深圳大学学报理工版

压裂井网气驱将油藏压裂与注气井网驱油结合，是当前致密低渗油藏提高采收率有效技术之一．水力裂缝及多相流动复杂性，使得基于精细油藏数值模拟的压裂井网气驱效果预测变得困难且耗时．提出一种基于均方根传播（rootmeansquarepropagation，RMSProp）深度学习的压裂井网气驱效果预测方法．通过建立压裂直井/水平井混合井网气驱数值模拟模型，引入高斯函数定量表征压裂水平井多级裂缝分布特征．利用正交试验筛选试验样本方案，自主编程实现数值模拟结果自动提取与数据处理，建立致密低渗油藏压裂井网气驱样本数据库．基于随机森林算法，筛选油藏地质、裂缝、生产等关键参数重要性特征，通过误差逆传播（backpropagation，BP）神经网络、长短期记忆单元（longshort-termmemory，LSTM）、双向长短期记忆单元（bi-directionallongshort-termmemory，BiLSTM）等深度学习算法，建立日产油、地层压力和采出程度预测代理模型，通过与油藏数值模拟对比，验证模型准确性．结果表明，BiLSTM算法在预测压裂井网气驱和压裂衰竭开发时效果最好.所提出的基于RMSProp的深度学习方法有效兼顾了模型实用性与精确性，为致密低渗油藏压裂井网气驱模拟预测提供了新途径．

Gas flooding in fractured well pattern is a combination of hydraulic fracturing and gas injection to displace reservoir oil within well pattern, and is one of the effective technologies for improving oil recovery in tight and low permeability reservoirs. However, due to the complexity of hydraulic fracturing and multi-phase flow, it becomes difficult and time-consuming to establish a fine reservoir numerical simulation to predict the performance of gas flooding in fractured well pattern. Therefore, we propose a deep learning method based on root mean square back propagation (RMSProp) for predicting the gas flooding effect of fracturing well pattern. Firstly, we establish a basic numerical simulation model of gas flooding in fractured vertical/horizontal hybrid well pattern in tight oil reservoir. Gaussian function curve is introduced to quantify the position of heel/toe of horizontal well and distributions of hydraulic fracture tips. The orthogonal test is then used to screen the test sample set, and the simulation results are auto-extracted and processed to build a database of gas flooding in tight oil reservoirs. The key parameters such as reservoir geology, fracture and production as input features are screened according to their importance in prediction by using the random forest algorithm. Then, we establish the prediction proxy models of daily oil production, reservoir pressure and field oil extraction (FOE) through deep learning algorithms such as back propagation (BP) neural network, long short-term memory (LSTM) and bi-directional long-short term memory (BiLSTM). Compared with the results of reservoir numerical simulation, the accuracy of the prediction proxy model is verified, and the results proves that the BiLSTM algorithm has the best performance in forecasting the gas flooding performance of fractured well pattern. The RMSprop-based deep learning method effectively takes into account the practicability and accuracy of the model andprovides a new approach to reservoir simulation and prediction of the fracturing well pattern gas flooding in tight and low permeability reservoirs.

引言
1 油藏数值模拟模型建立
2 深度学习预测模型建立
3 模型应用实例
结语

表1 模型基本参数Table 1 Basic parameters of the model

$图1 致密低渗油藏压裂井网注气开发数值模拟模型示意图Fig. 1 Schematic diagram of numerical simulation model of gas flooding in fractured well pattern in tight oil reservoirs.$

图1 致密低渗油藏压裂井网注气开发数值模拟模型示意图Fig. 1 Schematic diagram of numerical simulation model of gas flooding in fractured well pattern in tight oil reservoirs.

$图2 水平井水力裂缝分布示意图fracture tip of horizontal well (green dots).Fig. 2 Schematic diagram of horizontal well distribution. Gaussian function curve (dashed line) was introduced to calibrate the position of heel, toe and equivalent hydraulic$

图2 水平井水力裂缝分布示意图fracture tip of horizontal well (green dots).Fig. 2 Schematic diagram of horizontal well distribution. Gaussian function curve (dashed line) was introduced to calibrate the position of heel, toe and equivalent hydraulic

图3 单个LSTM神经元的运算Fig. 3 The operation of single LSTM neuron.

表2 试验变量及其取值范围Table 2 Test variables and the corresponding value range

$图4 水平井水力裂缝构建流程Fig. 4 Programming flow chart of horizontal well hydraulic fracture construction.$

图4 水平井水力裂缝构建流程Fig. 4 Programming flow chart of horizontal well hydraulic fracture construction.

图5 FOPR、FPR和FOE三种预测模型特征重要性暴风图Fig. 5 Characteristic importance storm chart prediction models of FOPR (blue column), FPR (red column), and FOE (green column).

表3 预测代理模型拟合结果Table 3 The fitting results of pridicted proxy models

图6 基于BiLSTM算法（a）FOPR代理模型、（b）FPR代理模型和（c）FOE代理模型真实值预测值对比Fig. 6 The comparison of true values (red solid line) with predicted values (green dashed line) of proxy models based on BiLSTM for (a) FOPR proxy model, (b) FPR proxy model, and (c) FOE proxy model.

[1]常晓宇. 2021年国内外油气行业发展报告[R].北京：国家高端智库中国石油集团经济技术研究院，2021. CHANG Xiaoyu. 2021 Domestic and foreign oil and gas industry development report [R]. Beijing: Institute of Economics and Technology of China National Petroleum Corporation, 2021. (in Chinese)
[2]房平亮，冉启全，鞠斌山.致密油藏压裂开发流固耦合数值模拟[J].中国矿业，2017，26（4）：140-145. FANG Pingliang, RAN Qiquan, JU Binshan. Coupled analysis of flow and deformation in tight oil reservoir fracturing and production simulation [J]. China Mining Magazine, 2017, 26 (4): 140-145. (in Chinese)
[3]LI Yuanjun, POPA A, JOHNSON A, et al. Dynamic layered pressure map generation in a mature water⁃flooding reservoir using artificial intelligence approach [C]// Proceedings of the SPE Western Regional Meeting. Garden Grove California, USA: Society of Petroleum Engi⁃neers, 2018: SPE-190042-MS. doi: 10. 2118/190042-MS Engineers, 2001: SPE-71345-MS. doi: 10. 2118/71345-MS
[4]ARPAT B G, CAERS J, HAAS A. Characterization of west-africa submarine channel reservoirs: a neural net⁃work based approach to integration of seismic data [C]//Proceedings of the SPE Annual Technical Conference and Exhibition. New Orleans, Louisiana: Society of Petroleum
[5]HEGEMAN P S, DONG C, VAROTSIS N, et al. Applica⁃tion of artificial neural networks to downhole fluid analysis [J]. SPE Reservoir Evaluation & Engineering, 2009, 12 (1): 8-13.
[6]POPA A S, PATEL A. Neural networks for production curve pattern recognition applied to cyclic steam optimiza⁃tion in diatomite reservoirs [C]// Proceedings of the SPE Western Regional Meeting. Bakersfield. California, USA:Society of Petroleum Engineers, 2012: SPE-153185-MS. doi: 10. 2118/153185-MS
[7]MAKHOTIN I, KOROTEEV D, BURNAEV E. Gradient boosting to boost the efficiency of hydraulic fracturing [J]. Journal of Petroleum Exploration and Production Tech⁃nology. 2019, 9(3): 1919-1925.
[8]LIU Kailei, XU Boyue, KIM C et al. Well performance from numerical methods to machine learning approach:applications in multiple fractured shale reservoirs [J]. Geofluids, 2021, 2021: 1-13. doi:10. 1155/2021/316945
[9]SONG Xuanyi, LIU Yuetian, XUE Liang, et al. Time-series well performance prediction based on long short-term memory (LSTM) neural network model [J]. Jour⁃nal of Petroleum Science and Engineering. 2020, 186:106682.
[10]KALAM S, ABU-KHAMSIN S A, AL-YOUSEF H Y, et al. A novel empirical correlation for waterflooding perfor⁃mance prediction in stratified reservoirs using artificial intelligence [J]. Neural Computing and Applications. 2021, 33(7): 2497-2514.
[11]PAL M. On application of machine learning method for history matching and forecasting of times series data from hydrocarbon recovery process using water flooding [J]. Petroleum Science and Technology. 2021, 39(15/16):519-549.
[12]DONG Peng, LIAO Xinwei, CHEN Zhiming, et al. An improved method for predicting CO2 minimum miscibility pressure based on artificial neural network [J]. Advances in Geo-Energy Research. 2019, 3(4): 355-364.
[13]张烈辉，贾鸣，张芮菡，等.裂缝性油藏离散裂缝网络模型与数值模拟[J].西南石油大学学报自然科学版，2017，39（3）：121-127. ZHANG Liehui, JIA Ming, ZHANG Ruihan, et al. Dis⁃crete fracture network modeling and numerical simulation of fractured reservoirs [J]. Journal of Southwest Petroleum University Science and Technology Edition, 2017, 39 (3):121-127. (in Chinese)
[14]袁彬，苏玉亮，丰子泰，等. 体积压裂水平井缝网渗流特征与产能分布研究[J]. 深圳大学学报理工版， 2013，30（5）：545-550. YUAN Bin, SU Yuliang, FENG Zitai, et al. Productivity distribution and flow characteristics of volume-fractured horizontal wells [J]. Journal of Shenzhen University Science and Engineering, 2013, 30(5): 545-550. (in Chi⁃nese)
[15]杨子清，陈文龙，杨军侠，等. 分段压裂水平井裂缝形态优化及产能特征研究[J]. 石油天然气学报， 2014，36（1）：99-103. YANG Ziqing, CHEN Wenlong, YANG Junxia, et al. The study on fracture morphology optimization and produc⁃tivity characteristics of staged fractured horizontal wells [J]. Journal of Oil and Gas Technology, 2014, 36(1): 99-103. (in Chinese)
[16]张剑.基于代理模型技术的高速列车性能参数设计及优化[D].成都：西南交通大学，2015. ZHANG Jian. The high-speed train performance parameter design and optimization based on surrogate model technology [D]. Chengdu: Southwest Jiaotong Uni⁃versity, 2015. (in Chinese)
[17]SIAMI-NAMINI S, TAVAKOLI N, NAMIN A S. The per⁃formance of LSTM and BiLSTM in forecasting time series [M]// BARU C, HUAN J, KHAN L, et al. IEEE Interna⁃tional Conference on Big Data. 2019: 3285-3292
[18]CALVETTE T, GURWICZ A, ABREU A C, et al. Fore⁃casting smart well production via deep learning and data driven optimization [C]//Proceedings of the Offshore Tech⁃nology Conference Brasil. Rio de Janeiro, Brazil: Society of Petroleum Engineers, 2019: OTC-29861-MS. doi:10. 4043/29861-MS
[19]GOODFELLOW I, COURVILLE A. Deep learning [M]. [S. l.]: MIT Press, 2016.
[20]ALIYUDA K, HOWELL J, HUMPHREY E. Impact of geological variables in controlling oil-reservoir perfor⁃mance: an insight from a machine-learning technique [J]. SPE Reservoir Evaluation & Engineering, 2020, 23(4):1314-1327.
[21]王文东，石梦翮，庄新宇，等.基于机器学习的井位及注采参数联合优化方法[J].深圳大学学报理工版， 2022，39（2）：126-133. WANG Wendong, SHI Menghe, ZHUANG Xinyu, et al. Joint optimization method of well location and injection-production parameters based on machine learning [J]. Journal of Shenzhen University Science and Engineering, 2022, 39(2): 126-133. (in Chinese)

备注

引言

1 油藏数值模拟模型建立

2 深度学习预测模型建立

3 模型应用实例

结语

期刊信息

备注

引言

1 油藏数值模拟模型建立

2 深度学习预测模型建立

3 模型应用实例

结 语

期刊信息

结语