基于深度学习的水平气井积液诊断与智能排采分析

Diagnosis of Liquid Loading in Horizontal Gas Wells and Intelligent Deliquification Analysis Based on Deep Learning

摘要: 为解决水平气井中后期因井底积液诊断不准、排采措施滞后而导致产能下降与经济性差的问题，提出一种融合井下传感器高频数据与多相管流机理的物理信息神经网络与双向长短期记忆网络(PINN-Bi-LSTM)耦合模型。将压力梯度偏微分方程作为物理约束正则项嵌入神经网络损失函数，准确预测气井开始特征，并构建“泡排–气举–速度管柱”三级智能排采优选矩阵。应用表明，模型在复杂工况下的积液诊断准确率达93.5%，优于常规积液诊断模型与纯数据驱动模型。实施智能决策系统后，单井平均自喷期延长14个月，操作成本降低18.2%，投资回收期缩短至0.67年。所提模型兼具数据挖掘与物理可解释性，为水平气井全生命周期智能排采与经济评价提供了一套可靠的技术方案。

Abstract: To address the challenges of inaccurate liquid-loading diagnosis and delayed drainage measures in the mid-to-late stages of horizontal gas wells, which lead to production decline and poor economic performance, this paper proposes a hybrid model coupling a physics-informed neural network with a bidirectional long short-term memory network (PINN-Bi-LSTM). To achieve precise characterization of liquid loading initiation, the partial differential equation describing pressure gradients is incorporated into the loss function of neural network, as a physics-informed regularization term. In addition, a three-level intelligent optimization framework for drainage strategy selection-encompassing foam lifting, gas lift, and velocity string technologies-is constructed. Field test results verify that, under dynamically changing and complex operating scenarios, the proposed model attains a liquid-loading diagnosis accuracy of 93.5%, which surpasses the performance of both traditional diagnostic models and approaches relying solely on data-driven methodologies. After the deployment of the intelligent decision-making system, the average flowing duration of each well has been prolonged by 14 months, operational expenses have been cut by 18.2%, and the investment payback period has been reduced to 0.67 years. By fusing data-driven efficiency with physical interpretability, the model presented herein delivers a reliable technical approach for intelligent drainage and full-lifecycle economic assessment of horizontal gas wells.