和LSTM-GOLCHOA进行了比较分析，同时也与使用人工神经网络（ANN）、自适应神经模糊推理系统（ANFIS）、梯度提升（GB）和自回归积分移动平均（ARIMA）的常见预测模型进行了比较。性能测试结果表明，LSTM-EPSCHOA模型在各方面均表现出色，训练期间其R2值为0.98，均方根误差（RMSE）为0.022，平均绝对百分比误差（MAPE）为0.701%，测试期间R2值为0.96，RMSE为0.025，MAPE为0.698%。与使用LSTM显式建模所取得的结果（R2为0.59，RMSE为0.101，MAPE为2.588%）相比，这些结果明显优于其他测量值。LSTM-EPSCHOA被证明比通常的模型具有更快的收敛速度和更低的误差测量值，这清楚地证明了其在解决测井数据不足问题方面的效率。这种新方法被认为具有许多有用的潜力，可以促进石油行业的测井解释。
