Volume 29,Issue 1,2022 Table of Contents
2022, 29(1):1-10.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.001
Abstract:
Given the operation flow and hot issues about the exploration and development of Shengli Oilfield，this paper detailed the research and application progress in big data and artificial intelligence（AI）technologies. Through continuous conquest of technical challenges，intelligent technologies were developed for multiple application scenarios such as the fault detection，horizon extraction，lithology identification，and logging interpretation. The efficiency of fault interpretation was increased by more than 10 times，and the lithology identification accuracy of logging sand-shale could be more than 90%. In terms of oil and gas development，intelligent application methods for injection-production response identification，development index prediction，and intelligent scheme optimization were explored and implemented. The efficiency of scheme optimization was enhanced by more than 5 times. The results show that the big data and AI technologies，able to provide efficient multi-dimensional and multi-scale data analysis，can not only greatly improve work efficiency but also promote the accuracy of geological modeling and engineering prediction of reservoirs. Considering the current problems in field application，we will focus on conquering core algorithms，formulating sample data standards，expanding the sample library，and constructing intelligent application platforms in the next step. In this way，we intend to gradually implement the all-around and full-process intelligent application of exploration and development，promote the development of intelligent technologies for oil and gas exploration and development，and assist the oil and gas industry in improving quality and efficiency.
Given the operation flow and hot issues about the exploration and development of Shengli Oilfield，this paper detailed the research and application progress in big data and artificial intelligence（AI）technologies. Through continuous conquest of technical challenges，intelligent technologies were developed for multiple application scenarios such as the fault detection，horizon extraction，lithology identification，and logging interpretation. The efficiency of fault interpretation was increased by more than 10 times，and the lithology identification accuracy of logging sand-shale could be more than 90%. In terms of oil and gas development，intelligent application methods for injection-production response identification，development index prediction，and intelligent scheme optimization were explored and implemented. The efficiency of scheme optimization was enhanced by more than 5 times. The results show that the big data and AI technologies，able to provide efficient multi-dimensional and multi-scale data analysis，can not only greatly improve work efficiency but also promote the accuracy of geological modeling and engineering prediction of reservoirs. Considering the current problems in field application，we will focus on conquering core algorithms，formulating sample data standards，expanding the sample library，and constructing intelligent application platforms in the next step. In this way，we intend to gradually implement the all-around and full-process intelligent application of exploration and development，promote the development of intelligent technologies for oil and gas exploration and development，and assist the oil and gas industry in improving quality and efficiency.
2022, 29(1):11-20.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.002
Abstract:
Fine reservoir description，with rich content and massive accumulated data，provides a solid foundation for the application of big data technology. Meanwhile，the transformation and rapid development of fine reservoir description from digital to intelligent need support from big data technology. The characteristics of big data in fine reservoir description are expounded from aspects of basic statistical data and comprehensive data on research results. In addition to the above data tables，the big data in fine reservoir description also includes logging interpretation charts，seismic interpretation data volumes，geological models，and other data volumes，as well as various result maps. In light of scientific research practices，the application of big data technology in fine reservoir description is discussed from the perspectives of automatic fine stratigraphic division and correlation，automatic bulk discriminant classification of reservoir sedimentary microfacies（or reservoir architecture），bulk fine logging reinterpretation，comprehensive quantitative reservoir evaluation by cluster analysis，and multipoint geostatistical 3D geological modeling. The problems with the application of the big data technology in fine reservoir description include the database construction for big data technology，information mining for the big data technology，representativeness of big data，fusion of various types of big data，security of big data application，and expansion of big data application fields. Future development mainly involves the construction of the big data application platforms，optimization of information mining methods oriented towards big data technology，quality control of big data，innovation of visualization technology in the big data，standardized management of massive big data on oilfield development，and exploratory application of the big data technology in fine reservoir description of the unconventional oil and gas.
Fine reservoir description，with rich content and massive accumulated data，provides a solid foundation for the application of big data technology. Meanwhile，the transformation and rapid development of fine reservoir description from digital to intelligent need support from big data technology. The characteristics of big data in fine reservoir description are expounded from aspects of basic statistical data and comprehensive data on research results. In addition to the above data tables，the big data in fine reservoir description also includes logging interpretation charts，seismic interpretation data volumes，geological models，and other data volumes，as well as various result maps. In light of scientific research practices，the application of big data technology in fine reservoir description is discussed from the perspectives of automatic fine stratigraphic division and correlation，automatic bulk discriminant classification of reservoir sedimentary microfacies（or reservoir architecture），bulk fine logging reinterpretation，comprehensive quantitative reservoir evaluation by cluster analysis，and multipoint geostatistical 3D geological modeling. The problems with the application of the big data technology in fine reservoir description include the database construction for big data technology，information mining for the big data technology，representativeness of big data，fusion of various types of big data，security of big data application，and expansion of big data application fields. Future development mainly involves the construction of the big data application platforms，optimization of information mining methods oriented towards big data technology，quality control of big data，innovation of visualization technology in the big data，standardized management of massive big data on oilfield development，and exploratory application of the big data technology in fine reservoir description of the unconventional oil and gas.
2022, 29(1):21-29.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.003
Abstract:
Lithology identification is a vital basic work in the field of oil and gas exploration and development. Tight sandstone reservoirs suffer from complex rock composition，diverse lithology，and limited lithology identification by conventional logging. As machine learning is powerful in data analysis，this paper proposed a gradient boosting decision tree GBDT）algorithm with strong generalization ability to cut down large manpower and material resource consumption in lithology identification. Taking Chang7 Member of Yanchang Formation in Ordos Basin as the research object，it selects eight logging parameters including acoustic time difference（AC），natural gamma ray（GR），resistivity（RT），clay content（SH），natural potential（SP），effective porosity（POR），water saturation（Sw），and density（DEN）through sensitive analysis to build a lithology identification model based on GBDT algorithm. Actual data were applied to verify and analyze the application effect. The accuracy of the GBDT algorithm can reach 92%，compared with that of other methods such as naive Bayes，random forest，support vector machine，and artificial neural network for lithology identification. The high-precision lithology identification model based on the GBDT algorithm provides a new solution for tight sandstone reservoir evaluation.
Lithology identification is a vital basic work in the field of oil and gas exploration and development. Tight sandstone reservoirs suffer from complex rock composition，diverse lithology，and limited lithology identification by conventional logging. As machine learning is powerful in data analysis，this paper proposed a gradient boosting decision tree GBDT）algorithm with strong generalization ability to cut down large manpower and material resource consumption in lithology identification. Taking Chang7 Member of Yanchang Formation in Ordos Basin as the research object，it selects eight logging parameters including acoustic time difference（AC），natural gamma ray（GR），resistivity（RT），clay content（SH），natural potential（SP），effective porosity（POR），water saturation（Sw），and density（DEN）through sensitive analysis to build a lithology identification model based on GBDT algorithm. Actual data were applied to verify and analyze the application effect. The accuracy of the GBDT algorithm can reach 92%，compared with that of other methods such as naive Bayes，random forest，support vector machine，and artificial neural network for lithology identification. The high-precision lithology identification model based on the GBDT algorithm provides a new solution for tight sandstone reservoir evaluation.
2022, 29(1):30-36.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.004
Abstract:
Most oilfields in China have entered the late stage of development，and the conventional oil and gas reserves are gradually exhausted. Therefore，unconventional oil and gas reservoirs such as low-resistance oil layers have become important targets of exploration. In complicated fault-block reservoirs，affected by multiple factors such as sedimentary microfacies，structure，and interlayer interference，manual identification is inaccurate and inefficient，simply relying on expert experience. In this regard，big data mining technology was adopted. Firstly，the low-resistance oil layers were screened and verified with sub-layer data as a pointcut through the combination of logging data and research results；then the relationships between oil-bearing-related parameters of sub-layers were analyzed by the parallel association rule algorithm；finally，all sub-layers were classified by clustering analysis algorithm，and the similarity on the sub-layers containing verified low-resistance oil layers were calculated. As a consequence，the low-resistance oil layers were identified. The analysis of substantial data from an oil field in the eastern region shows that the accurate identification method of low-resistance oil layers driven by big data can tap the potential of low-resistance oil layers，with an accuracy rate of 90%. The potential reservoirs selected in the oilfield were put into production，with a great oil increment. This method saved massive manpower，reduced development costs，and enhanced oil recovery.
Most oilfields in China have entered the late stage of development，and the conventional oil and gas reserves are gradually exhausted. Therefore，unconventional oil and gas reservoirs such as low-resistance oil layers have become important targets of exploration. In complicated fault-block reservoirs，affected by multiple factors such as sedimentary microfacies，structure，and interlayer interference，manual identification is inaccurate and inefficient，simply relying on expert experience. In this regard，big data mining technology was adopted. Firstly，the low-resistance oil layers were screened and verified with sub-layer data as a pointcut through the combination of logging data and research results；then the relationships between oil-bearing-related parameters of sub-layers were analyzed by the parallel association rule algorithm；finally，all sub-layers were classified by clustering analysis algorithm，and the similarity on the sub-layers containing verified low-resistance oil layers were calculated. As a consequence，the low-resistance oil layers were identified. The analysis of substantial data from an oil field in the eastern region shows that the accurate identification method of low-resistance oil layers driven by big data can tap the potential of low-resistance oil layers，with an accuracy rate of 90%. The potential reservoirs selected in the oilfield were put into production，with a great oil increment. This method saved massive manpower，reduced development costs，and enhanced oil recovery.
2022, 29(1):37-45.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.005
Abstract:
Sedimentary facies modeling is an important part of reservoir modeling and there are many methods to build sedimentary facies models. Traditional modeling methods need to describe the spatial structure information of variables by various parameters such as variogram and data patterns，and then reproduce the spatial structure in the realizations. With different strategies，the reservoir modeling based on Generative Adversarial Nets（GANs）learns a large number of images（models）to generate the model possessing highly similar characteristics with the learning samples. enerative Adversarial Nets based on the single image（SinGAN）only need one image for training to generate highly similar images，improving the traditional GANs. With the sedimentary microfacies diagram of two layers in N gas field as an example，the corresponding sedimentary facies model is built. Compared with the classical multiple-point geostatistics method Simpat，the SinGAN method obtains more similar spatial structure of sedimentary microfacies with that depicted by training images and has a good application prospect.
Sedimentary facies modeling is an important part of reservoir modeling and there are many methods to build sedimentary facies models. Traditional modeling methods need to describe the spatial structure information of variables by various parameters such as variogram and data patterns，and then reproduce the spatial structure in the realizations. With different strategies，the reservoir modeling based on Generative Adversarial Nets（GANs）learns a large number of images（models）to generate the model possessing highly similar characteristics with the learning samples. enerative Adversarial Nets based on the single image（SinGAN）only need one image for training to generate highly similar images，improving the traditional GANs. With the sedimentary microfacies diagram of two layers in N gas field as an example，the corresponding sedimentary facies model is built. Compared with the classical multiple-point geostatistics method Simpat，the SinGAN method obtains more similar spatial structure of sedimentary microfacies with that depicted by training images and has a good application prospect.
2022, 29(1):46-52.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.006
Abstract:
The Oilfield M in the Middle East is a dominated by dual-medium carbonate reservoir with well-developed fractures and strong heterogeneity. However，the conventional single method cannot finely characterize the distribution of fractures. The multi-information fusion modeling technology for fractures based on the neural network firstly depends on the neural network to predict the fracture density in a single well without imaging logging data. Secondly，the nonlinear fusion of multiple pre-stack seismic attributes including P-wave azimuthal anisotropy and seismic discontinuity detection is performed on the basis of the neural network to predict the development probability of 3D fracture density. The fracture density of a single well is taken as hard data，and the fracture density model is constructed within the dual constraints of strict variogram analysis and fracture density probability volume. Finally，the discrete fracture network model is construced with the geostatistical modeling method，which is coarsened to be equivalent to the fracture attribute model. The model fitting rate is applied to the decision-making optimization for Oilfield M development.It is preferable to use horizontal wells or highly deviated wells for development in the areas with relatively developed fractures，and the average daily oil production per well reaches thousands of barrels. The fracture development revealed by the new well is consistent with the results of pre-drilling prediction，and the output of a single production well is significantly higher than that of a previous development well.
The Oilfield M in the Middle East is a dominated by dual-medium carbonate reservoir with well-developed fractures and strong heterogeneity. However，the conventional single method cannot finely characterize the distribution of fractures. The multi-information fusion modeling technology for fractures based on the neural network firstly depends on the neural network to predict the fracture density in a single well without imaging logging data. Secondly，the nonlinear fusion of multiple pre-stack seismic attributes including P-wave azimuthal anisotropy and seismic discontinuity detection is performed on the basis of the neural network to predict the development probability of 3D fracture density. The fracture density of a single well is taken as hard data，and the fracture density model is constructed within the dual constraints of strict variogram analysis and fracture density probability volume. Finally，the discrete fracture network model is construced with the geostatistical modeling method，which is coarsened to be equivalent to the fracture attribute model. The model fitting rate is applied to the decision-making optimization for Oilfield M development.It is preferable to use horizontal wells or highly deviated wells for development in the areas with relatively developed fractures，and the average daily oil production per well reaches thousands of barrels. The fracture development revealed by the new well is consistent with the results of pre-drilling prediction，and the output of a single production well is significantly higher than that of a previous development well.
2022, 29(1):53-61.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.007
Abstract:
Because of the strong heterogeneity and complex pore types of the extremely thick carbonate reservoir in M Formation of H Oilfield in Iraq，the applicability of conventional permeability logging interpretation models is poor. To solve this problem，this paper proposes a hybrid simulated annealing-genetic algorithm-random forest（SA-GA-RF）algorithm permeability evaluation model with conventional logging data and derived parameters. Depending on the analysis of logging response characteristics，the permeability sensitive curve is determined，and the permeability evaluation model based on the geophysical logging data is constructed by a random forest（RF）algorithm. The simulated annealing-genetic algorithm （SA-GA）is used to optimize the parameters in the RF model，which thus eliminates the influence of key parameters in the RF algorithm on the model accuracy. This method is applied to evaluate the permeability of the study block，and the prediction results are compared with those of RF and the improved back-propagation（BP）neural network. The results show that the SA-GA-RF model for the permeability evaluation of complex carbonate reservoirs can take full advantage of the response characteristics of the conventional logging curves and reflect the trend of logging curves changing with depth. Particularly，it has good applicability in carbonate reservoirs with strong heterogeneity. Compared with the improved BP neural network，the SA-GA-RF model has distinctly enhanced accuracy. The correlation between the core permeability and the prediction result is up to 0.83，which is 0.15 higher than the accuracy of permeability evaluation by RF.
Because of the strong heterogeneity and complex pore types of the extremely thick carbonate reservoir in M Formation of H Oilfield in Iraq，the applicability of conventional permeability logging interpretation models is poor. To solve this problem，this paper proposes a hybrid simulated annealing-genetic algorithm-random forest（SA-GA-RF）algorithm permeability evaluation model with conventional logging data and derived parameters. Depending on the analysis of logging response characteristics，the permeability sensitive curve is determined，and the permeability evaluation model based on the geophysical logging data is constructed by a random forest（RF）algorithm. The simulated annealing-genetic algorithm （SA-GA）is used to optimize the parameters in the RF model，which thus eliminates the influence of key parameters in the RF algorithm on the model accuracy. This method is applied to evaluate the permeability of the study block，and the prediction results are compared with those of RF and the improved back-propagation（BP）neural network. The results show that the SA-GA-RF model for the permeability evaluation of complex carbonate reservoirs can take full advantage of the response characteristics of the conventional logging curves and reflect the trend of logging curves changing with depth. Particularly，it has good applicability in carbonate reservoirs with strong heterogeneity. Compared with the improved BP neural network，the SA-GA-RF model has distinctly enhanced accuracy. The correlation between the core permeability and the prediction result is up to 0.83，which is 0.15 higher than the accuracy of permeability evaluation by RF.
2022, 29(1):62-68.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.008
Abstract:
The relationship between underground salt bodies and oil and gas reservoirs is inseparable. The accurate recognition of salt bodies is of great significance to the oil and gas reservoir exploration and drilling path planning. In existing deep learning methods，the size of the receptive field is unchanged，and the convolution kernel cannot be dynamically adjusted for the feature capture according to the size of the salt body in a seismic image. As a result，the part of the global information is ignored，which results in poor recognition at the boundaries or narrow areas of the salt bodies. In response to the above problems，this paper proposes a new method based on U-Net with SKNet as the encoder to extract salt body features.It has a dynamic selection mechanism that allows the sizes of the receptive fields to be adjusted adaptively according to multiple scales of the input information. In addition，it combines the position and channel self-attention mechanism and the hyper column method for feature fusion. The improved U-Net method is used to evaluate the TGS salt body data set. The recognition results have an intersection over union（IoU）of 85.66% and a pixel accuracy of 96.1%.
The relationship between underground salt bodies and oil and gas reservoirs is inseparable. The accurate recognition of salt bodies is of great significance to the oil and gas reservoir exploration and drilling path planning. In existing deep learning methods，the size of the receptive field is unchanged，and the convolution kernel cannot be dynamically adjusted for the feature capture according to the size of the salt body in a seismic image. As a result，the part of the global information is ignored，which results in poor recognition at the boundaries or narrow areas of the salt bodies. In response to the above problems，this paper proposes a new method based on U-Net with SKNet as the encoder to extract salt body features.It has a dynamic selection mechanism that allows the sizes of the receptive fields to be adjusted adaptively according to multiple scales of the input information. In addition，it combines the position and channel self-attention mechanism and the hyper column method for feature fusion. The improved U-Net method is used to evaluate the TGS salt body data set. The recognition results have an intersection over union（IoU）of 85.66% and a pixel accuracy of 96.1%.
2022, 29(1):69-79.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.009
Abstract:
The fault interpretation is a crucial step in oil and gas exploration and development. Due to the increase in the number of collected 3D seismic data volumes，the manual and traditional methods can hardly interpret faults in data volumes in detail. To better meet the urgent needs for high-efficiency，high-precision，and high-resolution fault interpretation in oil and gas exploration and development，we propose a deep learning-based algorithm to realize the automatic and intelligent fault detection with seismic data. The forward modeling method is used to generate a large number of diversified training data in line with the actual situation，and at the same time，a complete training sample volume is constructed in combination with the existing fault interpretation results. On this basis，an optimized and simple three-dimensional（3D）convolutional neural network（CNN）model is designed to efficiently process large 3D seismic data volumes and obtain accurate fault detection results. We further apply scan processing of matched filtering to the fault detection results to enhance the fault probability volumes and at the same time，obtain an estimation of fault strikes and dips. Given the three fault attribute volumes，we finally utilize a region-growing algorithm to automatically construct all the fault surfaces in the data volumes.Compared with the conventional methods commonly used in the industry，our method is significantly superior to the conventional methods in robustness to noise，accuracy，and efficiency.
The fault interpretation is a crucial step in oil and gas exploration and development. Due to the increase in the number of collected 3D seismic data volumes，the manual and traditional methods can hardly interpret faults in data volumes in detail. To better meet the urgent needs for high-efficiency，high-precision，and high-resolution fault interpretation in oil and gas exploration and development，we propose a deep learning-based algorithm to realize the automatic and intelligent fault detection with seismic data. The forward modeling method is used to generate a large number of diversified training data in line with the actual situation，and at the same time，a complete training sample volume is constructed in combination with the existing fault interpretation results. On this basis，an optimized and simple three-dimensional（3D）convolutional neural network（CNN）model is designed to efficiently process large 3D seismic data volumes and obtain accurate fault detection results. We further apply scan processing of matched filtering to the fault detection results to enhance the fault probability volumes and at the same time，obtain an estimation of fault strikes and dips. Given the three fault attribute volumes，we finally utilize a region-growing algorithm to automatically construct all the fault surfaces in the data volumes.Compared with the conventional methods commonly used in the industry，our method is significantly superior to the conventional methods in robustness to noise，accuracy，and efficiency.
2022, 29(1):80-89.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.010
Abstract:
Given the shear wave（S-wave）velocity prediction problem，the advantages and disadvantages of the empirical formula method and petrophysical modeling were analyzed，and the principle of s-wave velocity prediction was discussed.On this basis，this paper proposed a deep feedforward neural network（DFNN）for S-wave velocity prediction. Starting with the relationship between compressional wave（P-wave）and S-wave velocities，this study expounded the feasibility of applying the DFNN to S-wave velocity prediction and explained the principle of this deep learning method. Five reservoir parameters（acoustic time difference，density，neutron porosity，shale content，and porosity）were chosen for deep neural network training with S-wave velocity，and a reliable S-wave prediction model was thereby built. The model was applied to S-wave velocity prediction in different research areas，and the results show that DFNN-based S-wave velocity prediction achieves effectively improved prediction accuracy and efficiency and has a wide application range. It can provide reliable S-wave data for pre-stack amplitude-versus-offset（AVO）analysis and pre-stack inversion，so it is worth practical application and promotion.
Given the shear wave（S-wave）velocity prediction problem，the advantages and disadvantages of the empirical formula method and petrophysical modeling were analyzed，and the principle of s-wave velocity prediction was discussed.On this basis，this paper proposed a deep feedforward neural network（DFNN）for S-wave velocity prediction. Starting with the relationship between compressional wave（P-wave）and S-wave velocities，this study expounded the feasibility of applying the DFNN to S-wave velocity prediction and explained the principle of this deep learning method. Five reservoir parameters（acoustic time difference，density，neutron porosity，shale content，and porosity）were chosen for deep neural network training with S-wave velocity，and a reliable S-wave prediction model was thereby built. The model was applied to S-wave velocity prediction in different research areas，and the results show that DFNN-based S-wave velocity prediction achieves effectively improved prediction accuracy and efficiency and has a wide application range. It can provide reliable S-wave data for pre-stack amplitude-versus-offset（AVO）analysis and pre-stack inversion，so it is worth practical application and promotion.
2022, 29(1):90-97.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.011
Abstract:
Machine learning and data mining possess excellent abilities of prediction，analysis，decision-making，and calculation and have achieved good results in the field of oil and gas exploration and development. On the basis of summarizing the reservoir prediction methods，this paper analyzes the applicability，advantages and disadvantages of different reservoir prediction methods. It utilizes the machine learning algorithm to predict the rock type，spatial distribution，porosity，permeability，and oil saturation of the reservoir by mining logging and seismic data. This method reveals evident advantages compared with seismic inversion reservoir prediction：first，mining a large amount of information contained in seismic data and multi-attribute fusion can improve the prediction accuracy；second，data-driven instead of experience-driven can simplify the workflow.
Machine learning and data mining possess excellent abilities of prediction，analysis，decision-making，and calculation and have achieved good results in the field of oil and gas exploration and development. On the basis of summarizing the reservoir prediction methods，this paper analyzes the applicability，advantages and disadvantages of different reservoir prediction methods. It utilizes the machine learning algorithm to predict the rock type，spatial distribution，porosity，permeability，and oil saturation of the reservoir by mining logging and seismic data. This method reveals evident advantages compared with seismic inversion reservoir prediction：first，mining a large amount of information contained in seismic data and multi-attribute fusion can improve the prediction accuracy；second，data-driven instead of experience-driven can simplify the workflow.
2022, 29(1):98-106.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.012
Abstract:
As the exploration and development of oil and gas fields become increasingly difficult，higher requirements are put forward for the accuracy of sandstone lithology prediction. Geostatistical methods with high longitudinal resolution have been more widely used，while the insufficient reliability facing cross-well prediction grows more evident. This paper proposed a method of sandstone lithology prediction based on optimized machine learning algorithm and attribute feature. Firstly，sandstone characteristics in the seismic data and attribute volume were clarified through fine well-to-seismic calibration. Then，the optimal machine learning algorithm was selected after optimizing the attribute features and determining the sensitive logging curves. Next，the K-fold cross validation was used to obtain the optimal combination of hyperparameters.Finally，the training model with high prediction accuracy and robustness was obtained through multiple iterations. This method was applied to the sandstone lithology prediction of the 5th sand group in the Upper Guantao Formation in the eastern slope of Chengdao. Results showed that the coincidence of well points is high，and the predicted extension pattern of cross-well sandstone is consistent with the seismic data，which proves the reliability of cross-well prediction.
As the exploration and development of oil and gas fields become increasingly difficult，higher requirements are put forward for the accuracy of sandstone lithology prediction. Geostatistical methods with high longitudinal resolution have been more widely used，while the insufficient reliability facing cross-well prediction grows more evident. This paper proposed a method of sandstone lithology prediction based on optimized machine learning algorithm and attribute feature. Firstly，sandstone characteristics in the seismic data and attribute volume were clarified through fine well-to-seismic calibration. Then，the optimal machine learning algorithm was selected after optimizing the attribute features and determining the sensitive logging curves. Next，the K-fold cross validation was used to obtain the optimal combination of hyperparameters.Finally，the training model with high prediction accuracy and robustness was obtained through multiple iterations. This method was applied to the sandstone lithology prediction of the 5th sand group in the Upper Guantao Formation in the eastern slope of Chengdao. Results showed that the coincidence of well points is high，and the predicted extension pattern of cross-well sandstone is consistent with the seismic data，which proves the reliability of cross-well prediction.
2022, 29(1):107-112.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.013
Abstract:
Fault recognition from seismic data is crucial to the seismic data interpretation，but with the expansion of exploration scale，the traditional artificial fault interpretation cannot meet the actual production needs. How to develop a high-precision fault recognition method and improve the operation speed of the method is an urgent problem for those skilled in the art. Therefore，an automatic fault recognition method based on the improved AlexNet model is proposed to treat fault recognition as binary classification of image recognition. First，instead of local response normalization（LRN），batch normalization is used to accelerate the model convergence. Then，the balanced cross entropy loss is introduced to solve the problem of unbalanced height between the fault and the non-fault in seismic data，which makes the model converge in the right direction. Finally，the convolution layer is adopted to replace the full connection layer，which greatly reduces the training parameters and speeds up the training. The prediction results of the theoretical data and actual data of the training model show that the improved AlexNet model fully learns the fault features and has the ability to identify faults from seismic data.
Fault recognition from seismic data is crucial to the seismic data interpretation，but with the expansion of exploration scale，the traditional artificial fault interpretation cannot meet the actual production needs. How to develop a high-precision fault recognition method and improve the operation speed of the method is an urgent problem for those skilled in the art. Therefore，an automatic fault recognition method based on the improved AlexNet model is proposed to treat fault recognition as binary classification of image recognition. First，instead of local response normalization（LRN），batch normalization is used to accelerate the model convergence. Then，the balanced cross entropy loss is introduced to solve the problem of unbalanced height between the fault and the non-fault in seismic data，which makes the model converge in the right direction. Finally，the convolution layer is adopted to replace the full connection layer，which greatly reduces the training parameters and speeds up the training. The prediction results of the theoretical data and actual data of the training model show that the improved AlexNet model fully learns the fault features and has the ability to identify faults from seismic data.
2022, 29(1):113-120.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.014
Abstract:
There are many development wells in mature oilfields with multiple series of strata in the vertical direction and complex oil–water relationships. Thus，the manual interpretation of stratigraphic correlation has heavy workload and multiple solutions. In conventional research based on big data，a sample prediction model is established at one time with multiple logging curves and layered sample labels by a selected machine learning algorithm. However，this method has low accuracy and difficult convergence. To tackle the problems，this paper proposes a small layer intelligent division method based on data-driven and cyclic sliding time window. The logging curves sensitive to geological stratification are selected as the characteristic parameters. To enrich the sample database，the paper collects sample data many times with the“window-topoint” circular sliding time window method. The optimal traimodel is obtained by optimizing the hyper-parameters of different machine learning algorithms. The model is used to predict the results of small layer partitioning. Analysis results show that when the length of the sliding time window is 20 ning and the step size is 2，the small layer intelligent division model based on the random forest method has a prediction accuracy of 88.4%，which is better than the conventional prediction method based on modeling at one time and achieves the best test effect.
There are many development wells in mature oilfields with multiple series of strata in the vertical direction and complex oil–water relationships. Thus，the manual interpretation of stratigraphic correlation has heavy workload and multiple solutions. In conventional research based on big data，a sample prediction model is established at one time with multiple logging curves and layered sample labels by a selected machine learning algorithm. However，this method has low accuracy and difficult convergence. To tackle the problems，this paper proposes a small layer intelligent division method based on data-driven and cyclic sliding time window. The logging curves sensitive to geological stratification are selected as the characteristic parameters. To enrich the sample database，the paper collects sample data many times with the“window-topoint” circular sliding time window method. The optimal traimodel is obtained by optimizing the hyper-parameters of different machine learning algorithms. The model is used to predict the results of small layer partitioning. Analysis results show that when the length of the sliding time window is 20 ning and the step size is 2，the small layer intelligent division model based on the random forest method has a prediction accuracy of 88.4%，which is better than the conventional prediction method based on modeling at one time and achieves the best test effect.
2022, 29(1):121-127.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.015
Abstract:
As the intelligent application of oil and gas exploration and development matures and application scenarios increase，large-scale application is drawing nearer. As a result，the distributed storage，efficient collection，and parallel com? puting of samples have become urgent requirements of the intelligent application of oil and gas exploration and development. The intelligent application of seismic exploration is an important part of that of oil and gas exploration and development. In view of the large amount of single file data in and the unstructured characteristic of seismic exploration data，this paper analyzes the collection requirements for seismic exploration big data samples，proposes a solution of large file segmentation and merging based on the Hadoop distributed file system（HDFS），and implements redundant storage of seismic exploration data in three dimensions to improve the efficiency of seismic exploration sample collection. The experimental results show that the HDFS-based triple redundant storage solution can effectively improve the efficiency in collecting seismic exploration big data samples under rapid growth in data amount and therefore meet the requirements for intelligent application of seismic exploration.
As the intelligent application of oil and gas exploration and development matures and application scenarios increase，large-scale application is drawing nearer. As a result，the distributed storage，efficient collection，and parallel com? puting of samples have become urgent requirements of the intelligent application of oil and gas exploration and development. The intelligent application of seismic exploration is an important part of that of oil and gas exploration and development. In view of the large amount of single file data in and the unstructured characteristic of seismic exploration data，this paper analyzes the collection requirements for seismic exploration big data samples，proposes a solution of large file segmentation and merging based on the Hadoop distributed file system（HDFS），and implements redundant storage of seismic exploration data in three dimensions to improve the efficiency of seismic exploration sample collection. The experimental results show that the HDFS-based triple redundant storage solution can effectively improve the efficiency in collecting seismic exploration big data samples under rapid growth in data amount and therefore meet the requirements for intelligent application of seismic exploration.
2022, 29(1):128-136.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.016
Abstract:
On the basis of the convolutional neural network（CNN）of deep learning theory，the calculation methods for the absolute permeability of sandstone digital cores and related important influencing factors were discussed. The study selected three representative sandstone samples，including digital cores of Bentheimer sandstone，Berea sandstone，and Doddington sandstone. First，the permeability of these samples was calculated through the N-S equation and the pore network model separately，and the difference was compared. Then，we explored the influence of three different subsample sizes on the mean absolute permeability and permeability components in different directions when the sandstone samples were cut into subsamples. On this basis，the original sandstone digital cores were cut by the size of 200×200×200 to obtain subsamples，and all subsamples were subjected to microscopic seepage simulation and calculations to produce the corresponding absolute permeability. Finally，the digital core subsample database for deep learning was constructed. Given this database，we discussed the selection methods of key parameters for CNN system construction，such as learning rates and dropout rates.Upon training and learning，the subsamples of the testing set were tested，and the difference between the predicted values and the true values was within 5%，which proves the effectiveness of the method.
On the basis of the convolutional neural network（CNN）of deep learning theory，the calculation methods for the absolute permeability of sandstone digital cores and related important influencing factors were discussed. The study selected three representative sandstone samples，including digital cores of Bentheimer sandstone，Berea sandstone，and Doddington sandstone. First，the permeability of these samples was calculated through the N-S equation and the pore network model separately，and the difference was compared. Then，we explored the influence of three different subsample sizes on the mean absolute permeability and permeability components in different directions when the sandstone samples were cut into subsamples. On this basis，the original sandstone digital cores were cut by the size of 200×200×200 to obtain subsamples，and all subsamples were subjected to microscopic seepage simulation and calculations to produce the corresponding absolute permeability. Finally，the digital core subsample database for deep learning was constructed. Given this database，we discussed the selection methods of key parameters for CNN system construction，such as learning rates and dropout rates.Upon training and learning，the subsamples of the testing set were tested，and the difference between the predicted values and the true values was within 5%，which proves the effectiveness of the method.
2022, 29(1):137-144.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.017
Abstract:
This research aims to establish a prediction method for the initial productivity of directional wells in offshore sandstone oil reservoirs. A total of 17 factors affecting the initial productivity were considered from aspects of the geology，engineering，and development. Given 2 700 sets of data from 45 directional wells in offshore sandstone oil reservoirs，the Spearman correlation coefficient，random forest，and recursive feature elimination algorithm were combined to rank the importance of these influencing factors. With the logic of reservoir engineering，the main controlling factors in the initial productivity were selected. On this basis，the initial productivity prediction model was constructed by the extreme gradient boosting（XGBoost）algorithm，and its loss function was improved by referring to the productivity formula to enhance its physical constraints of this data mining algorithm. The results show that the main controlling factors affecting initial productivity of directional wells in offshore sandstone oil reservoirs include the formation flow coefficient，porosity，variation coefficient of the formation flow coefficient between layers，vertical depth of an electric submersible pump（ESP），perforation thickness of reservoirs，borehole size，drawdown，frequency of ESP，and choke size. Furthermore，the average relative error of the XGBoost algorithm with physical constraints for predicting the initial productivity of five wells is 9.68%，and the average relative error of the XGBoost algorithm without physical constraints is 11.68%. Therefore，physical constraints can effectively improve the accuracy of the XGBoost algorithm in productivity prediction，and the proposed method can realize the accurate prediction of the initial productivity of directional wells in offshore sandstone oil reservoirs.
This research aims to establish a prediction method for the initial productivity of directional wells in offshore sandstone oil reservoirs. A total of 17 factors affecting the initial productivity were considered from aspects of the geology，engineering，and development. Given 2 700 sets of data from 45 directional wells in offshore sandstone oil reservoirs，the Spearman correlation coefficient，random forest，and recursive feature elimination algorithm were combined to rank the importance of these influencing factors. With the logic of reservoir engineering，the main controlling factors in the initial productivity were selected. On this basis，the initial productivity prediction model was constructed by the extreme gradient boosting（XGBoost）algorithm，and its loss function was improved by referring to the productivity formula to enhance its physical constraints of this data mining algorithm. The results show that the main controlling factors affecting initial productivity of directional wells in offshore sandstone oil reservoirs include the formation flow coefficient，porosity，variation coefficient of the formation flow coefficient between layers，vertical depth of an electric submersible pump（ESP），perforation thickness of reservoirs，borehole size，drawdown，frequency of ESP，and choke size. Furthermore，the average relative error of the XGBoost algorithm with physical constraints for predicting the initial productivity of five wells is 9.68%，and the average relative error of the XGBoost algorithm without physical constraints is 11.68%. Therefore，physical constraints can effectively improve the accuracy of the XGBoost algorithm in productivity prediction，and the proposed method can realize the accurate prediction of the initial productivity of directional wells in offshore sandstone oil reservoirs.
2022, 29(1):145-151.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.018
Abstract:
Deep learning has been widely used in various aspects of oil and gas field development. Nevertheless，purely data-driven deep learning models suffer from large amounts of required data，unstable prediction ability，and weak generalization ability，and such models fail to consider the physical laws underlying the data. For dynamic reservoir pressure prediction，a deep neural network model for pressure field prediction driven jointly by reservoir seepage physics and data was constructed，and the mathematical seepage model of heterogeneous reservoirs was added to the loss function through regularization，which enabled the model to conform to both the results of data training and the constraint of the seepage physics equations. The results show that the jointly driven deep neural network model can achieve efficient learning and accurate prediction of pressure field data. Compared with the purely data-driven deep neural network model，the jointly driven model reduces the error（L2）between predicted and reference values by 93.1% and increases the decision coefficient（R2）by 20.3%.In the case of noisy observed data，the jointly driven model can still maintain high stability with strong noise resistance.
Deep learning has been widely used in various aspects of oil and gas field development. Nevertheless，purely data-driven deep learning models suffer from large amounts of required data，unstable prediction ability，and weak generalization ability，and such models fail to consider the physical laws underlying the data. For dynamic reservoir pressure prediction，a deep neural network model for pressure field prediction driven jointly by reservoir seepage physics and data was constructed，and the mathematical seepage model of heterogeneous reservoirs was added to the loss function through regularization，which enabled the model to conform to both the results of data training and the constraint of the seepage physics equations. The results show that the jointly driven deep neural network model can achieve efficient learning and accurate prediction of pressure field data. Compared with the purely data-driven deep neural network model，the jointly driven model reduces the error（L2）between predicted and reference values by 93.1% and increases the decision coefficient（R2）by 20.3%.In the case of noisy observed data，the jointly driven model can still maintain high stability with strong noise resistance.
2022, 29(1):152-159.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.019
Abstract:
The application of agent models based on deep learning is a new direction of oil and gas reservoir simulation.Given the huge time cost of high-precision full-order oil and gas reservoir simulation，this paper adopts an Embed to Control（E2C）model to construct a deep learning network through the architecture of“encoder+linear converter+decoder”. Data of the pressure field and saturation field at the original moment are integrated with well control constraints to generate the field data at a new moment. The PY35-1 Gas Field in the east of the South China Sea is discussed as an example to test the differences between the simulation results of the E2C model and those of the traditional numerical simulator. The test results show that the E2C model has smaller errors，with a relative error in the saturation field of less than 5% and an average relative error in the pressure field of 8%. Under the same CPU，the E2C model takes 16 s to run 100 cases，which is 375 times faster than the traditional numerical simulator（the time cost is 6 000 s）. In conclusion，the E2C model can greatly reduce the time cost under the condition of ensuring simulation precision.
The application of agent models based on deep learning is a new direction of oil and gas reservoir simulation.Given the huge time cost of high-precision full-order oil and gas reservoir simulation，this paper adopts an Embed to Control（E2C）model to construct a deep learning network through the architecture of“encoder+linear converter+decoder”. Data of the pressure field and saturation field at the original moment are integrated with well control constraints to generate the field data at a new moment. The PY35-1 Gas Field in the east of the South China Sea is discussed as an example to test the differences between the simulation results of the E2C model and those of the traditional numerical simulator. The test results show that the E2C model has smaller errors，with a relative error in the saturation field of less than 5% and an average relative error in the pressure field of 8%. Under the same CPU，the E2C model takes 16 s to run 100 cases，which is 375 times faster than the traditional numerical simulator（the time cost is 6 000 s）. In conclusion，the E2C model can greatly reduce the time cost under the condition of ensuring simulation precision.
2022, 29(1):160-167.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.020
Abstract:
The multi-scale complex fracture networks are formed by volume fracturing in tight oil reservoirs. At the same time，the uncertainty of reservoir parameters is significantly enhanced due to the strong heterogeneity. It is very important to accurately characterize fractures and reduce the uncertainty of reservoir parameters for building an accurate reservoir numerical simulation model. Therefore，a numerical simulation model for CO2 huff-n-puff in tight oil reservoirs was constructed according to the embedded discrete fracture model（EDFM）. By doing this，the multi-scale complex fracture network of tight oil reservoirs after fracturing was effectively characterized. With the ensemble Kalman filter（EnKF）method，the intelligent historical matching was carried out for CO2 huff-n-puff production data of tight oil reservoirs，and the physical properties and fracture parameters of reservoirs were estimated，which reduced the uncertainty of model parameters. The results show that the EDFM-based simulation model for tight oil reservoirs can accurately characterize multi-scale complex fracture networks and is suitable for treating complex fractures in tight oil reservoirs after fracturing. Pseudo-and real-vertices exist in the production curve calculated by the multi-scale medium model after effective medium treatment，which is consistent with the actual CO2 huff-n-puff production law of tight oil reservoirs. Upon repeated iterative matching with the EnKF intelligent history matching method，the convergence of the ensemble parameter curve of the initially realized multi-scale fractured medium model is enhanced，and the simulated production and production history of two horizontal wells，i.e.，Well J1 and Well J2，are well matched.
The multi-scale complex fracture networks are formed by volume fracturing in tight oil reservoirs. At the same time，the uncertainty of reservoir parameters is significantly enhanced due to the strong heterogeneity. It is very important to accurately characterize fractures and reduce the uncertainty of reservoir parameters for building an accurate reservoir numerical simulation model. Therefore，a numerical simulation model for CO2 huff-n-puff in tight oil reservoirs was constructed according to the embedded discrete fracture model（EDFM）. By doing this，the multi-scale complex fracture network of tight oil reservoirs after fracturing was effectively characterized. With the ensemble Kalman filter（EnKF）method，the intelligent historical matching was carried out for CO2 huff-n-puff production data of tight oil reservoirs，and the physical properties and fracture parameters of reservoirs were estimated，which reduced the uncertainty of model parameters. The results show that the EDFM-based simulation model for tight oil reservoirs can accurately characterize multi-scale complex fracture networks and is suitable for treating complex fractures in tight oil reservoirs after fracturing. Pseudo-and real-vertices exist in the production curve calculated by the multi-scale medium model after effective medium treatment，which is consistent with the actual CO2 huff-n-puff production law of tight oil reservoirs. Upon repeated iterative matching with the EnKF intelligent history matching method，the convergence of the ensemble parameter curve of the initially realized multi-scale fractured medium model is enhanced，and the simulated production and production history of two horizontal wells，i.e.，Well J1 and Well J2，are well matched.
2022, 29(1):168-174.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.021
Abstract:
Accurately identifying the main controlling factors of the fracturing performance in coalbed methane（CBM）wells and then effectively guiding the optimization of repeated fracturing schemes are the keys to improving the repeated fracturing productivity in CBM wells. Relying on the geological and engineering data in the research block，the feature selection based on Copula mutual information and cross-validation（CBFS-CV）algorithm was adopted to identify the main controlling factors that affect the fracturing performance. In combination with the gradient boosting regression model for productivity prediction and inspection，an improved identification algorithm was formed for CBM wells. This algorithm can effectively reduce the redundant features and increase correlation，and thus determine the optimal number of features. The results show that the coal structure，reservoir parameters（gas content，gas saturation，critical reservoir ratio），and operation displacement parameters（maximum operation displacement）are the three main controlling factors that affect the fracturing performance in the research block. The gradient boosting regression model verifies that the prediction coincidence rate of the main controlling factors identified by the CBFS-CV algorithm reaches 88%，which proves the effectiveness of the algorithm. Moreover，the main controlling factors of the typical well in this block were analyzed based on the above results，and the plugging removal solution with nitrogen foam was applied to the problems of poor coal structure and coal powder plugging. After field operation，the daily gas production increased from 288 m3/d to 805 m3/d，and the fracturing performance was significantly improved.
Accurately identifying the main controlling factors of the fracturing performance in coalbed methane（CBM）wells and then effectively guiding the optimization of repeated fracturing schemes are the keys to improving the repeated fracturing productivity in CBM wells. Relying on the geological and engineering data in the research block，the feature selection based on Copula mutual information and cross-validation（CBFS-CV）algorithm was adopted to identify the main controlling factors that affect the fracturing performance. In combination with the gradient boosting regression model for productivity prediction and inspection，an improved identification algorithm was formed for CBM wells. This algorithm can effectively reduce the redundant features and increase correlation，and thus determine the optimal number of features. The results show that the coal structure，reservoir parameters（gas content，gas saturation，critical reservoir ratio），and operation displacement parameters（maximum operation displacement）are the three main controlling factors that affect the fracturing performance in the research block. The gradient boosting regression model verifies that the prediction coincidence rate of the main controlling factors identified by the CBFS-CV algorithm reaches 88%，which proves the effectiveness of the algorithm. Moreover，the main controlling factors of the typical well in this block were analyzed based on the above results，and the plugging removal solution with nitrogen foam was applied to the problems of poor coal structure and coal powder plugging. After field operation，the daily gas production increased from 288 m3/d to 805 m3/d，and the fracturing performance was significantly improved.
2022, 29(1):175-180.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.022
Abstract:
The complicated geological conditions and the strong horizontal and vertical heterogeneity of the reservoir severe? ly restrict the water injection development efficiency of 22F Well Block in offshore Chengdao Oilfield，Shengli Oil Province.Accurately identifying the water injection situation of layers in a reservoir is an important prerequisite for reservoir management. It also has important guiding significance for the formulation of a reasonable water injection development plan. Therefore，a data-driven method is proposed for water injection profile prediction in this paper. The Extreme Gradient Boosting （XGBoost）algorithm is used to construct a model for making water injection profile prediction，with which the evolution of the water injection profile of each water injection well during the entire development period is predicted using the geological parameters and dynamic production data of the reservoir. As a result，high-quality data can be provided for rational production allocation and injection-production scheme adjustments. The application results in 22F Well Block of offshore Chengdao Oilfield show that the proposed method can accurately inverse and predict the water injection profile with an average relative error of 0.04，a determination coefficient of 0.87，and a root mean squared error of 3.12. Compared with the KH splitting method，the model in this paper yields a predicted value more consistent with the actual water absorption. This demonstrates that the proposed method can better reflect the actual water absorption of the reservoir and lays a solid foundation for fine stratified water injection and intelligent development of oilfields.
The complicated geological conditions and the strong horizontal and vertical heterogeneity of the reservoir severe? ly restrict the water injection development efficiency of 22F Well Block in offshore Chengdao Oilfield，Shengli Oil Province.Accurately identifying the water injection situation of layers in a reservoir is an important prerequisite for reservoir management. It also has important guiding significance for the formulation of a reasonable water injection development plan. Therefore，a data-driven method is proposed for water injection profile prediction in this paper. The Extreme Gradient Boosting （XGBoost）algorithm is used to construct a model for making water injection profile prediction，with which the evolution of the water injection profile of each water injection well during the entire development period is predicted using the geological parameters and dynamic production data of the reservoir. As a result，high-quality data can be provided for rational production allocation and injection-production scheme adjustments. The application results in 22F Well Block of offshore Chengdao Oilfield show that the proposed method can accurately inverse and predict the water injection profile with an average relative error of 0.04，a determination coefficient of 0.87，and a root mean squared error of 3.12. Compared with the KH splitting method，the model in this paper yields a predicted value more consistent with the actual water absorption. This demonstrates that the proposed method can better reflect the actual water absorption of the reservoir and lays a solid foundation for fine stratified water injection and intelligent development of oilfields.
2022, 29(1):181-189.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.023
Abstract:
Timely and accurate monitoring of the working conditions of oil wells is of great significance to the safe and efficient production of oilfields and the enhanced oil recovery. With the continuous deepening of oilfield informatization construction，real-time collection of dynamic monitoring data regarding oil well production such as indicator diagrams has been realized，and massive amounts of data have been accumulated and urgently need to be further explored and utilized. A new generation of artificial intelligence technology based on“big data+deep learning”is expected to break through the limitations of existing technologies and lead the upgrade of working condition diagnosis technology for oil wells. To this end，first，relying on more than 40 million sets of historical dynamic monitoring data covering oil wells in the different reservoirs，we prepared a large-scale dataset for working condition diagnosis of oil wells，which covered 5 categories and 37 different types of working conditions. On this basis，we selected the convolutional neural network algorithm and designed a personalized convolutional neural network（OWDNet）for working condition diagnosis of oil wells which contained more than 59 million learnable parameters in 26 layers. The OWDNet was trained using the above-mentioned working condition diagnosis dataset. After 10 epochs，the training accuracy was up to 99.7%，and the verification accuracy reached 98.9%. Furthermore，an intelligent working condition diagnosis system for oil wells was developed，and more than 5 million working condition diagnoses have been completed on site.The application accuracy of working condition diagnosis is 90%，and timely alarms are achieved. With this system，oil well production management and control were more reasonable and efficient，and working conditions of oil wells continued to improve. The proportion of continuous and stable production wells increased from 68% to 88%. The research provided a useful demonstration for the high-value application of oilfield big data.
Timely and accurate monitoring of the working conditions of oil wells is of great significance to the safe and efficient production of oilfields and the enhanced oil recovery. With the continuous deepening of oilfield informatization construction，real-time collection of dynamic monitoring data regarding oil well production such as indicator diagrams has been realized，and massive amounts of data have been accumulated and urgently need to be further explored and utilized. A new generation of artificial intelligence technology based on“big data+deep learning”is expected to break through the limitations of existing technologies and lead the upgrade of working condition diagnosis technology for oil wells. To this end，first，relying on more than 40 million sets of historical dynamic monitoring data covering oil wells in the different reservoirs，we prepared a large-scale dataset for working condition diagnosis of oil wells，which covered 5 categories and 37 different types of working conditions. On this basis，we selected the convolutional neural network algorithm and designed a personalized convolutional neural network（OWDNet）for working condition diagnosis of oil wells which contained more than 59 million learnable parameters in 26 layers. The OWDNet was trained using the above-mentioned working condition diagnosis dataset. After 10 epochs，the training accuracy was up to 99.7%，and the verification accuracy reached 98.9%. Furthermore，an intelligent working condition diagnosis system for oil wells was developed，and more than 5 million working condition diagnoses have been completed on site.The application accuracy of working condition diagnosis is 90%，and timely alarms are achieved. With this system，oil well production management and control were more reasonable and efficient，and working conditions of oil wells continued to improve. The proportion of continuous and stable production wells increased from 68% to 88%. The research provided a useful demonstration for the high-value application of oilfield big data.
2022, 29(1):190-196.
DOI: 10.13673/j.cnki.cn37-1359/te.2022.01.024
Abstract:
Kick is one of the most common accidents in drilling operations. If the kick monitoring and diagnosis are not timely，serious well control risks such as blowouts can be caused. Due to the massive monitoring data in drilling fields，directly using these data as the input of a kick diagnosis model will increase the complexity of the model and thus affect the accuracy of the model. Moreover，the labeling cost of kick samples is high in the process of constructing the diagnosis model. To solve these problems，this paper develops an intelligent diagnosis method for kick diagnosis，which is based on the kernel principal component analysis-semi-supervised extreme learning machine（KPCA-SSELM）. Firstly，KPCA is adopted to integrate the drilling parameters，and the principal components are extracted to reflect the core information of the original data. Then，the SSELM algorithm is employed for model training. Finally，the field drilling data is compared with the results of SSELM，KPCA-ELM，and other models to verify the effectiveness of the model. The results show that the proposed model based on KPCA-SSELM has a higher kick diagnosis rate and model generalization ability than other models. The semi-supervised learning method can make full use of the information contained in the unlabeled data to train the network when the number of labeled samples is relatively small，which can further improve the model performance and has a good application prospect.
Kick is one of the most common accidents in drilling operations. If the kick monitoring and diagnosis are not timely，serious well control risks such as blowouts can be caused. Due to the massive monitoring data in drilling fields，directly using these data as the input of a kick diagnosis model will increase the complexity of the model and thus affect the accuracy of the model. Moreover，the labeling cost of kick samples is high in the process of constructing the diagnosis model. To solve these problems，this paper develops an intelligent diagnosis method for kick diagnosis，which is based on the kernel principal component analysis-semi-supervised extreme learning machine（KPCA-SSELM）. Firstly，KPCA is adopted to integrate the drilling parameters，and the principal components are extracted to reflect the core information of the original data. Then，the SSELM algorithm is employed for model training. Finally，the field drilling data is compared with the results of SSELM，KPCA-ELM，and other models to verify the effectiveness of the model. The results show that the proposed model based on KPCA-SSELM has a higher kick diagnosis rate and model generalization ability than other models. The semi-supervised learning method can make full use of the information contained in the unlabeled data to train the network when the number of labeled samples is relatively small，which can further improve the model performance and has a good application prospect.
Online Submit
Reviewer
Editor
Reader
Quick Search
External Links
