基于物理引导深度学习的超深碳酸盐岩储层横波时差预测方法

doi:10.3969/j.issn.2096-1693.2026.01.033

石油科学通报 ›› 2026, Vol. 11 ›› Issue (5): 1651-1672. doi: 10.3969/j.issn.2096-1693.2026.01.033

基于物理引导深度学习的超深碳酸盐岩储层横波时差预测方法

陈思程¹^,²^,³(), 张恒¹^,²^,³, 孔令运¹^,²^,³^,⁴^,^*(), 刘昭茜¹^,²^,³

¹ 油气勘探开发理论与技术湖北省重点实验室(中国地质大学(武汉))，武汉 430074
² 中国地质大学(武汉)构造与油气资源教育部重点实验室，武汉 430074
³ 中国地质大学(武汉)资源学院，武汉 430074
⁴ 深层地热富集机理与高效开发全国重点实验室，武汉 430074

收稿日期:2026-04-20 修回日期:2026-06-22 出版日期:2026-10-15 发布日期:2026-10-31
通讯作者: ^*孔令运(1990年—)，博士，特任教授、博士生导师，主要从事页岩油气储层岩石力学、地质工程一体化等领域的研究，lingyun.kong@cug.edu.cn。
作者简介:陈思程(2005年—)，从事资源环境大数据研究，csc2326@qq.com。
基金资助:
国家级大学生创新创业训练计划项目(202610491013);中央高校基本科研业务费专项资金资助项目(2021259)

Physics-guided deep learning for shear-wave transit-time prediction in ultra-deep carbonate reservoirs

CHEN Sicheng¹^,²^,³(), ZHANG Heng¹^,²^,³, KONG Lingyun¹^,²^,³^,⁴^,^*(), LIU Zhaoqian¹^,²^,³

¹ Hubei Provincial Key Laboratory of Oil and Gas Exploration and Development Theory and Technology, China University of Geosciences (Wuhan), Wuhan 430074, China
² Key Laboratory of Tectonics and Petroleum Resources, Ministry of Education, China University of Geosciences (Wuhan), Wuhan 430074, China
³ School of Resources, China University of Geosciences (Wuhan), Wuhan 430074, China
⁴ State Key Laboratory of Deep Geothermal Enrichment Mechanisms and Efficient Development, Wuhan 430074, China

Received:2026-04-20 Revised:2026-06-22 Online:2026-10-15 Published:2026-10-31
Contact: *lingyun.kong@cug.edu.cn。

摘要/Abstract

摘要：

在超深碳酸盐岩勘探中，横波时差(DTS)是具有重要工程意义的岩石物理参数，但能够直接获取横波信息的偶极子声波测井并非常规配置，受成本、时效与井下风险等限制，难以实现全井段覆盖，因此利用常规测井曲线智能预测DTS具有重要的意义和价值。中国西部某大型叠合盆地超深缝洞型碳酸盐岩储层，经历多期岩溶改造与断裂控制，纵波时差(AC)与DTS之间不保持稳定的单值映射关系，研究区6口井的AC-DTS Pearson相关系数介于0.42~0.96，常规预测模型普遍存在井内拟合精度高、跨井外推误差放大的泛化难题。针对这一问题，本文以AC、DEN、GR、CNL、RD、RS六条常规测井曲线为输入，搭建了融合一维卷积、双向长短期记忆网络与注意力机制的深度序列模型(CNN-BiLSTM-Attention)，以Castagna关系为岩石物理软先验、固定双神经网络融合为主模型，并将域自适应门控作为探索方案来缓解跨井响应漂移；评价联合R²、RMSE、MAE、MAPE、Pearson相关系数与Kling-Gupta效率系数(KGE)，并以留一井交叉验证(LOWO)作为衡量未见井外推能力的核心口径。在合并的9533个测试样本上，模型全样本合并决定系数(Pooled R²)为0.818，RMSE为9.79 μs/ft，MAE为4.69 μs/ft，MAPE为3.88%，Pearson相关系数为0.918，KGE为0.756；在更严格的LOWO口径下，中位与均值R²为0.516和0.565，明显优于仅采用Castagna公式的纯物理基线，后者的留一井外推中位R²仅为0.204。误差传播分析表明，DTS预测误差会传递至动态杨氏模量、泊松比和归一化脆性指数等依赖DTS计算的地质力学参数，井级误差相关系数约为0.87~0.95，说明DTS高值段和孔洞—裂缝强烈发育段不仅是横波预测难点，也是工程参数评价的不确定性集中区。同时，单看R²易受窄动态范围小分母效应误导，需结合绝对误差与KGE一并判断。该流程在多数典型井上能够稳定恢复DTS主趋势，为缺失横波资料的超深缝洞型碳酸盐岩井段提供了一条物理引导、可跨井迁移的横波时差获取途径。

关键词: 横波时差预测, 超深碳酸盐岩, 缝洞型储层, 深度学习, CNN-BiLSTM-Attention, 岩石物理先验, 跨井泛化

Abstract:

In ultra-deep carbonate exploration, shear-wave transit time (DTS) is a petrophysical parameter of considerable engineering significance. However, dipole sonic logging—the primary means of acquiring shear-wave data—is often excluded from conventional logging suites because of cost, time, and downhole risks, making full-interval coverage difficult. Therefore, predicting DTS from conventional logs has substantial practical value. In the ultra-deep fractured-vuggy carbonate reservoirs of a large superimposed basin in western China, which experienced multiple phases of karst modification and fault control, compressional-wave transit time (AC) and DTS lack a stable single-valued mapping, with AC-DTS Pearson correlation coefficients across the six study wells ranging from 0.42 to 0.96. Consequently, conventional models generalize poorly, with high within-well accuracy but amplified errors when extrapolated to unseen wells. To address this issue, six conventional logs (AC, DEN, GR, CNL, RD, and RS) are utilized as inputs to construct a deep sequence model, CNN-BiLSTM-Attention, integrating one-dimensional convolution, bidirectional long short-term memory networks, and an attention mechanism. The Castagna relation is incorporated as a soft rock-physics prior, a fixed dual-neural-network fusion is adopted as the main model, and domain-adaptive gating is introduced as an exploratory scheme to mitigate cross-well response drift. The evaluation incorporates the coefficient of determination R², root-mean-square error RMSE, mean absolute error MAE, mean absolute percentage error MAPE, Pearson correlation coefficient, and Kling-Gupta efficiency KGE, utilizing leave-one-well-out (LOWO) cross-validation as the core measure of extrapolation to unseen wells. Across the 9533 pooled test samples, the model attains a pooled coefficient of determination Pooled R² of 0.818, an RMSE of 9.79 μs/ft, MAE of 4.69 μs/ft, MAPE of 3.88%, and Pearson and KGE values of 0.918 and 0.756, respectively. Under the stricter LOWO protocol, the median and mean R² reach 0.516 and 0.565, distinctly outperforming the physics-only baseline derived solely from the Castagna formula, for which the LOWO median R² is merely 0.204. Error propagation analysis indicates that DTS prediction errors are transferred to derived geomechanical parameters (dynamic Young's modulus, Poisson's ratio, and normalized brittleness index) that rely on DTS, with well-level error correlation coefficients of 0.87~0.95. This demonstrates that high-DTS intervals and zones with strongly developed vugs and fractures are not only challenging for shear-wave prediction but also represent areas of concentrated uncertainty for engineering parameter evaluation. Meanwhile, evaluating R² in isolation can be misleading due to the small-denominator effect of narrow dynamic ranges, necessitating joint evaluation with absolute error and KGE. Overall, the workflow stably recovers the principal DTS trend in most typical wells, providing a physics-guided and cross-well-transferable approach to acquiring shear-wave transit time for ultra-deep fractured-vuggy carbonate intervals lacking shear-wave data.

Key words: shear-wave transit-time (DTS) prediction, ultra-deep carbonate reservoirs, fractured-vuggy reservoirs, deep learning, CNN-BiLSTM-Attention, rock-physics prior, cross-well generalization

中图分类号:

陈思程, 张恒, 孔令运, 刘昭茜. 基于物理引导深度学习的超深碳酸盐岩储层横波时差预测方法[J]. 石油科学通报, 2026, 11(5): 1651-1672.

CHEN Sicheng, ZHANG Heng, KONG Lingyun, LIU Zhaoqian. Physics-guided deep learning for shear-wave transit-time prediction in ultra-deep carbonate reservoirs[J]. Petroleum Science Bulletin, 2026, 11(5): 1651-1672.

https://sykxtb.cup.edu.cn/CN/Y2026/V11/I5/1651

图/表 23

参考文献 40

[1]	何登发, 贾承造, 赵文智, 等. 中国超深层油气勘探领域研究进展与关键问题[J]. 石油勘探与开发, 2023, 50(6): 1162-1172.
	[He D F, Jia C Z, Zhao W Z, et al. Research progress and key issues of ultra-deep oil and gas exploration in China[J]. Petroleum Exploration and Development, 2023, 50(6): 1162-1172.]
[2]	马永生, 蔡勋育, 云露, 等. 塔里木盆地顺北超深层碳酸盐岩油气田勘探开发实践与理论技术进展[J]. 石油勘探与开发, 2022, 49(1): 1-17.
	[Ma Y S, Cai X Y, Yun L, et al. Practice and theoretical and technical progress in exploration and development of Shunbei ultra-deep carbonate oil and gas field, Tarim Basin, NW China[J]. Petroleum Exploration and Development, 2022, 49(1): 1-17.]
[3]	焦方正. 塔里木盆地深层碳酸盐岩缝洞型油藏体积开发实践与认识[J]. 石油勘探与开发, 2019, 46(3): 552-558.
	[Jiao F Z. Practice and knowledge of volumetric development of deep fractured-vuggy carbonate reservoirs in Tarim Basin, NW China[J]. Petroleum Exploration and Development, 2019, 46(3): 552-558.]
[4]	Castagna J P, Batzle M L, Eastwood R L. Relationships between compressional-wave and shear-wave velocities in clastic silicate rocks[J]. Geophysics, 1985, 50(4): 571-581.
[5]	Rajabi M, Bohloli B, Gholampour Ahangar E. Intelligent approaches for prediction of compressional, shear and Stoneley wave velocities from conventional well log data: A case study from the Sarvak carbonate reservoir in the Abadan Plain (Southwestern Iran)[J]. Computers & Geosciences, 2010, 36(5): 647-664.
[6]	Bagheripour P, Gholami A, Asoodeh M, et al. Support vector regression based determination of shear wave velocity[J]. Journal of Petroleum Science and Engineering, 2015, 125: 95-99.
[7]	罗功伟, 安小平, 姚卫华, 等. 非常规油气储层测井智能解释应用现状与发展趋势[J]. 石油科学通报, 2025, 10(5): 908-925.
	[Luo G W, An X P, Yao W H, et al. Application status and development trends of artificial intelligence in logging interpretation for unconventional oil and gas reservoirs[J]. Petroleum Science Bulletin, 2025, 10(5): 908-925.]
[8]	Greenberg M L, Castagna J P. Shear-wave velocity estimation in porous rocks: Theoretical formulation, preliminary verification and applications[J]. Geophysical Prospecting, 1992, 40(2): 195-209.
[9]	Li Y, Downton J. The applications of amplitude versus offset in carbonate reservoir: Re-examining the potential[C]//SEG Technical Program Expanded Abstracts 2000, Calgary, Alberta, Canada, 2000: 166-169.
[10]	Zahmatkesh I, Soleimani B, Kadkhodaie A, et al. Estimation of DSI log parameters from conventional well log data using a hybrid particle swarm optimization-adaptive neuro-fuzzy inference system[J]. Journal of Petroleum Science and Engineering, 2017, 157: 842-859.
[11]	Mehrad M, Ramezanzadeh A, Bajolvand M, et al. Estimating shear wave velocity in carbonate reservoirs from petrophysical logs using intelligent algorithms[J]. Journal of Petroleum Science and Engineering, 2022, 212: 110254.
[12]	Feng G, Zeng H H, Xu X R, et al. Shear wave velocity prediction based on deep neural network and theoretical rock physics modeling[J]. Frontiers in Earth Science, 2023, 10: 1025635.
[13]	Feng G, Liu W Q, Yang Z, et al. Shear wave velocity prediction based on 1D CNN-BiLSTM network with attention mechanism[J]. Frontiers in Earth Science, 2024, 12: 1376344.
[14]	Cova D, Liu Y. Shear wave velocity prediction using bidirectional recurrent gated graph convolutional network with total information embeddings[J]. Frontiers in Earth Science, 2023, 11: 1101601.
[15]	Hazbeh O, Rajabi M, Tabasi S, et al. Determination and investigation of shear wave velocity based on one deep/machine learning technique[J]. Alexandria Engineering Journal, 2024, 92: 358-369.
[16]	Li J Y, Lin Y T, Gui Z X, et al. Inception-Attention-BiLSTM hybrid network: A novel approach for shear wave velocity prediction utilizing well logging[J]. Applied Sciences, 2025, 15(5): 2345.
[17]	孙正心, 金衍, 孟翰, 等. 基于深度学习数据融合的测井数据精细表征[J]. 石油科学通报, 2025, 10(1): 75-86.
	[Sun Z X, Jin Y, Meng H, et al. Fine characterization of logging data based on the deep learning data fusion[J]. Petroleum Science Bulletin, 2025, 10(1): 75-86.]
[18]	Kheirollahi H, Shad Manaman N, Leisi A. Robust estimation of shear wave velocity in a carbonate oil reservoir from conventional well logging data using machine learning algorithms[J]. Journal of Applied Geophysics, 2023, 211: 104971.
[19]	Moreno-Torres J G, Raeder T, Alaiz-Rodríguez R, et al. A unifying view on dataset shift in classification[J]. Pattern Recognition, 2012, 45(1): 521-530.
[20]	Zhou K Y, Liu Z W, Qiao Y, et al. Domain generalization: A survey[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(4): 4396-4415.
[21]	Zhang T, Chai H, Wang H, et al. Interpretable machine learning model for shear wave estimation in a carbonate reservoir using LightGBM and SHAP: A case study in the Amu Darya right bank[J]. Frontiers in Earth Science, 2023, 11: 1217384.
[22]	Esmaeilpour S, Ispas I. Data-driven shear wave velocity prediction in a carbonate reservoir using conventional well logs[C]. 57th US Rock Mechanics/Geomechanics Symposium, Atlanta, Georgia, USA, 2023: ARMA- 2023-0180.
[23]	邬德刚, 吴胜和, 张玉飞, 等. 小样本条件下的储层物性参数智能解释方法研究[J]. 石油科学通报, 2025, 10(2): 378-391.
	[Wu D G, Wu S H, Zhang Y F, et al. Research on intelligent interpretation methods for reservoir physical parameters under few-shot conditions[J]. Petroleum Science Bulletin, 2025, 10(2): 378-391.]
[24]	Wang Y, Chen H X, Liu Y L, et al. The structural and diagenetic coupling controls the distribution of deep carbonate rock reservoirs in the southern of Tahe oilfield, Tarim Basin[J]. Geosciences, 2025, 15(11): 435.
[25]	Lucia F J. Rock-fabric/petrophysical classification of carbonate pore space for reservoir characterization[J]. AAPG Bulletin, 1995, 79(9): 1275-1300.
[26]	Wang Z Z, Njiekak G, Schmitt D R, et al. Empirical rock physics relationships on carbonate dry-frame elastic properties[J]. Petroleum Science, 2021, 18(3): 783-806.
[27]	Gupta H V, Kling H, Yilmaz K K, et al. Decomposition of the mean squared error and NSE performance criteria: Implications for improving hydrological modelling[J]. Journal of Hydrology, 2009, 377(1-2): 80-91.
[28]	Austin G I, Pe’er I, Korem T. Distributional bias compromises leave-one-out cross-validation[J]. Science Advances, 2025, 11(48): eadx6976.
[29]	Raissi M, Perdikaris P, Karniadakis G E. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations[J]. Journal of Computational Physics, 2019, 378: 686-707.
[30]	Chen Y T, Zhang D X. Physics-constrained deep learning of geomechanical logs[J]. IEEE Transactions on Geoscience and Remote Sensing, 2020, 58(8): 5932-5943.
[31]	Zhao L X, Liu J Y, Xu M H, et al. Rock-physics-guided machine learning for shear sonic log prediction[J]. Geophysics, 2024, 89(1): D75-D87.
[32]	Karpatne A, Atluri G, Faghmous J H, et al. Theory-guided data science: A new paradigm for scientific discovery from data[J]. IEEE Transactions on Knowledge and Data Engineering, 2017, 29(10): 2318-2331.
[33]	周长所, 袁俊亮, 丁智强, 等. 物理与工程先验融合的智能地层压力预测方法[J]. 石油科学通报, 2026, 11(2): 398-414.
	[Zhou C S, Yuan J L, Ding Z Q, et al. An intelligent formation pressure prediction method integrating physical and engineering priors[J]. Petroleum Science Bulletin, 2026, 11(2): 398-414.]
[34]	Koenker R W, Bassett G. Regression quantiles[J]. Econometrica, 1978, 46(1): 33-50.
[35]	Feng G, Yang Z, Xu X R, et al. Shear wave velocity prediction for fractured limestone reservoirs based on artificial neural network[J]. Geophysical Prospecting, 2024, 72(7): 2646-2665.
[36]	Fu X G, Wei Y H, Su Y, et al. Shear wave velocity prediction using long short-term memory network with generative adversarial mechanism[J]. PLoS ONE, 2025, 20(6): e0325271.
[37]	Zhang Y, Zhong H R, Wu Z Y, et al. Improvement of petrophysical workflow for shear wave velocity prediction based on machine learning methods for complex carbonate reservoirs[J]. Journal of Petroleum Science and Engineering, 2020, 192: 107234.
[38]	Deng Z Y, Zhou D S, Dong H Z, et al. Deep learning for predicting porosity in ultra-deep fractured-vuggy reservoirs from the Shunbei oilfield in Tarim Basin, China[J]. Scientific Reports, 2024, 14: 29605.
[39]	Zhang Y, Zhang C L, Ma Q Y, et al. Automatic prediction of shear wave velocity using convolutional neural networks for different reservoirs in Ordos Basin[J]. Journal of Petroleum Science and Engineering, 2022, 208: 109252.
[40]	Pearce T, Brintrup A, Zaki M, et al. High-quality prediction intervals for deep learning: A distribution-free, ensembled approach[C]. Proceedings of the 35th International Conference on Machine Learning, Stockholm, 2018.

井名	样本数N	井类别	AC-DTS相关性
A1	2769	高相关井	0.8771
A2	1157	高相关井	0.9610
A3	1711	中相关井	0.6060
A4	1804	中相关井	0.4239
A5	1413	中相关井	0.5332
A6	679	中相关井	0.5969

井名	样本数N	井类别	AC-DTS相关性
A1	2769	高相关井	0.8771
A2	1157	高相关井	0.9610
A3	1711	中相关井	0.6060
A4	1804	中相关井	0.4239
A5	1413	中相关井	0.5332
A6	679	中相关井	0.5969

编号	标定方案	参数来源	Castagna-only	固定双NN框架	主要说明
V02	前2井等权	训练井样本数前2井等权OLS	中位0.204 均值0.357 Pooled 0.784	中位0.516 均值0.565 Pooled 0.818	本文最终统一采用的折内物理先验
V00	固定参数	历史固定参数(1.16, 36.91)	中位0.353 均值0.382 Pooled 0.650	中位0.501 均值0.555 Pooled 0.772	历史固定参数，仅作对照
V10	加权	单井OLS后按加权	中位0.193 均值0.360 Pooled 0.762	中位0.486 均值0.539 Pooled 0.801	弱小井补偿
V16	GC石灰岩	Greenberg-Castagna石灰岩文献先验	中位0.091 均值0.152 Pooled 0.772	中位0.485 均值0.564 Pooled 0.816	外部文献先验
V06	加权	单井OLS后按加权	中位0.202 均值0.362 Pooled 0.779	中位0.478 均值0.535 Pooled 0.787	温和大井偏置
V03	前3井等权	训练井样本数前3井等权OLS	中位0.259 均值0.353 Pooled 0.749	中位0.472 均值0.555 Pooled 0.787	扩大到前3井
V09	等权平均	单井OLS后等权平均	中位0.199 均值0.362 Pooled 0.772	中位0.468 均值0.533 Pooled 0.811	不引入样本量偏置
V04	N²加权	单井OLS后按N²加权	中位0.209 均值0.365 Pooled 0.789	中位0.456 均值0.543 Pooled 0.805	强大井偏置
V11	1/N加权	单井OLS后按1/N加权	中位0.184 均值0.357 Pooled 0.747	中位0.456 均值0.543 Pooled 0.789	强小井补偿
V13	Huber池回归	5训练井样本池Huber回归	中位0.212 均值0.324 Pooled 0.746	中位0.449 均值0.531 Pooled 0.792	稳健回归，对异常点更宽容
V01	前1井等权	训练井样本数最大1井的单井OLS	中位0.216 均值0.372 Pooled 0.805	中位0.436 均值0.548 Pooled 0.827	最强单井主导
V14	RANSAC池回归	5训练井样本池RANSAC回归	中位0.325 均值0.326 Pooled 0.633	中位0.430 均值0.507 Pooled 0.769	稳健回归，对离群点更严格
V05	样本池OLS	5训练井样本池合并OLS	中位0.214 均值0.313 Pooled 0.752	中位0.425 均值0.511 Pooled 0.793	全样本池线性拟合
V15	GC白云岩	Greenberg-Castagna白云岩文献先验	中位0.021 均值-0.169 Pooled 0.708	中位0.418 均值0.523 Pooled 0.833	外部文献先验
V07	log(N)加权	单井OLS后按log(1+N)加权	中位0.200 均值0.362 Pooled 0.775	中位0.417 均值0.533 Pooled 0.813	对大井增益进一步压缩
V12	相关性加权	单井OLS后按max(r,0)加权	中位0.204 均值0.358 Pooled 0.771	中位0.416 均值0.529 Pooled 0.812	用AC-DTS相关性加权
V08	sigmoid加权	单井OLS后按sigmoid((N-μ)/σ)加权	中位0.208 均值0.365 Pooled 0.784	中位0.416 均值0.531 Pooled 0.805	非线性样本量加权

编号	标定方案	参数来源	Castagna-only	固定双NN框架	主要说明
V02	前2井等权	训练井样本数前2井等权OLS	中位0.204 均值0.357 Pooled 0.784	中位0.516 均值0.565 Pooled 0.818	本文最终统一采用的折内物理先验
V00	固定参数	历史固定参数(1.16, 36.91)	中位0.353 均值0.382 Pooled 0.650	中位0.501 均值0.555 Pooled 0.772	历史固定参数，仅作对照
V10	加权	单井OLS后按加权	中位0.193 均值0.360 Pooled 0.762	中位0.486 均值0.539 Pooled 0.801	弱小井补偿
V16	GC石灰岩	Greenberg-Castagna石灰岩文献先验	中位0.091 均值0.152 Pooled 0.772	中位0.485 均值0.564 Pooled 0.816	外部文献先验
V06	加权	单井OLS后按加权	中位0.202 均值0.362 Pooled 0.779	中位0.478 均值0.535 Pooled 0.787	温和大井偏置
V03	前3井等权	训练井样本数前3井等权OLS	中位0.259 均值0.353 Pooled 0.749	中位0.472 均值0.555 Pooled 0.787	扩大到前3井
V09	等权平均	单井OLS后等权平均	中位0.199 均值0.362 Pooled 0.772	中位0.468 均值0.533 Pooled 0.811	不引入样本量偏置
V04	N²加权	单井OLS后按N²加权	中位0.209 均值0.365 Pooled 0.789	中位0.456 均值0.543 Pooled 0.805	强大井偏置
V11	1/N加权	单井OLS后按1/N加权	中位0.184 均值0.357 Pooled 0.747	中位0.456 均值0.543 Pooled 0.789	强小井补偿
V13	Huber池回归	5训练井样本池Huber回归	中位0.212 均值0.324 Pooled 0.746	中位0.449 均值0.531 Pooled 0.792	稳健回归，对异常点更宽容
V01	前1井等权	训练井样本数最大1井的单井OLS	中位0.216 均值0.372 Pooled 0.805	中位0.436 均值0.548 Pooled 0.827	最强单井主导
V14	RANSAC池回归	5训练井样本池RANSAC回归	中位0.325 均值0.326 Pooled 0.633	中位0.430 均值0.507 Pooled 0.769	稳健回归，对离群点更严格
V05	样本池OLS	5训练井样本池合并OLS	中位0.214 均值0.313 Pooled 0.752	中位0.425 均值0.511 Pooled 0.793	全样本池线性拟合
V15	GC白云岩	Greenberg-Castagna白云岩文献先验	中位0.021 均值-0.169 Pooled 0.708	中位0.418 均值0.523 Pooled 0.833	外部文献先验
V07	log(N)加权	单井OLS后按log(1+N)加权	中位0.200 均值0.362 Pooled 0.775	中位0.417 均值0.533 Pooled 0.813	对大井增益进一步压缩
V12	相关性加权	单井OLS后按max(r,0)加权	中位0.204 均值0.358 Pooled 0.771	中位0.416 均值0.529 Pooled 0.812	用AC-DTS相关性加权
V08	sigmoid加权	单井OLS后按sigmoid((N-μ)/σ)加权	中位0.208 均值0.365 Pooled 0.784	中位0.416 均值0.531 Pooled 0.805	非线性样本量加权

分析层面	代表指标/变量	与固定双NN框架 LOWO均值R²的 Spearman ρ	与BiLSTM残差分支均值R²的 Spearman ρ	主要发现
纯基线拟合优度	Castagna-only LOWO均值R²	+0.260	+0.157	相关性较弱，单看纯基线R²不能解释固定双NN框架强弱
整体pooled 拟合	Castagna-only pooled R²	+0.221	+0.066	Pooled较高并不保证未见井固定双NN框架最优
相关性指标	Castagna-only平均Pearson r (各方案恒约为0.671)	近似常数	近似常数	各方案本质是AC的线性变换，相关性几乎不区分方案
综合一致性	Castagna-only LOWO均值KGE	-0.618	-0.662	“更像观测”的单基线不一定更利于残差学习
直接控制变量	BiLSTM残差分支LOWO均值R²	+0.924	—	固定双NN框架表现主要受残差分支质量控制
主成分结果	PC1/PC2解释率52.3%/40.2%	PC1 ρ=-0.150； PC2 ρ=-0.328	—	未发现单一“纯基线主成分”能强解释固定双NN框架，说明不存在单一“纯基线控制指标”

选择文件类型/文献管理软件名称

选择包含的内容

基于物理引导深度学习的超深碳酸盐岩储层横波时差预测方法

Physics-guided deep learning for shear-wave transit-time prediction in ultra-deep carbonate reservoirs

RichHTML

PDF

可视化

被引次数

摘要/Abstract

引用本文

使用本文

图/表 23

参考文献 40

相关文章 15

编辑推荐

Metrics

本文评价

配置	定义	LOWO 中位R²	LOWO 均值R²	Pooled R²	主要说明
V02 Castagna-only	仅DTS_cast = a×AC + b	0.204	0.357	0.784	V02折内top2标定的纯物理基线
BiLSTM残差分支(单分支)	仅保留BiLSTM残差分支	0.505	0.560	0.842	模型关键模块对比中的最强单分支
四分支等权平均	树模型残差分支，树模型直接预测分支，BiLSTM直接预测分支，BiLSTM残差分支各0.25	0.337	0.487	0.799	检验简单平均是否足够
固定双NN融合主模型	式(3)融合	0.516	0.565	0.818	最终统一主线口径
域自适应Top-2 门控探索线	四分支Top-2门控的探索口径	0.516	0.553	0.815	探索线，保留作对照但不作为最终主结果

井名	样本数	R²	RMSE/μs·ft^-1	MAE/μs·ft^-1	±5%命中率/ ±10%命中率	±5 μs·ft^-1命中率/ ±10 μs·ft^-1命中率
A1	2769	0.813	6.170	4.365	72.6%/91.6%	72.2%/89.2%
A2	1157	0.857	18.858	12.123	50.0%/69.9%	49.5%/62.3%
A3	1711	0.591	2.782	1.864	92.3%/98.6%	93.3%/98.5%
A4	1804	0.314	10.260	3.649	86.5%/90.2%	86.8%/90.2%
A5	1413	0.442	6.076	2.545	92.9%/97.7%	92.9%/97.5%
A6	679	0.375	14.476	7.638	52.9%/80.9%	54.9%/80.1%
Overall	9533	0.818	9.794	4.686	77.6%/90.1%	77.8%/88.4%

验证口径	训练/测试设置	组合数	中位R²	均值R²	加权/组合R²	RMSE/μs·ft^-1	MAE/μs·ft^-1	说明
LOWO	训练5井，测试1井，循环6次	6	0.516	0.565	0.598	9.77	5.36	主评价方式，反映单口未见井外推能力
LKWO3	训练3井，测试3井，共20种组合	20	0.690	0.682	0.684	10.29	5.05	多井留出补充验证，观察组合缺失条件下的整体表现
LKWO5	训练1井，测试5井，共6种组合	6	0.790	0.738	0.736	11.03	5.48	极少训练井条件下的补充观察，统计口径为组合级

窗口长度L/点	等效厚度/m	LOWO 中位R²	LOWO 均值R²	加权R²	Pooled R²	备注
4	0.5	0.501	0.538	0.567	0.756	短窗口
8	1.0	0.495	0.553	0.583	0.801	短窗口，整体次优
16	2.0	0.430	0.527	0.552	0.808	中短窗口
32	4.0	0.516	0.565	0.598	0.818	最佳综合表现
48	6.0	0.406	0.506	0.549	0.791	性能转折
64	8.0	0.359	0.483	0.518	0.802	过长窗口，整体退化
128	16.0	0.387	0.480	0.510	0.803	过长窗口，继续退化
256	32.0	0.402	0.470	0.487	0.787	超长窗口，井间差异增大
512	64.0	0.293	0.430	0.453	0.805	超长窗口，中位表现最低

分析内容	主要设置	关键结果	对主模型选择的影响
固定融合权重扫描	w=0.00~1.00(步长0.02)，按式 (3)，对w∈[0,1]扫描	w=0.72位于稳定高值区间0.68~0.98；均值最优在w=0.86、中位最优在w=0.82，但与0.72差异很小	支持0.72/0.28属平台区经验设置，而非单一尖锐最优
训练域内验证选权	外层LOWO，内层leave-one-training-well-out，严格后处理口径选权	外折选权范围0.30~0.60、均值0.42；训练域内验证选权均值/中位0.516/0.491，未全面稳定超过固定0.565/0.516	作为敏感性诊断与未来扩展方向，不升为主模型
域自适应 Top-2 门控	仅按训练井表现排序分支；另核验偏向A6的并列判定规则	理想门控(oracle)均值0.581、Pooled 0.846；域自适应门控均值0.553、Pooled 0.815；A6在域自适应门控下0.290，偏向A6规则下升至0.369	门控作为部署探索线；A6诊断有价值，但不足以替代主线
四分支全局权重搜索	树模型残差分支/树模型直接预测分支/ BiLSTM直接预测分支/ BiLSTM残差分支共同参与全局权重搜索	最优解退化为双NN：BiLSTM直接分支0.52、BiLSTM残差分支0.48、树模型两分支均为0，均值/中位0.546/0.510，与两分支最优{0.52,0.48}一致	树分支适合作候选与对照，而非固定主线融合核心
多随机种子	8个随机种子下的域自适应门控复验(0/7/21/32/42/64/84/123)	均值0.534±0.016，加权0.569±0.015，中位0.470±0.024，区间0.437~0.514	随机种子42可作代表结果，不写成唯一最优；低相关井会放大中位波动

模型	闭集均值R²	闭集加权R²	LOWO中位R² (最优对照)	LOWO均值R² (最优对照)	LOWO加权R² (最优对照)
MLP	0.847	0.868	0.274	0.380	0.397
HistGBR	0.842	0.862	0.317	0.394	0.403
ExtraTrees	0.823	0.841	0.306	0.418	0.438
LGB	0.819	0.838	0.299	0.391	0.406
CatBoost	0.799	0.816	0.308	0.393	0.404
RF	0.798	0.816	0.342	0.425	0.443
XGB	0.797	0.815	0.257	0.377	0.399
SVR	0.796	0.830	0.272	0.383	0.402
CNN-BiLSTM-Attention(主模型)	0.786	0.816	0.516	0.565	0.598
GBR	0.754	0.773	0.240	0.372	0.395
TabNet	0.703	0.736	0.287	0.414	0.434

代表井	主要误差表现	关键指标	可能地质控制因素	证据来源
A1(取心井)	趋势与幅度恢复良好，可作高质量井基准	R²=0.813、r=0.905	岩性单一、孔隙度低且集中、基质致密均质，响应域与训练井相近	4.3/4.4节+物性
A2	高DTS异常段持续低估／幅度压缩	R²=0.857、r=0.969、MAE=12.12 μs/ft	局部裂缝-孔洞组合及充填/流体差异，使该段幅度超出训练域(无取心)	图5及4.2节
A4(取心井)	振幅压缩(趋势可跟随、起伏恢复不足)	R²=0.314、KGE α≈0.53	岩性多样、孔隙度跨度大、局部孔洞-裂缝发育，储集空间类型纵向切换	4.3/4.4节+物性
A6	局部误差尖峰／误差集中	R²=0.375、β≈1.03、MAE=7.64 μs/ft	局部复杂非均质段超出训练井相似响应域(无取心)	图6及4.2节

派生参数	加权MAE	6井平均 MAE	与井级DTS 误差相关	工程含义
动态杨氏模量E/GPa	3.29	3.46	0.866	影响刚度估计与力学分层，是井壁稳定与压裂计算的输入
动态泊松比v	0.021	0.024	0.951	依赖纵横波速度比，直接影响应力分析与脆塑性判断
归一化脆性指数BI	10.39	11.03	0.951	宜作相对比较，不宜脱离地应力与静态标定单独决策

[1]	谭智鹏, 刘志斌, 王智琨, 张寓淇, 黄敏珂, 王博弘, 李敬法, 苏越. 基于有限元和深度学习的掺氢天然气缺陷管道失效压力预测[J]. 石油科学通报, 2026, 11(5): 1544-1556.
[2]	张飞, 戚蒿, 李丰清, 陈欣, 张晓东, 陈通, 潘艳芝. 基于DCGAN数据增强与CNN-LSTM混合模型的水下温压传感器故障诊断研究[J]. 石油科学通报, 2026, 11(4): 1190-1204.
[3]	李晓蓉, 苏飞宇, 李仨兴, 赵杨, 司晓宇, 冯永存. 基于分布式光纤声传感技术的水泥环流体泄漏工况诊断[J]. 石油科学通报, 2025, 10(6): 1330-1349.
[4]	张皓惟, 刘钰铭, 侯加根, 刘浩辰, 陈齐, 刘沛沛. 走滑断裂带差异变形及内部结构特征对碳酸盐岩缝洞型储层的控制机理研究—以塔里木盆地富满地区F_Ⅰ6断裂带为例[J]. 石油科学通报, 2025, 10(6): 1152-1166.
[5]	宋逸辰, 曾联波, 姚迎涛, 谭笑林, 毛哲, 曹东升, 龚飞. 走滑断裂控制的碳酸盐岩缝洞型储层非均质性及其影响因素[J]. 石油科学通报, 2025, 10(6): 1114-1129.
[6]	赵昭阳, 赵建国, 欧阳芳, 马铭, 闫博鸿, 张宇. 基于3D TransUnet模型的断层识别方法[J]. 石油科学通报, 2025, 10(5): 878-891.
[7]	孟翰, 张振鑫, 韩雪银, 林伯韬, 金衍. 融合钻头磨损与预训练机制的机械钻速预测方法研究[J]. 石油科学通报, 2025, 10(5): 1030-1046.
[8]	孙正心；金衍；孟翰；郭旭洋. 基于深度学习数据融合的测井数据精细表征[J]. , 2025, 10(1): 75-86.
[9]	曹倩雯；李维；林伯韬；金衍；韩雪银；张家豪. 融合旋转位置编码与掩码条件随机场的钻井工程命名实体智能识别方法[J]. , 2024, 9(5): 750-763.
[10]	王正；宋先知；李根生；潘涛；李臻；祝兆鹏. 基于MLP-CNN的固井质量智能评价方法[J]. , 2024, 9(5): 724-736.
[11]	胡迅；侯加根；刘钰铭. 多源断控岩溶型溶洞训练数据集构建和生成对抗网络三维建模应用[J]. , 2024, 9(3): 422-433.
[12]	杨柳青；王守东；杜宝强. 基于注意力机制的无监督学习地震数据随机和不规则噪声衰减方法[J]. , 2024, 9(1): 35-49.
[13]	高洋；孙郧松；王文闯；李国发. 基于层序统计结构和空间地质结构的深度学习高分辨率处理方法[J]. , 2023, 8(3): 290-302.
[14]	宋先知；姚学喆；李根生；肖立志；祝兆鹏. 基于LSTM-BP神经网络的地层孔隙压力计算方法[J]. , 2022, 7(1): 12-23.
[15]	刘珈铨；侯磊；毕新忠；段闯；任赆慈. 基于机器视觉的注水泵智能监控方法研究[J]. , 2021, 6(2): 282-291.

模态框（Modal）标题

选择文件类型/文献管理软件名称

选择包含的内容

基于物理引导深度学习的超深碳酸盐岩储层横波时差预测方法

Physics-guided deep learning for shear-wave transit-time prediction in ultra-deep carbonate reservoirs

RichHTML

PDF

可视化

被引次数

摘要/Abstract

引用本文

使用本文

图/表 23

参考文献 40

相关文章 15

编辑推荐

Metrics

本文评价