物理—数据双驱动的压裂压力实时预测方法

A physics and data dual-driven method for real-time fracturing pressure prediction

胡晓东 刘俊仪 王天宇 周福建 卢旭涛 易普康 陈超

HU Xiaodong LIU Junyi WANG Tianyu ZHOU Fujian LU Xutao YI Pukang CHEN Chao

中国石油大学(北京) 人工智能学院, 北京 102249

College of Artificial Intelligence, China University of Petroleum (Beijing), Beijing 102249, China

井口压力预测存在压力波动剧烈、干扰因素多以及影响机理复杂等问题。现阶段研究中，由于对复杂的地层条件、裂缝特征及流体动力学过程的过度简化，传统物理模型难以捕捉多重非线性变化和突发波动，导致在真实施工环境下的预测精度和实时响应能力受到局限。而人工智能模型尽管具有较强的非线性拟合能力，但往往缺乏对压力波动的物理机理的深入理解，对地层和施工参数的敏感性不足，导致在极端或动态变化的条件下稳定性较差、解释性不足。针对这一难题，提出了一种物理—数据双驱动的压力曲线的预测方法对未来压力趋势进行预测。首先，构建了基于长短期记忆(LSTM)神经网络的智能模型，融合缝内支撑剂床平衡高度计算结果与井场实时泵注数据作为模型输入，预测了未来60 s的压力数据；其次，结合传统井口压力反演方法，使用小波变换分解智能模型与传统模型预测结果，利用LSTM模型整体趋势与压力反演计算方法(IPC)模型中突变点特征，重构了兼顾整体趋势和局部波动的井口压力预测曲线。结果表明，相比LSTM模型，IPC和LSTM的小波融合模型未来60 s井口压力预测的均方根误差(RMSE)和均方绝对误差(MAE)分别下降了37.87%和15.29%，预测结果能够精准捕捉现场施工的压裂压力变化，为现场施工提供更为可靠的指导和决策依据。

Wellhead pressure prediction is challenging due to problems such as drastic pressure fluctuations, numerous disturbing factors, and complex influencing mechanisms. Current research often adopts traditional physical models which find it difficult to capture multiple nonlinear changes and sudden fluctuations due to the oversimplification of complex formation conditions, fracture characteristics, and fluid dynamics processes, limiting their prediction accuracy and real-time responsiveness in actual operations. Artificial intelligence (AI) models, despite their strong nonlinear fitting capabilities, often lack an in-depth understanding of the physical mechanisms underlying pressure fluctuations and are less sensitive to formation and operational parameters, resulting in poor stability and insufficient interpretability under extreme or dynamically changing conditions. To address these challenges, a physics and data dual-driven prediction method was proposed to predict future pressure trends. An intelligent model based on a long and short-term memory (LSTM) neural network was constructed, integrating the equilibrium height calculations of the proppant bed within the fracture and real-time pumping data at the wellsite as model inputs to predict pressure for the next 60 seconds. Then, combined with traditional wellhead pressure inversion method, wavelet transform was used to decompose predictions from both the intelligent and traditional models. The overall trend of the LSTM model and the characteristics of mutation point in the inverse pressure calculation (IPC) model were utilized to reconstruct the wellhead pressure prediction curves that could balance the overall trend and local fluctuations. Results showed that compared to pure LSTM model, the wavelet fusion model of IPC and LSTM reduced the root mean square error (RMSE) and mean absolute error (MAE) by 37.87% and 15.29%, respectively, in wellhead pressure prediction for the next 60 seconds. The fusion model can accurately capture fracturing pressure changes during field operations, providing more reliable guidance and decision support for field operations.

国家自然科学基金“超深层强应力下力—化联合作用水力裂缝缝网扩展与控制机理” U23B2084

https://doi.org/10.11781/sysydz2024061323