面向盲井预测的多尺度时序融合网络密度测井曲线重构方法

A Multi-Scale Temporal Fusion Network for Density Log Reconstruction in Blind-Well Prediction

摘要: 密度测井曲线(DEN)是储层评价、孔隙度估算和岩性解释的重要基础资料，但在实际测井过程中，受井眼条件、仪器故障或测井项目不完整等因素影响，DEN曲线常出现缺失或异常。为提高盲井条件下DEN曲线重构精度，本文提出一种多尺度TCN-ECA-BiGRU-Transformer网络，并将该任务构建为基于滑动窗口的中心点回归问题。模型以常规测井曲线、中心深度信息和电阻率比特征等9个可获得输入变量为基础，预测滑动窗口中心位置的DEN值。该模型融合了多尺度时序卷积和高效通道注意力机制，用于提取局部深度方向特征，引入残差BiGRU与跳跃融合结构，用于增强双向上下文表征能力，并采用基于RoPE位置编码的Transformer，以建模长距离深度依赖关系。实验基于大庆油田12口井的7394个有效样本开展，采用按井划分策略，并以T252井作为独立盲井测试集。结果表明，所提出模型的RMSE、MAE和R²分别为0.0296 g/cm³、0.0198 g/cm³和0.9290，优于多种机器学习和深度学习基线模型。进一步的留一井交叉验证结果显示，模型的平均RMSE、MAE和R²分别为0.0316 g/cm³、0.0213 g/cm³和0.9167。上述结果表明，所提出的时序建模框架能够有效提升盲井DEN曲线重构精度，可为测井曲线缺失段补全和异常曲线修复提供支持。

Abstract: Density logs (DEN) are essential for reservoir evaluation, porosity estimation, and lithology interpretation, but they are often incomplete or abnormal owing to borehole conditions, tool failure, or limited logging programs. To improve blind-well DEN reconstruction, this study proposes a multi-scale TCN-ECA-BiGRU-Transformer network and formulates the task as sliding-window center-point regression. Nine available inputs, including conventional well logs, center-depth information, and a resistivity_ratio feature, are used to predict the DEN value at the window center. The model integrates multi-scale temporal convolution and efficient channel attention for local depth-wise feature extraction, residual BiGRU and skip fusion for bidirectional contextual representation, and a RoPE-based Transformer for long-range dependency modeling. Experiments were conducted on 7394 valid samples from 12 wells in the Daqing Oilfield, using well-wise splitting and an independent blind test on well T252. The proposed model achieved RMSE, MAE, and R² values of 0.0296 g/cm³, 0.0198 g/cm³, and 0.9290, respectively, outperforming machine-learning and deep-learning baselines. Leave-one-well-out validation yielded average RMSE, MAE, and R² values of 0.0316 g/cm³, 0.0213 g/cm³, and 0.9167. These results indicate that the proposed temporal modeling framework improves blind-well DEN reconstruction and can support missing-interval completion and anomalous-log repair.