摘要：

储层压后产量精准预测对于优化压裂施工参数和评估储层改造效果具有重要意义。然而，实际工区开发过程中，由于现场样本数量有限，传统数据驱动模型容易出现过拟合问题，导致复杂储层条件下预测可靠性不足。针对上述问题，本文提出了一种融合数据增强与迁移学习的储层压后产量预测方法，以提高小样本条件下压后产量预测的准确性和稳定性。首先，综合考虑储层地质特征、工程地质参数以及压裂施工条件等多方面影响因素，构建包含多源特征参数的压后产量预测数据集。针对样本不足导致模型训练受限的问题，引入带梯度惩罚的条件生成对抗网络(Conditional Wasserstein Generative Adversarial Network with Gradient Penalty，CWGAN-GP)，通过学习真实样本的数据分布特征生成高质量扩充数据，提高训练样本的多样性和代表性。随后，结合最大相关最小冗余(Maximum Relevance Minimum Redundancy，mRMR)与皮尔逊(Pearson)相关分析的混合特征选择方法，筛选影响储层压后产量的关键控制因素，降低冗余信息对模型预测性能的干扰。在此基础上，构建知识约束下数据联合驱动的注意力机制双向门控循环单元(Attention-based Bidirectional Gated Recurrent Unit，Attention-BiGRU)模型，将储层工程知识融入深度学习过程，提高预测结果的合理性和稳定性。同时，引入迁移学习策略，减弱不同储层之间数据分布差异对模型性能的影响，提升目标区块小样本条件下模型的泛化能力。研究结果表明，所提出模型在测试集上的决定系数(R²)达到0.93，相较于传统机器学习模型及单一深度学习模型均表现出更高的预测精度。对比实验和消融实验进一步证明，各模块均能够有效提升模型预测性能，并表现出协同增益作用。本文方法有效解决了小样本条件下储层压后产量预测精度不足的问题，为储层产量评价和油气井生产分析提供了一种可靠的方法。

关键词: 低渗油藏, 迁移学习, 注意力机制, 门控循环神经网络, 产量预测

Abstract:

Accurate prediction of post-fracturing production is essential for optimizing hydraulic fracturing parameters and evaluating reservoir stimulation effectiveness. However, the limited availability of field samples in practical reservoirs often causes conventional data-driven models to suffer from overfitting and insufficient prediction reliability under complex reservoir conditions. To address this issue, this study proposes a post-fracturing production prediction framework integrating data augmentation and transfer learning to improve prediction accuracy and stability under small-sample conditions. A multi-source dataset is first constructed by comprehensively considering reservoir geological characteristics, engineering geological parameters, and fracturing operation conditions. To overcome the limitation of insufficient training samples, a Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP) is employed to learn the distribution characteristics of real samples and generate high-quality augmented data, thereby improving the diversity and representativeness of training samples. Subsequently, a hybrid feature selection method combining Maximum Relevance Minimum Redundancy (mRMR) and Pearson correlation analysis is developed to identify dominant controlling factors affecting post-fracturing production and reduce the interference of redundant information. Based on the selected key features, an attention-based Bidirectional Gated Recurrent Unit (Attention-BiGRU) model jointly driven by data and knowledge constraints is established. By integrating reservoir engineering knowledge into the deep learning process, the proposed model enhances the rationality and stability of prediction results. Furthermore, transfer learning is introduced to alleviate the influence of data distribution differences between different reservoirs and improve the generalization capability of the model under limited target samples. The results demonstrate that the proposed model achieves an R² value of 0.93 on the testing dataset, outperforming conventional machine learning models and individual deep learning models. Comparative experiments and ablation studies further verify that each module contributes to improving prediction performance, showing clear synergistic effects among the integrated components. The proposed method effectively addresses the challenge of inaccurate post-fracturing production prediction caused by insufficient samples and provides a reliable approach for reservoir productivity evaluation and oil and gas production analysis.

Key words: low permeability reservoir, transfer learning, Attention Mechanism, gated recurrent neural network, production prediction