基于TPE优化的时空图神经网络油藏产量动态预测

doi:10.3969/j.issn.2096-1693.2025.03.019

石油科学通报 ›› 2025, Vol. 10 ›› Issue (5): 983-996. doi: 10.3969/j.issn.2096-1693.2025.03.019

基于TPE优化的时空图神经网络油藏产量动态预测

张博维¹^,²(), 刘月田²^,³^,^*(), 黄晋江²^,³, 薛亮²^,³, 宋来明⁴

1 中国免费靠逼视频(北京)人工智能学院，北京 102249
2 中国免费靠逼视频(北京)油气资源与工程全国重点实验室，北京 102249
3 中国免费靠逼视频(北京)石油工程学院，北京 102249
4 中海油研究总院有限责任公司，北京 100028

收稿日期:2025-03-06 修回日期:2025-06-18 出版日期:2025-10-15 发布日期:2025-10-21
通讯作者: *刘月田(1965年—)，博士，教授，主要从事渗流力学与油藏开发、油藏模拟与人工智能、裂缝性油藏渗流与开发方面的研究，lyt51@163.com。
作者简介:张博维(2000年—)，博士研究生，主要研究方向为人工智能在油气田开发中的应用研究，15304381793@163.com。
基金资助:
中国海洋石油有限公司联合研究院科技项目“基于流场适配性的砂岩油藏开发动态智能分析与评价方法”(CL2O22RCPS2018XNN)

Reservoir production dynamics prediction using TPE-optimized spatio-temporal graph neural networks

ZHANG Bowei¹^,²(), LIU Yuetian²^,³^,^*(), HUANG Jinjiang²^,³, XUE Liang²^,³, SONG Laiming⁴

1 College of Artificial Intelligence, China University of Petroleum, Beijing 102249, China
2 State Key Laboratory of Petroleum Resources and Engineering, China University of Petroleum,Beijing 102249, China
3 College of Petroleum Engineering, China University of Petroleum, Beijing 102249, China
4 CNOOC Research Institute, Beijing 100028, China

Received:2025-03-06 Revised:2025-06-18 Online:2025-10-15 Published:2025-10-21
Contact: *lyt51@163.com

摘要/Abstract

摘要：

油田开发过程中，准确的产量动态预测可以为油田生产措施调整、开发决策优化提供重要帮助。地下井网系统复杂的空间结构和动态随机的时变特征影响产量动态预测方法对注采井时空关系的有效学习，同时现有预测方法未能考虑多参数跨时间步的时空响应关系，导致井组多生产动态时序特征提取和关联分析存在局限性，制约产量预测精度的提升。本文考虑多节点多生产动态时序特征，建立时空图神经网络多井产量动态预测方法及基于树结构的贝叶斯算法的模型参数优化策略，有效聚合邻近节点多元信息，提高油藏产量预测精度和鲁棒性。该模型利用某海上水驱油藏生产数据验证。结果表明，优化后的模型精度较高，具有较好的产量趋势及置信区间预测；通过对比实验证明该模型可以有效实现多生产动态信息利用，提高预测精度，预测精度相较前人方法均方误差降低23.67%~56.96%，分位数损失函数降低18.31%~59.58%。研究成果可用于水驱油藏产量动态预测，为油藏生产决策提供可靠支持。

关键词: 产量预测, 多井预测, 时空图神经网络, 时空图建模, TPE优化策略, 概率预测

Abstract:

In the process of oilfield development, accurate production performance prediction can provide crucial support for adjusting production measures and optimizing development strategies. The complex spatial structure of underground well networks and the dynamic stochastic time-varying characteristics hinder the effective learning of spatiotemporal relationships between injection and production wells in existing prediction methods. Furthermore, current approaches fail to account for the cross-time-step spatiotemporal response relationships among multiple parameters, resulting in limitations in extracting temporal features and conducting correlation analysis of multi-well production performance sequences. These constraints ultimately restrict the improvement of production prediction accuracy. This study proposes a spatiotemporal graph neural network-based multi-well production forecasting method, incorporating a Tree-structured Parzen Estimator (TPE)-driven model parameter optimization strategy. The approach effectively aggregates multivariate information from neighboring nodes, enhancing reservoir production prediction accuracy and robustness. The model is validated using production data from an offshore waterflood reservoir. Results demonstrate that the optimized model achieves high accuracy, with improved production trend and confidence interval predictions. Comparative experiments confirm the model’s effectiveness in leveraging multi-dynamic information, significantly improving prediction accuracy. Specifically, the mean squared error is reduced by 23.67%~56.96%, and the quantile loss function decreases by 18.31%~59.58% compared to existing methods. The proposed framework provides reliable support for waterflood reservoir production forecasting and decision-making.

Key words: production prediction, multi-well Prediction, spatio-temporal graph neural networks, spatio-temporal graph modeling, TPE optimization strategy, probabilistic forecasting

中图分类号:

TE341
TP18

张博维, 刘月田, 黄晋江, 薛亮, 宋来明. 基于TPE优化的时空图神经网络油藏产量动态预测[J]. 石油科学通报, 2025, 10(5): 983-996.

ZHANG Bowei, LIU Yuetian, HUANG Jinjiang, XUE Liang, SONG Laiming. Reservoir production dynamics prediction using TPE-optimized spatio-temporal graph neural networks[J]. Petroleum Science Bulletin, 2025, 10(5): 983-996.

http://sykxtb.dgqinyehang.com/CN/Y2025/V10/I5/983

图/表 15

图1 时间特征提取模块结构示意图

Fig. 1 Schematic diagram of the time feature extraction module structure

图2 时空图卷积模型STGCN结构
注：图中V代表时空图数据，$V_{t} \in \mathbb{R}^{n \times C_{i}}$代表时间步t时n个节点的特征，C_i是数据的特征通道数，$E=\left\{w_{i, j}\right\}$表示图节点的邻接矩阵，w_i,j表示节点之间的连接关系，$\hat{V}_{t+i} \in \mathbb{R}^{n \times C_{0}}$代表时间步t+i时n个节点的特征，C_o_j是各模块内卷积核数量，j=1, 2, 3。

Fig. 2 Spatio-temporal Graph Convolutional Network (STGCN) model structure

图3 油藏各井生产动态时空信息图

Fig. 3 Dynamic spatio-temporal information map for spatial production

图4 TPE-STG模型参数优化策略

Fig. 4 Parameter optimization strategy

图5 参数组合与目标函数分布等高线图
注：图中组合损失函数值的大小通过蓝色深浅表示，损失函数值越小蓝色越深，代表不同参数组合下的模型性能越好；横纵坐标轴为待优化的超参数，每个子图为每2个超参数组合下的损失函数值分布，通过随时函数值的分布直观表现出当前组合下的。

Fig. 5 Contour plot of parameter combinations and objective function distribution

图6 参数组合与目标函数关系

Fig. 6 Relationship between parameter combinations and the objective function

表1 具体参数及调整范围表

Table 1 Specific parameters and adjustment range

参数	初次优化范围	精确优化范围
滑动窗口大小	10~32	10~12
批处理大小	8~128	8~32
初始学习率	1e^-3~1e^-6	1e^-3~1e^-4
权重衰减系数	1e^-3~1e^-6	1e^-3~1e^-4
梯度裁剪阈值	0.5~10	0.5,10
训练周期	20~100	60~100
激活函数选择	Elu,ReLU,LeakyReLU	ReLU,LeakyReLU
优化器	SGD,RMSgrop,Adam及其变式	Adam,RAdam
Lr衰减系数	0.1~1	0.6~1
Lr调整周期	10~50	10~30
动量因子	0~1	0~0.5

表2 基础油藏模型参数

Table 2 Parameter of the basic reservoir model

基础参数	范围	均值
初始压力/MPa	23.4~23.9	23.67
初始含水饱和度	0.2~1.0	0.393
水平方向渗透率/mD	0.5~498.8	162.04
垂直方向渗透率/mD	0.4~99.7	43.34
孔隙度	0.01~0.21	0.123
深度/m	2355~2416	2383.4

图7 基础油藏模型PUNQ-S3各井产量动态预测

Fig. 7 Production dynamic prediction for each well in the PUNQ-S3 reservoir model

图8 训练集与验证集产量动态预测损失函数对比

Fig. 8 Comparison of loss functions for production dynamic prediction between training and validation sets

表3 模型参数设置

Table 3 Model hyperparameter settings

参数	取值
GCN的神经元数量	32
GLU的神经元数量	32
FCN的神经元数量	64
滑动窗口大小	10
批处理大小	8
初始学习率	9.06e^-4
权重衰减系数	2.12e^-4
梯度裁剪阈值	0.5
训练周期	100
激活函数选择	leaky_relu, relu
优化器	RAdam
Lr衰减系数	0.3
Lr调整周期	50

图9 TPE-STG深度神经网络趋势及产量置信区间预测结果对比

Fig. 9 Comparison of TPE-STG deep neural network trend and yield confidence interval prediction results

图10 实际油藏各井产量动态预测损失函数对比

Fig. 10 Comparison of loss functions for production dynamic prediction in actual reservoirs

图11 累计产油量预测趋势及置信区间预测结果对比

Fig. 11 The cumulative oil production measurement trend and confidence interval prediction have been compared

图12 实际油藏产量动态预测多模型损失函数对比

Fig. 12 Comparison of multi-model loss functions for actual reservoir production dynamic prediction

参考文献 29

[1]	DU S Y, ZHAO X G, XIE C Y, et al. Data-driven production optimization using particle swarm algorithm based on the ensemble-learning proxy model[J]. Petroleum Science, 2023, 20(5): 2951-2966.
[2]	李荣光, 金龙, 孙伶, 等. 时间序列统计法预测中国石油石化领域大数据算法发展趋势[J]. 石油钻采工艺, 2024: 1-17.
	[LI R G, JIN L, SUN L, et al. Forecasting trends in big data algorithms for China’s petroleum and petrochemical sector: a time series statistical approach[J]. Oil Drilling & Production Technology, 2024: 1-17.]
[3]	张蕾, 窦宏恩, 王天智, 等. 基于集成时域卷积神经网络模型的水驱油田单井产量预测方法[J]. 石油勘探与开发, 2022, 49(5): 996-1004. doi: 10.11698/PED.20210825
	[ZHANG L, DOU H E, WANG T Z, et al. A novel integrated temporal convolutional neural network model for predicting single-well production in waterflooded oilfields[J]. Petroleum Exploration and Development, 2022, 49(5): 996-1004.]
[4]	ZHONG Z, SUN A Y, REN B, et al. A deep-learning-based approach for reservoir production forecast under uncertainty[J]. SPE Journal, 2021, 26(03): 1314-1340.
[5]	LI X, MA X, XIAO F, et al. Small-sample production prediction of fractured wells using multitask learning[J]. SPE Journal, 2022, 27(3): 1504-1519.
[6]	WANG L, YAO Y, LUO X, et al. A critical review on intelligent optimization algorithms and surrogate models for conventional and unconventional reservoir production optimization[J]. Fuel, 2023, 350: 128826.
[7]	SAGHEER A, KOTB M. Time series forecasting of petroleum production using deep LSTM recurrent networks[J]. Neurocomputing, 2019, 323: 203-213.
[8]	DU J, ZHENG J, LIANG Y, et al. A deep learning-based approach for predicting oil production: a case study in the United States[J]. Energy, 2024, 288: 129688.
[9]	薛亮, 顾少华, 王嘉宝, 等. 基于粒子群优化和长短期记忆神经网络的气井生产动态预测[J]. 石油钻采工艺, 2021, 43(4): 525-531.
	[XUE L, GU S H, WANG J B, et al. Dynamic production forecasting in gas wells using particle swarm optimization and long short-term memory neural networks[J]. Oil Drilling & Production Technology, 2021, 43(4): 525-531.]
[10]	王娟, 梅启亮, 邹永玲, 等. 基于多参数时间序列及粒子群优化算法的油藏产量动态建模预测方法[J]. 石油钻采工艺, 2023, 45(2): 190-196.
	[WANG J, MEI Q L, ZOU Y L, et al. Reservoir production dynamic modeling and prediction using multi-parameter time series and particle swarm optimization algorithms[J]. Oil Drilling & Production Technology, 2023, 45(2): 190-196.]
[11]	林伯韬, 郭建成. 人工智能在石油工业中的应用现状探讨[J]. 石油科学通报, 2019, 4(4): 403-413.
	[LIN B T, GUO J C. Discussion on current application of artificial intelligence in petroleum industry[J]. Petroleum Science Bulletin, 2019, 4(4): 403-413.]
[12]	ZHUANG X, WANG W, SU Y, et al. Multi-objective optimization of reservoir development strategy with hybrid artificial intelligence method[J]. Expert Systems with Applications, 2024, 241: 122707.
[13]	张雷, 杜立滨, 王聪, 等. 基于历史调控经验的强化学习油藏生产优化方法[J]. 科学技术与工程, 2024, 24(31): 13342-13350.
	[ZHANG L, DU L B, WANG C, et al. A reinforcement learning-based reservoir production optimization method using historical control experience[J]. Science Technology and Engineering, 2024, 24(31): 13342-13350.]
[14]	刘合, 李艳春, 贾德利, 等. 人工智能在注水开发方案精细化调整中的应用现状及展望[J]. 石油学报, 2023, 44(9): 1574-1586. doi: 10.7623/syxb202309014
	[LIU H, LI Y, JIA D, et al. Application status and prospects of artificial intelligence in fine-tuning water injection development plans[J]. Acta Petrolei Sinica, 2023, 44(9): 1574-1586.]
[15]	匡立春, 刘合, 任义丽, 等. 人工智能在石油勘探开发领域的应用现状与发展趋势[J]. 石油勘探与开发, 2021, 48(1): 1-11. doi: 10.11698/PED.2021.01.01
	[KUANG L, LIU H, REN Y, et al. Application status and development trends of artificial intelligence in petroleum exploration and development[J]. Petroleum Exploration and Development, 2021, 48(1): 1-11.]
[16]	李艳春, 贾德利, 王素玲, 等. 基于深度生成网络的时变井控下油藏动态预测代理模型[J]. 石油勘探与开发, 2024, 51(5): 1114-1125. doi: 10.11698/PED.20240269
	[LI Y, JIA D, WANG S, et al. A surrogate model for time-varying well control-based reservoir dynamic prediction using deep generative networks[J]. Petroleum Exploration and Development, 2024, 51(5): 1114-1125.]
[17]	胡晓东, 涂志勇, 罗英浩, 等. 拟合函数—神经网络协同的页岩气井产能预测模型[J]. 石油科学通报, 2022, 7(3): 394-405.
	[HU X, TU Z, LUO Y, et al. Shale gas well productivity prediction model with fitted function-neural network cooperation. Petroleum Science Bulletin, 2022, 7(3): 394-405.]
[18]	ATTANASI E D, FREEMAN P A, COBURN T C. Well predictive performance of play-wide and subarea random forest models for Bakken productivity[J]. Journal of Petroleum Science and Engineering, 2020, 191: 107150.
[19]	DONG Y, QIU L, LU C, et al. A data-driven model for predicting initial productivity of offshore directional well based on the physical constrained eXtreme gradient boosting (XGBoost) trees[J]. Journal of Petroleum Science and Engineering, 2022, 211: 110176.
[20]	RAJULA H S R, VERLATO G, MANCHIA M, et al. Comparison of conventional statistical methods with machine learning in medicine: diagnosis, drug development, and treatment[J]. Medicina, 2020, 56(9): 455.
[21]	LEE K, LIM J, YOON D, et al. Prediction of shale-gas production at duvernay formation using deep-learning algorithm[J]. SPE Journal, 2019, 24(06): 2423-2437.
[22]	TEMIZEL C, CANBAZ C H, SARACOGLU O, et al. Production forecasting in shale reservoirs using lstm method in deep learning[C]. 2020.
[23]	WERNECK R de O, PRATES R, MOURA R, et al. Data-driven deep-learning forecasting for oil production and pressure[J]. Journal of Petroleum Science and Engineering, 2022, 210: 109937.
[24]	AL-ALI Z A A H, HORNE R. Probabilistic well production forecasting in volve field using temporal fusion transformer deep learning models[C/OL]// Gas & Oil Technology Showcase and Conference. OnePetro, 2023[2024-01-11].
[25]	樊冬艳, 杨灿, 孙海, 等. 基于时间序列相似性与机器学习方法的页岩气井产量预测[J]. 中国免费靠逼视频学报(自然科学版), 2024, 48(3): 119-126.
	[FAN D Y, YANG C, SUN H, et al. Shale gas well production forecasting based on time series similarity and machine learning methods[J]. Journal of China University of Petroleum (Edition of Natural Science), 2024, 48(3): 119-126.]
[26]	韩江峡, 薛亮, 位云生, 等. 基于深度自回归神经网络的多井产量概率预测[J]. 石油科学通报, 2024, 9(4): 679-689.
	[HAN J X, XUE L, WEI Y S, et al. Multi-well production probability prediction based on deep autoregressive neural network[J]. Petroleum Science Bulletin, 2024, 9(4): 679-689.]
[27]	HUANG H, GONG B, SUN W. A deep-learning-based graph neural network-long-short-term memory model for reservoir simulation and optimization with varying well controls[J]. SPE Journal, 2023, 28(6): 2898-2916.
[28]	RAHMANIFARD H, GATES I. A Comprehensive review of data-driven approaches for forecasting production from unconventional reservoirs: best practices and future directions[J]. Artificial Intelligence Review, 2024, 57(8): 213. doi: 10.1007/s10462-024-10865-5 pmid: 39050688
[29]	SHERSTINSKY A. Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network[J]. Physica D: Nonlinear Phenomena, 2020, 404: 132306.

选择文件类型/文献管理软件名称

选择包含的内容

基于TPE优化的时空图神经网络油藏产量动态预测

Reservoir production dynamics prediction using TPE-optimized spatio-temporal graph neural networks

RichHTML

PDF

可视化

被引次数

摘要/Abstract

引用本文

使用本文

图/表 15

参考文献 29

相关文章 2

编辑推荐

Metrics

本文评价

[1]	从梦泽, 薛亮, 韩江峡, 苗得雨, 刘月田. 基于大语言模型的气井产量预测方法[J]. 石油科学通报, 2025, 10(5): 1056-1068.
[2]	韩江峡；薛亮；位云生；齐亚东；王军磊；陈海洋；刘月田. 基于深度自回归神经网络的多井产量概率预测[J]. , 2024, 9(4): 679-689.

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TPE优化的时空图神经网络油藏产量动态预测

Reservoir production dynamics prediction using TPE-optimized spatio-temporal graph neural networks

RichHTML

PDF

可视化

被引次数

摘要/Abstract

引用本文

使用本文

图/表 15

参考文献 29

相关文章 2

编辑推荐

Metrics

本文评价