《Processes》:Special Issue on “Modeling, Simulation and Control of Energy Systems”
Dimitris Ipsakis and
Andreas Yiotis
编辑推荐:
本研究利用机器学习算法重新解释最初通过递减曲线分析(DCA)绘制的现有数据集,旨在无需获取新的现场数据即可提高预测精度。研究人员编制了历史生产记录,包括月度油/气产量、井底压力和累积产量,通过最小二乘优化将其拟合至Arps方程,并将关键递减参数(如初始产量、名
本研究利用机器学习算法重新解释最初通过递减曲线分析(DCA)绘制的现有数据集,旨在无需获取新的现场数据即可提高预测精度。研究人员编制了历史生产记录,包括月度油/气产量、井底压力和累积产量,通过最小二乘优化将其拟合至Arps方程,并将关键递减参数(如初始产量、名义递减率和双曲指数)作为输入数据。训练并验证了四种机器学习模型:人工神经网络(ANN)、支持向量机(SVM)和线性回归(LR),采用80/20训练-测试划分和5折交叉验证。采用均方误差(MSE)、均方根误差(RMSE)、平均绝对误差(MAE)和决定系数(R2)评估模型。ANN表现最佳,在独立测试集上达到近乎完美的预测精度(R2≈1),对于石油产量率的误差值较低(MSE = 0.0012 Ncm2/月2,RMSE = 0.035 Ncm/月,MAE = 0.028 Ncm/月)。气体产量率和压力也取得了类似精度。这些结果反映了所分析数据集中存在的强且高度规则的关系,而非精确的零误差拟合。ANN的多层架构有效捕捉了Arps参数与瞬态流动状态之间的非线性交互,优于经验和物理约束方法。线性回归表现良好(R2 = 0.98,RMSE = 0.15 Ncm/月),但在高递减场景下表现不佳,无法准确建模指数尾部。SVM偏差最大(RMSE = 0.42 Ncm/月,R2 = 0.89),归因于在稀疏、噪声递减数据中的核敏感性。随机森林(RF)表现居中(R2 = 0.97)。这种ANN驱动方法通过自动化参数调节和不确定性量化重新定义了递减分析,与经典Arps方法相比,将预测误差降低了85%。
研究背景方面,产量预测一直是油藏工程的核心,直接影响油田开发规划、储量估算和经济决策。递减曲线分析(DCA)作为最广泛使用的油气产量预测工具,基于Arps提出的指数、双曲和调和三种经验模型,能够有效从历史井数据中预测生产趋势。然而,传统DCA存在若干局限性:假设恒定场操作、边界主导流条件、驱动机制不变,且无法准确捕捉瞬态流动状态,DCA参数与渗透率、孔隙度等关键储层属性之间缺乏直接联系。这些限制在非常规储层中尤为突出,导致预测偏离较大。为解决上述问题并避免复杂且昂贵的地质建模、历史拟合等传统流程,研究人员开展了本研究,旨在通过机器学习算法重新解释Arps递减参数,在统一框架下系统比较经典DCA与多种机器学习方法的预测性能。论文发表在《Processes》。
关键技术方法方面,研究人员使用了三种机器学习算法:线性回归(LR)、支持向量机回归(SVM)和人工神经网络(ANN),在MATLAB回归学习环境下实现。数据来自罗马尼亚某成熟气田的300个实验数据点,包含压力与产量率测量值。数据集经清洗、归一化后随机分为训练集(70%)、验证集(15%)和测试集(15%)。性能评价采用均方误差(MSE)、均方根误差(RMSE)、平均绝对误差(MAE)和决定系数(R2)。ANN采用前馈神经网络结构,压力预测网络含一个隐藏层(10个神经元),产量预测网络含两个隐藏层(各10个神经元),使用Levenberg-Marquardt反向传播算法训练。
研究结果部分,原文包含以下小标题:
**1. Introduction**:研究人员通过文献分析指出传统DCA的局限性,并回顾了机器学习在产量预测中的应用(如ML-DCA、贝叶斯框架、随机森林与拉伸指数衰减模型、混合物理神经网络等)。该部分为研究动机奠定基础。
**2. Bibliometric Analysis and Identification of Research Gaps**:研究人员在Web of Science核心合集中进行文献检索,发现仅有25篇同行评审论文明确结合了Arps递减模型与机器学习,表明该领域研究稀疏且碎片化。关键词共现网络显示传统递减曲线分析与机器学习术语之间存在结构分离,且机器学习处于连接传统聚类与产量预测聚类的桥接位置但连接薄弱。时间演化分析表明机器学习辅助预测是新兴方向。主题图将递减分析曲线定位在“新兴或衰退”象限,表明其低中心性。研究人员据此确认本研究的创新性:在统一框架下直接比较经典DCA与多种机器学习算法。
**3. Arps Decline Models**:研究人员介绍了Arps三种经验模型(指数、双曲、调和),并将其应用于300个数据点预测第365、730、1095、1460天的产量与压力。结果显示指数预测相关性最佳,但第300天流量高估约2500 Nm3;双曲预测(b=0.5)高估最终值;调和预测数值脱离实际。压力无法用Arps方法直接现场关联。因此研究人员决定产量率同时用DCA和机器学习,压力仅用机器学习。
**4. Machine-Learning Algorithms**:研究人员构建了LR、SVM、ANN三种模型,分别预测压力与产量率。通过性能指标比较,ANN在压力(R2=1,MSE≈0)和产量率(R2=1,MSE=0.0012 Ncm2/月2)上均达到最佳,误差最低;LR在压力预测表现良好(R2≈0.99),但产量预测在高递减场景下精度下降(R2=0.98,RMSE=0.15 Ncm/月);SVM表现最差(R2=0.89,RMSE=0.42 Ncm/月),因核函数对稀疏噪声数据敏感。ANN的多层架构有效捕捉了油藏非线性动态行为。
**5. Conclusions**:研究结论部分原文为:“This study presents a systematic comparison between classical Decline-Curve Analysis (DCA) and data-driven machine-learning techniques for forecasting gas production rates and reservoir pressure, using a unified dataset from a mature Romanian gas field. By reinterpreting Arps decline parameters through multiple machine-learning algorithms, the research directly addresses a documented gap in the literature where empirical decline models and artificial intelligence methods are rarely evaluated side by side within a consistent framework. The results demonstrate that Artificial Neural Networks (ANN) significantly outperform traditional Arps-based decline models as well as other machine-learning approaches, including Linear Regression (LR) and Support Vector Machines (SVM). ANN achieved the lowest prediction errors across all performance indices and exhibited an almost perfect agreement with experimental data (R2 = 1) for both pressure and flow forecasting. This high predictive accuracy reflects the strong and consistent relationships present in the analyzed dataset. The results demonstrate the capability of ANN models to capture the complex nonlinear behavior associated with reservoir depletion, which is only partially represented by conventional empirical decline equations. Among the machine-learning methods evaluated, Linear Regression performed adequately for pressure estimation but lacked robustness in flow prediction, particularly under nonlinear decline conditions. SVM yielded the weakest overall performance, with higher prediction errors and sensitivity to data distribution. In contrast, ANN consistently provided stable and accurate predictions, highlighting its suitability for practical reservoir engineering applications. From a broader perspective, the bibliometric analysis confirms that research integrating decline-curve analysis with machine learning remains sparse and fragmented. By embedding classical Arps models within an ANN-based predictive framework, this study effectively bridges an emerging research theme with a foundational production forecasting methodology, contributing to the modernization of decline-curve analysis within the digital oilfield paradigm. The proposed ANN-assisted forecasting approach offers a practical, data-efficient alternative to complex reservoir simulation, requiring no additional field data while substantially reducing forecasting uncertainty. Its implementation can support improved decision-making in reserves estimation, production optimization, and field development planning. Beyond the case study presented, the proposed methodology has broader applicability due to its data-driven nature and ability to model complex nonlinear relationships. The methodology may be applied to other petroleum engineering applications, such as estimation fluid and fuel properties. Furthermore, the methodology can be applied to the modeling and optimization of industrial processes where experimental data are available, including process performance prediction, equipment behavior analysis, and system monitoring. These examples highlight the versatility of the proposed approach for a wide range of engineering problems involving experimental datasets and complex system dynamics.”