Abstract
Nowadays, people’s demand for underground mineral resources is increasing, and geological disasters have occurred frequently in recent years. Geological disasters refer to geological effects or geological phenomena that are formed under the action of natural or man-made factors, causing loss of human life and property, and damage to the environment; such as landslides, collapses, mudslides, and ground subsidence. Under such a background, people must accelerate the exploration of complex geological structures. This paper is aimed at using the methods and concepts of deep reinforcement learning. Deep learning is to learn the inherent laws and representation levels of sample data. The information obtained during these learning processes is of great help to the interpretation of data such as text and images. In this way, the fine geology of complex fault-block reservoirs is modeled and studied. Geological structures and phenomena are discussed through convolutional neural network models and computer techniques. At the same time, the multitask bird recognition network is used to extract and classify geological images, so as to construct geological model maps with different spatial structures. Finally, the quality of the fault reconstruction model, the calculation of reservoir geological simulation reserves, and the evaluation of the water injection development effect of complex fault blocks are analyzed. In the evaluation of the development effect of water injection in complex fault blocks, comparing the relationship curve between the actual comprehensive water content and the oil recovery factor with the standard curve, the comprehensive water content of the initial block increased rapidly. Through timely and dynamic water allocation and comprehensive management, the water cut rising speed is controlled. The current comprehensive water cut of the reservoir is between 60% and 80%, the actual curve is between 25% and 35%, and the estimated waterflooding recovery is about 30%.
1. Introduction
The geological model of complex fault-block reservoirs should take the fine interpretation of faults as a breakthrough point. Reasonable combination of fault points, calculation of fault distance, and division of fault blocks are the basis for understanding complex fault block reservoirs and establishing 3D geological models. The complex fault block reservoir is one of many types of reservoirs. The characteristics of this type of reservoir are it not only exhibits strong reservoir heterogeneity (including affecting the integrity of injection and production between wells) but also the complexity of the nature of the boundary (including the adaptability of the well pattern) and the complexity of the oil-water system. The characteristics of fault-block reservoirs lead to large differences between complex fault-block reservoirs and ordinary reservoirs in terms of water injection development and water flooding effect evaluation.
Due to its geological structural characteristics and a series of special structures, the injection-production well pattern is imperfect, which affects the rapid energy consumption of the reservoir over a period of time, which is faster than the productivity decline of ordinary oil reservoirs. On the other hand, the lower water flooding the degree of control affects the macroscopic swept volume, leading to a faster rate of water cut. Second, the complex fault block reservoir contains dense natural fractures, which is difficult for water injection development, so it has the characteristics of low recovery. Therefore, modeling research on fine geology of complex fault-block reservoirs is becoming more and more important.
Geological modeling of complex fault block reservoirs has always been a hot and difficult research topic. Chen et al.’s classification is one of the hot issues in hyperspectral remote sensing research. In the past two decades, a large number of methods have been proposed to deal with the classification of hyperspectral data. He introduced the concept of deep learning to the classification of hyperspectral data for the first time. First, they verify the applicability of stacked autoencoders in accordance with the classic classification method based on spectral information. Second, a classification method based on spatial dominant information is proposed. Then, a new deep learning framework is proposed to fuse these two features to obtain the highest classification accuracy. The framework is a mixture of principal component analysis (PCA), deep learning architecture, and logistic regression. Specifically, as a deep learning architecture, stacked autoencoders are designed to obtain useful advanced functions. The experimental results of widely used hyperspectral data show that the classifier constructed under this deep learning-based framework has a good classification effect. However, most algorithms do not extract deep features hierarchically [1]. Xue et al.’s research object is a complex fault block reservoir provided by the China Petroleum Engineering Design Competition. Reservoir characteristics are recorded, including stratigraphic characteristics, vertical changes, and profile characteristics (thickness, sand percentage, and combined percentage). Through the comprehensive analysis of the structural pattern and reservoir characteristics, three-dimensional quantitative modeling of the reservoir was carried out on of geostatistics as the theoretical guide. Then, combining layering, structure, physics, and borehole trajectory data, a high-resolution layered reservoir model of the oilfield was established. The established 3D geological model integrates all well and structural information and provides a basic model for subsequent sedimentary microfacies modeling and physical property modeling. Finally, a series of cross-sections were established in sequence, namely, three-dimensional fence diagrams, connected well section cross-section diagrams, and well group section diagrams. However, its various complex geological construction maps need to be further studied [2]. In the field of X Tan 3D geological modeling, researchers often pay attention to modeling methods and workflow, but ignore the quantitative evaluation of the model. If the evaluation scope is narrowed to the same reservoir type, the comparability and practicality of quantitative evaluation will become apparent. The evaluation system should include three parts: data verification, geological understanding, and process inspection. Data verification is mainly to use actual data to test the accuracy of local predictions, and geological understanding is to test whether the global estimation of the model conforms to geological principles and prior insights. They are quantitative verification from different angles, complement each other, and are the key to quantitative evaluation. Process inspection is also a necessary condition to avoid contingency. Taking complex fault-block sandstone reservoirs as an example, based on feedback from experts in the petroleum industry, a multiparameter three-dimensional geological model quantitative evaluation criterion was established. However, the practical scope of these standards remains to be discussed and discussed [3]. Cui et al. propose a new advanced coupled multistable stochastic resonance method with two first-order multistable stochastic resonance systems, namely, CMSR, to detect motor bearing faults. The comparison with MSR shows that CMSR can achieve higher output signal-to-noise ratio. It is more beneficial to extract weak signal features and realize fault detection. At the same time, the method also has practical application value for engineering rotating machinery [4]. Zhang et al. proposed a three-partition state alphabet-based sequential pattern (Tri-SASP) for MTS. Experimental results on four real-world datasets show that (1) the discovered Tri-SASP and temporal rules can enrich human cognition. (2) Two three-partition strategies can bring us very meaningful and diverse Tri-SASP. (3) Both algorithms are efficient and scalable [5]. Mahmood et al.’s study discussed the effect of five different sizes of sand on the ultimate stress (MPa) of polymer-modified hand cement grouting sand using two different test standards (ASTM and BS). Based on dispersion index (SI), objective function (OBJ) evaluation, and training and testing datasets, the compressive strength of cement grouting sand can be well predicted using NLR and ANN models. The compressive strength tested using the BS standard is 71% higher than the compressive strength of the same mixture tested using the ASTM standard [6]. Firebaugh developed a platform for precise two-dimensional particle manipulation via acoustic forces from Chladni patterns and resonant microscale membranes. The project consists of two distinct phases: (i) macroscale manipulation in air using Chlarney plates; (ii) microscale manipulation in liquid using microscale membranes. The results show that the control method developed at the macroscale can be implemented and used with good precision and accuracy at the microscale [7].
The innovations of this paper are (1) through the methods and concepts of deep reinforcement learning, deep learning is to learn the inherent laws and representation levels of sample data, and the information obtained in these learning processes is of great help to the interpretation of data such as text, images, and sounds. Modeling and research on the fine geology of complex fault-block reservoirs. (2) Geological structures and phenomena are discussed through the convolutional neural network model and computer technology. (3) The multitask bird recognition network is used to extract and classify geological images to construct geological model maps with different spatial structures. (4) The quality of the fault reconstruction model, the calculation of reservoir geological simulation reserves, and the evaluation of the water injection development effect of complex fault blocks are analyzed.
2. Research Method for Fine Geological Modeling of Complex Fault Block Reservoirs Based on Deep Learning
2.1. Deep Learning
In recent years, driven by the rapid development of GPU-based parallel computing technology and the promotion of large-scale data sets, the method represented by deep learning has gradually made huge breakthroughs in multiple application fields of computers, especially in the field of computer vision and nature language processing field. In the field of computer vision, convolutional neural network (CNN) can well aim at the data characteristics of image modalities, establish the relationship between pixels, gradually extract the semantic information in the image, and finally realize the image content. For text sequence data, the method based on long short-term memory network (LSTM) is good at capturing the relationship between sequential data and extracting the inner meaning of the text sequence. Based on the above two mainstream modeling methods, a series of research on multimodal data has gradually become a research hotspot. Convolutional neural network (CNN) was first proposed by LeCun in 1998 and was successfully applied to handwritten digit recognition tasks on checks. Its network structure LeNet is shown in Figure 1.
As shown in Figure 1, the LeNet network contains basic elements such as convolution, downsampling, and nonlinear activation. The original image is convolutional operation to model the pixel relationship in the spatial neighborhood; the spatial resolution of the image is reduced through continuous downsampling operation, while the semantic information of the image is gradually abstracted; and under the action of the nonlinear activation function, the original image features are mapped to other feature spaces to enhance the expressive ability of the network [8]. The original image is input into the LeNet network, and through the abovementioned convolution, downsampling, and nonlinear operations, the recognition result of handwritten digits can be output finally. Although LeNet has a simple structure, it already contains the basic elements of convolutional neural networks, laying a solid foundation for the subsequent development of convolutional neural networks. Although LeNet has been used in handwritten digit recognition tasks in 1998, convolutional neural networks have not been applied on a large scale due to the development of computing power and the scale of image databases. It was not until the emergence of AlexNet in 2012 that the abovementioned status quo was changed [9, 10].
On the one hand, the emergence of AlexNet benefited from the substantial development of parallel computing capabilities supported by GPU hardware, which accelerated the training process of convolutional neural networks; on the other hand, the millions of large-scale image databases represented by ImageNet were used as volumes. The training of the product neural network provides a large amount of data, which promotes the network to be adequately trained, increases the feature expression ability of the network, and avoids the network from falling into an overfitting state due to the small data size [11].
Compared to LeNet, AlexNet has made the following improvements to the basic elements of convolutional neural networks. First of all, on the convolutional elements, AlexNet correspondingly increases the depth of the convolutional network and the number of channels of the convolutional layer [12]. This change has brought a huge improvement to the performance of the convolutional neural network, which is conducive to the convolutional neural network to learn more complex and expressive high-dimensional image features. Second, the maximum pooling operation (max pooling) is introduced on the downsampling method, which retains the maximum activation value of the image feature in the spatial neighborhood [13]. Finally, in the selection of the activation function, a linear rectification unit (ReLU) is used to replace the traditional sigmoid function, which avoids the phenomenon of gradient disappearance and reduces the corresponding amount of calculations and optimizes the problems in the training process as a whole.
And with the increase in the depth and complexity of the AlexNet network, a training strategy to deal with network overfitting is also proposed for the first time and provides a reference for the later training of deep convolutional neural networks [14, 15]. On the one hand, by introducing the data augmentation method (data augmentation), the original image is flipped, random cropping, and color dithering, which enriches the different manifestations of the same image; on the other hand, it proposes the dropout method, which is used in training. During this period, some neurons are randomly selected and cut out, and the value of the neuron is set to zero. The above training strategy can avoid the deep convolutional neural network from falling into an overfitting state due to too many parameters when the amount of data is limited [16, 17].
So far, AlexNet has opened the era of large-scale use of deep convolutional neural networks to model image data, and it has attracted the attention of a large number of researchers. How to better and more effectively express the inherent characteristics of images through sophisticated network structure design has become a research hotspot.
2.2. Complex Fault Block Reservoir
Complicated fault block reservoir is one of many types of reservoirs. It refers to the accumulation of oil and gas resources, etc., which are accumulated in the traps formed by complex and chaotic faults and surrounding rock structures. The characteristics of this type of oil reservoir are it not only exhibits strong reservoir heterogeneity, including affecting the integrity of injection and production between wells, but also exhibits complex boundary properties, including affecting the adaptability of well patterns and oil and water the complexity of the system [18]. Due to the wide existence of this type of reservoir, it is of great significance for the development of fault-block reservoirs, focusing on the characteristics of “small, broken, lean, scattered, and narrow” reservoirs and the difference between water injection development rules and development dynamic characteristics [19]. It is not difficult to investigate the differences in water injection development rules and development dynamic characteristics of large and medium-sized reservoirs in complex fault-block reservoirs. It is not difficult to conclude that the fluid flow rules, water injection development theory, microscopic water drive oil mechanism, and water drive efficiency under the same conditions are consistent. Second, there may be complex oil-water systems and strong heterogeneity of the reservoir, which will affect the adaptability of the injection-production well pattern [20, 21]. (1)Geological characteristics of complex fault block reservoirs
Complex fault blocks have a certain degree of heterogeneity in terms of geological characteristics. In terms of structural characteristics, due to the cutting and extension of fault blocks, the process of forming oil reservoir traps in the block has the characteristics of fine fragmentation, which is limited by boundary cutting and the scale of sand bodies. A series of factors affect [22, 23], and fault cutting has a great impact on the sand body distribution and sedimentary phase changes of the oil layer. It is difficult to form a complete fault block reservoir, and a certain degree of injection-production system makes the well pattern control not high, which affects the macroscopic swept volume, and thus it affects the recovery efficiency of fault-block reservoirs. Therefore, for complex fault-block reservoirs, the difference in geological factors represented by strong heterogeneity is the influence of fault-block reservoirs the main factors of development [24]. (2)Evaluation of water injection development effect
The current development of the block has entered a higher water-cut stage. The oilfield is facing development difficulties such as increasing difficulty in mining, high dispersion of remaining oil in fault blocks, relatively limited remaining recoverable reserves, deteriorating geological conditions, and accelerating production decline. Flooding has highlighted the contradictions in the development of the oilfield strata, the potential tapping effect has become poor, and the injection-production relationship has become complex. Therefore, how to control the growth of water cut, effectively develop the oilfield, and maintain a stable annual oil production has become a concern for the effect of water injection development. The evaluation of development effects includes evaluations for water injection development, such as water cut and water retention, as well as evaluations for changes in fluid production. It is not only for the entire region, but for the characteristics of fault block oil layer cutting and fine fragmentation, but also for different well groups. To evaluate the effect of water injection development [25, 26]. To develop oilfields by water flooding, it is necessary to continuously evaluate the development effect at different stages of development in order to put forward effective adjustment opinions. During the evaluation process, due to changes in mining conditions, well pattern planning, and different development stages, development indicators are also different [27].
Analyzing the change law of water cut during reservoir development is of great significance for evaluating the effect of oilfield development. The current methods for analyzing water content include water content gray GM (1,1) model, logistic cycle model, generalized usher model, and five models that characterize the relationship between water content and recovery degree (concave type, concave-S transition type, S type, S-convex transition type, and convex type). The main controlling factors affecting the rise of water cut are reservoir heterogeneity and the difference in water absorption status of water wells [28]. The main factors affecting the change law of water cut in low permeability oilfields are fluid properties (oil-water viscosity ratio) and reservoir physical properties (permeability, starting pressure). The lower the permeability, the greater the starting pressure gradient, the faster the relative permeability of the oil phase decreases, and the slower the increase of the water phase permeability [29].
Reasonable reservoir production pressure difference affects oil and fluid production capacity of oil wells. The indexes describing the production capacity of oil wells include oil production index, fluid production index, dimensionless oil production, fluid production index, rice oil production index, etc. The fluid production index is the ratio of the fluid production volume to the production pressure difference [30]. The dimensionless fluid production index refers to the ratio of the fluid production index at a certain stage of development to the fluid production index at the initial stage of oilfield development (under the condition of irreducible water saturation). Through the oil-water two-phase seepage law, we can deduce the dimensionless oil production and fluid production index at different water-bearing stages [31]. The change law of conventional oil reservoirs is with the increase of water cut, the dimensionless oil production and fluid production index all have an upward trend, which indicates that with development the low fluid recovery index is unfavorable to the stable production of oil reservoirs, and the effective driving effect of injected water is reduced [32].
Reservoir recovery factor is an important basis for formulating development plans, and it is of great significance to predict the recovery factor based on the actual situation of the reservoir. Before the oilfield was put into development, it mainly relied on less exploration data, core experiments, or analogy to the same type of reservoir, through analogy to the type of reservoir, reservoir fluid viscosity, porosity, permeability, well pattern density and permeability variation coefficient, etc. Parameters select appropriate values, and the empirical formula method is usually derived from the regression of the production materials in the study area, with a small scope of application and low reliability; in the middle and late stages of development, when the mine data is rich, it can pass type A, B, C, and D. The water drive law curve predicts the recovery factor. This method is widely applicable to water drive sandstone reservoirs. When the output of the oil field is in the decline stage in the later stage of development, the decline curve (exponential decline, hyperbolic decline, and harmonic decline) can also be used to predict the recovery rate. Tong’s chart method can also be applied to oil recovery prediction in the later stage of development. In addition, based on the establishment of a fine geological model of the study area, numerical simulation software is used to predict the recovery factor under the current mining conditions when the history matching meets the accuracy requirements. This method comprehensively considers the geological structure fluctuations, development factors, and reservoirs. The effect of physical properties and fluid properties, the prediction results are more accurate and reliable.
2.3. Multitask Bird Recognition Network (MTBN)
Multitask bird network (MTBN) is the joint learning of image recognition and image caption. On the one hand, the improvement of the quality of the underlying image features is conducive to simultaneously improving the accuracy of the image recognition task and the quality of the word generation in the picture-viewing task. On the other hand, the constraints of the joint learning of the top-level multitask can in turn continuously promote the underlying image features learning. And in the image caption task, we also considered the existence of individual text descriptions (individual text descriptions) and shared text descriptions (shared text descriptions). The specific description of the individual description text encourages each image to pay more attention to the learning of its own detailed features, making the image features more refined; and sharing the content of the description text allows the deep convolutional neural network to pay more attention to the confusing image pair. The differences between the details make the image features more distinguishable. Under the joint guidance of these two texts, the learned image features have strong detailed information and discriminative information at the same time, which is beneficial to promote the performance of image recognition tasks.
Convolutional neural networks need to learn from a large amount of data in order to extract key features more efficiently. Therefore, when training a convolutional neural network, it is necessary to learn from the implicit training data as much as possible, and the obtained accuracy rate is more convincing. It is the many advantages of convolutional neural network that make it widely used. We first operate on the image data through a deep convolutional neural network (CNN) to extract image features; then send the image features directly to the classifier module (classifier) to determine the category of the image. At the same time, for each picture, we send its image characteristics and the corresponding individual description text to the individual text processing module. For each pair of images in a confusing category, we send the pair of image features and the corresponding shared description text into the shared text processing module. These three modules will generate corresponding loss functions, then jointly constrain the learning process of the entire network, and jointly promote and improve the underlying image features. Below, we will describe the specific expansion of the abovementioned modules.
In the deep convolutional neural network (CNN) module, we use classic neural networks to extract image features. Convolutional network is composed of common convolution layer (convolution layer), batch normalization layer (batch normalization layer), pooling layer (pooling layer), rectification linear unit (Relu layer), and other basic units combined. The original image is sent to the deep convolutional network, and the semantic feature a of the image is gradually extracted through the function of the above basic unit, where a , , , and represent the length and width of the image feature, respectively, and high.
In the classifier module, we apply the global average pooling (GAP) operation to the image feature to remove the spatial relationship in the image feature and obtain the feature , with high-level semantic information. RD as shown in the following formula:
Among them, represents the value in the -th row and -th column of the image feature . The semantic feature obtained by the above operation is sent to the subsequent classification, and the predicted probability of image recognition is obtained, where , represents the number of image categories. Specific as shown in the following formula:
where represents the image classifier. Calculate the cross-entropy loss function value from the label predicted by the classifier and the real label of the image, as shown in the following formula:
Among them, represents whether the image belongs to the -th type of image, and the value is 0 or 1, and represents the probability that the image is predicted to be the -th type of image. In the individual text description module, we input the image feature into the long short-term memory network (LSTM) as the start of the sequence, output the predicted words of the network in turn, and finally form a description of the image content.
Suppose the input image feature , and the output word sequence , where
represents the length of the generated word sequence, represents the generated -th word, and represents the size of the vocabulary. At the same time, we decompose the image feature into the following form:
At time , it is necessary to predict and calculate the weight coefficient at that time, the original image feature is weighted by this weight, and the degree of contribution of the image feature to the predicted word of the LSTM module at different times is changed, as shown in the following formula:
Among them, and , respectively, represent the weight coefficients before and after normalization, and is obtained by normalizing by the softmax function. Where formula (6) represents that at time , the previous LSTM module hidden layer output and the vector of the image feature at position are input, and the fat module is transformed, and the output at the spatial position at time weight eti. We need to integrate the normalization layer into the current faster RCNN architecture. In our implementation, we follow ParseNet’s layer definition. There are also two ROI pooling layers to extract features from the third and fourth convolutional feature maps. The two ROI pooling layers, together with the original ROI pooling layer from the last, the fifth convolutional feature map, then independently pass the data through the normalization layer. Therefore, at time , the original image feature is weighted by the above weight coefficient to obtain the corresponding context feature vector , as shown in the following equation:
Among them, indicates that the image features of different spatial positions have different contributions to word prediction at time . With the aforementioned context feature vector , the hidden layer output ht corresponding to the LSTM module at time and the word predicted at that time can be predicted, as shown in the following formula:
Through the above operations, we can predict the corresponding word at each moment, compare it with the real word, and calculate the corresponding loss function , as shown in the following formula:
Among them, the first item in is the cross-entropy loss value calculated by comparing the predicted words at all moments with the real words. The second term is the constraint term. The weight value of the network at all moments in space position is as close to 1 as possible. With the constraints of the loss function, it can promote the continuous improvement of the quality of the underlying image features, so that the semantic content expressed by the image features can correspond to the individual description text, and the detailed information of the image features can be enhanced. In the shared text description module, we weighted the image features of the easily confusing category, respectively, and spliced them into a context vector and used the spliced context vector to guide the generation of the shared text sequence. Under the supervision of the shared description text, it helps the network distinguish the differences in the image features of the easily confusing categories. At time , the weight coefficients and corresponding to the image feature are calculated, respectively, , as shown in the following formula:
Among them, is the weight coefficient normalized by the softmax function corresponding to the image feature . Among them, , share the fat module. After calculating the weight coefficient of the corresponding image feature, the corresponding context feature at time can be calculated, as shown in the following equation:
Among them, means joining the weighted features of and to form the context feature . With the context feature at time , the output of the hidden layer of the LSTM module at time and the predicted word can be obtained, and finally, the corresponding loss function value is calculated, as shown in the following equation:
The image features are extracted through the underlying deep convolutional neural network, and the image features are sent to the classifier module, the individual description text module, and the shared description text module, respectively; under the joint drive of these three task-driven tasks, improvement from top to bottom. The quality of image features, thereby ultimately improving the accuracy of image recognition. Optimizing the loss function of the entire framework can finally be jointly optimized through the loss function described by the following equation.
3. Fine Geological Construction Model of Complex Fault Block Reservoir Based on Deep Learning
3.1. Fine Structural Modeling of Complex Fault Block Reservoirs
3.1.1. Structural Modeling Ideas
The essence of 3D geological modeling is to reconstruct, reproduce, and abstract the geological bodies, geological phenomena, and geological processes in the natural world. To achieve this goal, 3D visualization technology must be used. With the essence of 3D geological modeling, it can be seen that the 3D model of geological body should retain the structural characteristics of geological body. These features include the spatial form, location, and distribution of geological body. It should also have complete topological relations and geological semantics. The theory and methods of 3D geological modeling mainly include five aspects, namely, standardization of geological source data, 3D visualization data model of geological bodies, structural model methods of geological bodies, management of geological attribute data, and visualization of geological graphics.
3.1.2. Basic Process
In order to get a better modeling effect, the 3D attribute data volume line can be preprocessed before the isosurface is extracted. The main way of preprocessing is to perform interpolation processing. The purpose of interpolation preprocessing is to obtain more data points with attribute values and coordinate values in the three-dimensional space. Based on the above research, the modeling process is as follows:
(1) Obtain the profile contour line. This paper mainly studies the construction of complex geological bodies based on the contour line method, so the input of the algorithm is a series of profile contour lines. (2) Construct a four-dimensional function. A four-dimensional function is constructed for each point on the plane where the profile contour line is located, and the coordinates of each point and its attribute value are obtained. (3) Calculate the attribute value. The calculation of the attribute value is mainly to obtain the symbol of the attribute value by judging the position relationship between the point and the contour line and calculate the distance from the point to the contour line to obtain the value of the attribute value. (4) Generate the attribute data body. Calculate the attribute value for the points of a certain plane contour line to obtain a two-dimensional attribute data body, and calculate the attribute value for all the plane contour points, then, the two-dimensional attribute data body is raised to three-dimensional, and a three-dimensional attribute data body is obtained. (5) Interpolation preprocessing. Not all points in the three-dimensional attribute data body have attributes. In order to get a better modeling effect, the attribute data volume needs to be interpolated. Commonly used interpolation algorithms include inverse distance weighted interpolation, B-spline interpolation, kriging interpolation, and so on. (6) Extract the isosurface. Extract the isosurface from the attribute data volume, and use CGAL to complete the extraction of the isosurface. The drawing of isosurface should first find the intermediate primitives, then triangulate the intermediate primitives to generate a series of triangles, and connect the triangles to generate the isosurface. (7) Generate geological blocks. The extracted isosurface constitutes the geological block model, completing the structure of the complex body model.
3.2. Construction of Four-Dimensional Function
For each point on the plane where the profile contour line is, an attribute value needs to be given to generate the attribute data body. The attribute value of each point is mainly generated by constructing a four-dimensional function, that is, for each point, the coordinate value of the point and the attribute value of the point should be obtained. For each point on the plane where the profile contour line is located, a field function is defined.
The attribute value of each point on the plane where the contour line is located is the value of the field function . It can be seen from the definition of the field function that if the point is above a certain contour line, the attribute of the point. The value is 0; if the point is within a certain contour line, the attribute value of the point is less than 0; if the point is outside all contour lines, the attribute value of the point is greater than 0.
3.3. Determination of Attribute Value
After constructing the four-dimensional function, it is necessary to determine its attribute value for each point on the plane where the profile contour line is located. In this paper, the distance function is used to construct a four-dimensional function, so the determination of the attribute value of each point is transformed into the calculation of the distance function . According to the definition of the distance function, the attribute value inside the contour is less than 0, the attribute value outside the contour is greater than 0, and the attribute value above the contour is equal to 0. Therefore, to determine the attribute value of a point, you should first determine the relationship between the point and the contour line and then calculate the distance from the point to the contour line.
3.4. Generate Attribute Body Data
In this paper, the key to the complex closed geological body structure based on attribute volume data is the generation of attribute volume data. The generation effect of volume data directly affects the modeling effect of the geological block after the final isosurface extraction. In this article, the generation of attribute volume data is only achieved by constructing four-dimensional functions.
After constructing a four-dimensional function with a given distance function, it is necessary to obtain the attribute value of the point by calculating the distance function. Among them, by judging the relationship between the point and the contour line, the sign of the attribute value of the point is judged; by calculating the distance between the point and the contour line, the value of the attribute value of the contour line is obtained. Judge the positional relationship between the point and the contour line and calculate the distance between the point and the contour line. Then, the coordinate value and attribute value of the point on the plane where the profile contour line is located are determined, and the attribute volume data is successfully generated. Starting from 3D geological modeling, the geological body modeling of complex fault block reservoirs has been studied in depth, and through the constructed method of contour surface extraction of attribute body data, the problem of multiple bifurcations existing between contours can be exploited to achieve better modeling results.
4. Fine Geological Modeling of Complex Fault Block Reservoirs Based on Deep Learning
4.1. Quality of Fault Reconstruction Model
In the research on the reconstruction of 3D geological bodies, the construction of simple layered geological bodies is relatively mature, but due to the multiple value problems that may exist in complex geological interfaces and the sparse original data of geological modeling, it is impossible to directly use general point cloud weights. The geological interface method requires different reconstruction methods according to different geological conditions. The existing three-dimensional fracture geological model construction methods mainly include “local interpolation method” and “global interpolation method.”
As shown in Figure 2, the overall method is used to regrid the inverse fault test model. First, the three-dimensional uniform sampling points as shown in the figure are calculated, and these sampling points are directly triangulated (Figure a) and interpolated. (Figure b). In the result image, it can be observed that there are more obvious intersections and sharp grids at the locations where the elevation values on both sides of the fault are significantly different, and the regriding effect is not ideal.
(a)
(b)
(c)
(d)
4.2. Reserve Calculation of Reservoir Geological Modeling
Through the abovementioned reservoir modeling process, different simulation realizations can be performed to obtain the reservoir attribute parameter models realized by different simulations of the Sanjianfang Formation in the Ling 2 District of the Qiuling Oilfield, the effective reservoirs can be screened, and the geological construction of the research block can be calculated separately. The reserves of each layer under different simulation models are shown in Figure 3. It can be seen that the average crude oil reserves calculated by each model are relatively close, and the error is small.
4.3. Evaluation of the Effect of Water Injection Development in Complex Fault Blocks
The influence of crude oil underground viscosity: according to the methods and methods of water injection development at home and abroad, when the viscosity of local crude oil is greater than 5 mPa·s, it will have an impact on conventional water injection development. If it is greater than 70 mPa·s, there will be great difficulties in water injection development. By comparing the successful experience of waterflooding development at home and abroad, it is believed that if the reservoir chooses waterflooding development for perfection, its underground crude oil viscosity should not exceed 190 mPa.s.
As shown in Figure 4, under the same interfacial tension condition, when the viscosity of crude oil increases, the corresponding oil displacement efficiency during the production process also decreases; the oil displacement efficiency is inversely proportional to the oil-water viscosity ratio, and the residual oil displacement efficiency is inversely proportional to the oil-water viscosity ratio. The oil saturation can be seen from the figure showing an upward trend. The increase in residual oil saturation and the decrease in oil displacement efficiency are relatively sudden before the oil-water viscosity ratio is 345, and the oil-water viscosity ratio shows a gentle decline and an upward trend between 345-1600. When the oil-water viscosity ratio is greater than 1500, the oil displacement efficiency and residual oil saturation can be seen to be very small, it looks like a straight line on the whole, and the water injection development effect is not obvious, as shown in Figure 5.
It can be seen from Figure 5 that the crude oil viscosity of this research block is 76.3 mPa.s, which is more suitable for water injection development from the perspective of viscosity. The formation water viscosity is 0.5 mPa.s, its oil-water viscosity ratio is 152.6, its oil displacement efficiency is better, the residual oil saturation is lower than normal, and the study area is more suitable for water injection development.
The water cut of the Niuxintuo oil layer increases sharply with the increase of the oil reservoir’s recovery. At present, the Niuxintuo oil layer has a recovery rate of 15.01%, and the water cut has reached 72.4%. Due to the low degree of oil production, the later period of the curve is not obvious, and it can still be seen from the curve shape that the development effect of this block is not good. In the initial stage of water injection development, as the water cut increases, the rate of water cut rises. In the later stage, the oil wells are controlled by water to stabilize the oil. Through water transfer and layer replenishment measures, the current curve gradually moves closer to the theoretical curve, indicating the current overall development deviation of the reservoir. But there is a trend of gradually getting better in the later period, and the rising rate of water content shows a downward trend.
Comparing the relationship curve between the actual comprehensive water cut and the recovery degree of the reservoir with the standard curve, the result is shown in Figure 6. The comprehensive water cut of the initial block rises faster. Through timely and dynamic water allocation and comprehensive management, the rate of water cut rise is controlled., It can be seen from the curve that the current comprehensive water cut of the reservoir is between 60% and 80%, the actual curve is between 25% and 35%, and the estimated waterflooding recovery rate is about 30%.
Through the statistics of each small layer from the early stage of development to the present, the oil produced by each small layer in the entire water injection development process, the changes in the degree of production, and the changes in the original oil saturation and the current remaining oil saturation are shown in Figure 7, and the statistics of water injection are shown in Figure 7. The oil production of each small layer has been developed to date. The water injection effect of each small layer is evaluated by the changes in the reserves, production level, water production, and water content of each small layer from the early stage of development to the present, and it can be seen that the NIII1 included in the NIII layer, NIII2 has a larger oil content, a large amount of production, and a greater degree of recovery than NI and NII, indicating that the water injection development effect of NII2, NIII1, and NIII2 layers is better than that of NI2 and NI3 layers..
As shown in Figure 8, the development status of each microphase at this stage: through statistics of each microphase from the initial stage of development to the present, the oil produced by each microphase in the entire water injection development process, the change in the degree of production, and the original oil saturation and the current remaining. The change in oil saturation, the comparison of the oil production of each microphase from the development of water injection to the present, through the changes in the reserves, recovery, water production, and water content of each microphase from the initial stage of development to the present, the water injection effect of each microphase is carried out. Evaluation, it can be seen that the cumulative oil production and production degree of the channel microfacies are the largest, and it can be judged that the microfacies has the best water injection effect, followed by the side of the channel, the estuary dam, and the sheet sand.
Water injection adjustment measures for optimal well groups in the whole area: for complex fault block oil layers, adjustments are not only made to well groups in finely divided fault blocks, on the other hand, for the whole area, the well groups should also be selected from a macroscopic perspective to be used. The effect of some well groups affects the development effect of other well groups. Adjust the stratification of the above water wells, adjust the water injection volume of each layer, and apply the blocking method to the selected 19 well groups (Figure 9), and the cumulative production before and after the water transfer is obtained. The oil and production levels change, and the 19 well groups are optimized to adjust the water injection volume, adjust the stratification situation and the blocking measures according to their characteristics. After the adjustment, the cumulative oil production increase is obvious, and the cumulative oil production is 193,400 tons increased by 2.6%.
4.4. Comprehensive Plan Index Evaluation Forecast
As shown in Figure 10, by comparing the remaining oil volume of each small layer predicted 20 years after no measures and the comprehensive plan, it is concluded that after the comprehensive plan measures, the oil increase of the NI2-NII1 small layer is obvious, and the cumulative oil increase of each layer is .
Due to the uniform method used in the remeshing process, the calculation of the three-dimensional mesh vertices is independent and data-free, and each CUDA thread can process the vertices of a region. The main process can be divided into six steps: area division, data reading, data encapsulation, GRID construction, area remeshing, and output. Regional vertex division is the basis of the parallel process. The large-scale grid 3D vertex data is divided into discrete unit regions according to the resolution requirements, and there is no data overlap between adjacent units; the scheduler allocates multiple divided task units on the CPU side. Threads perform coarse-grained parallel processing, and the regrid calculation of each task area is mapped to the corresponding thread block. The regrid calculation process of each vertex is assigned to the threads on the GPU in an atomic task for fine-grained parallelism. The experimental platform: Microsoft Visual Studio 2010, GPU side graphics card is nVIDIAGeForceGT720, CUDAC4.1, CPU is CoreI5-6500, RAM 8 G. Experimental data: local river model A, B, C in a certain area of Nanjing, the number of vertices is 1405, 1939, 2576, and the running time of GPU-CPU coordinated operation and CPU is compared. It can be seen from Table 1 that as the number of model triangle vertices increases, GPU parallel resources are effectively used, and the acceleration effect is more obvious.
As shown in Table 2, the in situ stress data of the fault block mainly refers to the microresistivity scanning imaging and interactive multilevel subarray acoustic logging data in the southern part of the new station. The rock tensile strength refers to the test data of the Hailar Oilfield, which is 5 ~ 10 MPa between. After calculating , the artificial fracture first opens along the direction of the natural fracture. On this basis, the calculation of the limit angle of other wells with data is carried out, and the angle is above 48°.
The oil production rate is predicted by the reservoir numerical simulation data, and the predicted oil production rate of the D165 block is counted. It can be seen from the prediction result in Figure 11 that the annual oil production rate of this block is relatively low and has been declining.
Figure 12 shows the probability distribution histogram between the reservoir porosity and permeability model simulation parameter data of the Ling 2 block of the fault oil field in the study area, the coarsening data at the well point, and the original well data. Compare these two probabilities. From the distribution histogram, it can be seen that their distribution laws are similar, that is, the probability distribution trend of the simulated data volume and the original data is the same, which indicates that the simulation accuracy of the built reservoir parameter model is better, and the simulated value of the model is faithful to the original attributes. Value data are reflecting the changing law of reservoir attribute parameters.
5. Conclusion
This article mainly focuses on the study of fine geological modeling of complex fault block reservoirs based on deep learning. Starting from the 3D geological modeling, this paper conducts an in-depth study on the modeling of complex fault block reservoir geological bodies. Through the isosurface extraction method of attribute volume data constructed in this paper, the existence of multiple bifurcation problems between contour lines can be used. Achieve better modeling results. In addition, this paper also analyzes the relationship between water cut and recovery degree of complex fault block reservoir geology. We need to control water cut rising speed reasonably and improve recovery rate through timely and dynamic water allocation and comprehensive management. At the same time, it is obtained by comparing and analyzing the relationship curve between the actual comprehensive water content and the oil recovery factor and the standard curve. The current comprehensive water cut of the reservoir is between 60% and 80%, the actual curve is between 25% and 35%, and the estimated waterflooding recovery factor is about 30%. The innovation of this article lies in the application of deep learning to geological modeling, and the use of convolutional neural networks for multilevel analysis to meet the completeness standards and requirements of modeling. There are still shortcomings in this paper. More factors are not considered in the selection of complex fault block reservoirs. The representativeness is poor, and further improvement is needed. It is hoped that the research in this article can provide some theoretical support and help for the geological modeling research of complex fault blocks.
Data Availability
The data that support the findings of this study are available from the corresponding author upon reasonable request.
Conflicts of Interest
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Acknowledgments
This work was supported by The National Natural Science Foundation of China (41602154) and Projects of Talents Recruitment of GDUPT (2018rc01).