阅读说明：本技术 一种基于深度混合神经网络的储层预测方法 (Reservoir prediction method based on deep hybrid neural network ) 是由 王俊 曹俊兴 蒋旭东 于 2021-08-13 设计创作，主要内容包括：本发明提供了一种基于深度混合神经网络的地震储层预测方法,所述方法包括：首先,根据研究层段的解释层位计算目标层段的多种地震属性,提取已钻井的储层分类信息；然后,将提取的地震属性通过聚类分析技术进行优选,选取对含油气储层敏感的地震属性；基于此,以测井数据和地震数据为驱动,利用井点的储层分类数据为标签,优选的地震属性为输入,训练深度混合神经网络深度学习模型；最后,利用训练好的反演模型进行从已知到未知的地震储层直接反演预测。本发明综合了井点数据与地震数据,可较好的挖掘原始数据中蕴含的信息,实现无井区的地震储层预测,所预测的油气储层边界更加清晰,预测结果与实际情况基本吻合。(The invention provides a seismic reservoir prediction method based on a deep hybrid neural network, which comprises the following steps: firstly, calculating various seismic attributes of a target interval according to an interpretation horizon of a study interval, and extracting classification information of a drilled reservoir; then, the extracted seismic attributes are optimized through a clustering analysis technology, and seismic attributes sensitive to oil-gas reservoirs are selected; based on the method, the logging data and the seismic data are used as driving, reservoir classification data of well points are used as labels, and the optimal seismic attributes are used as input, so that a deep hybrid neural network deep learning model is trained; and finally, carrying out direct inversion prediction from a known to an unknown seismic reservoir by using the trained inversion model. The invention integrates well point data and seismic data, can better mine information contained in original data, realizes seismic reservoir prediction without a well area, and has clearer predicted oil and gas reservoir boundary and basically consistent prediction result with actual situation.)

1. A reservoir prediction method based on a deep hybrid neural network is characterized by comprising the following steps:

step 1, accurately calibrating a horizon according to exploratory well information and seismic information to obtain seismic data and well data;

step 2, calculating a plurality of seismic attributes according to the interpretation horizon of the study interval, and extracting the classification information of the drilled reservoir;

step 3, the seismic attributes extracted in the step 2 are optimized through a clustering analysis technology, seismic attributes sensitive to an oil-gas reservoir are selected, and influences of irrelevant seismic attributes are reduced;

step 4, based on the step 2 and the step 3, training a deep hybrid neural network deep learning seismic reservoir inversion prediction model by using the logging data and the seismic data as driving, using reservoir classification data of well points as training labels and using preferred seismic attribute data as training input;

and 5, taking seismic attribute data of the non-well area as input, and performing inversion prediction from a known seismic reservoir to an unknown seismic reservoir by using the deep learning model trained in the step 4.

2. The reservoir prediction method based on the deep hybrid neural network as claimed in claim 1, wherein: on the basis of accurate calibration of a target layer, acquiring seismic data and well data, utilizing a seismic interpretation layer position for constraint, and extracting seismic multiple attributes of the target layer section, including seismic attributes such as amplitude, frequency, phase and the like; on the basis, establishing a corresponding relation between the seismic attributes and the reservoir classification; and extracting various seismic attribute data of the well points according to the well point coordinates, and establishing a well point position reservoir classification and various seismic attribute database.

3. The reservoir prediction method based on the deep hybrid neural network as claimed in claim 1, wherein: the deep hybrid neural network deep learning model is formed by stacking a convolutional neural network and a bidirectional cyclic neural network, and the intrinsic characteristics between seismic attribute data and reservoir type information are mined to the greatest extent by fully utilizing the advantages of the two networks; the convolutional neural network is used as a feature extractor for extracting seismic data information, the feature information extracted by the convolutional neural network is used as the input of the bidirectional cyclic neural network to consider the time sequence relation of stratum sedimentation, the stratum sedimentation is converted into internal hidden layer information, and then the hidden information is converted into reservoir type data through a full connecting layer.

4. The reservoir prediction method based on the deep hybrid neural network as claimed in claim 1, wherein: the deep hybrid neural network prediction method mainly comprises three parts of data standardization, model establishment and prediction, analysis and evaluation. The model building and predicting part is a core module of the model and mainly comprises network building and data prediction, and the reservoir prediction is directly carried out by utilizing seismic data through training a complex mapping relation between network approximation input and target output by a training data set.

5. The reservoir prediction method based on the deep hybrid neural network as claimed in claim 1, wherein: the method integrates well point data and seismic data, can better mine effective information contained in the original seismic data, avoids influences caused by transverse and longitudinal heterogeneity of a reservoir to the maximum extent, realizes direct inversion prediction from a known seismic reservoir to an unknown seismic reservoir, ensures that the predicted seismic oil-gas reservoir boundary is clearer, and ensures that a prediction result is basically consistent with an actual situation.

Technical Field

The invention relates to the field of geophysical exploration of petroleum, in particular to a seismic reservoir inversion prediction method based on a depth hybrid neural network.

Background

Seismic exploration is one of the most common oil and gas exploration methods at present, and seismic data are used as media of underground geological conditions and always occupy a crucial position. With the continuous deepening of oil and gas exploration and development, an exploration object is gradually changed from a conventional oil and gas reservoir to a fractured oil and gas reservoir and a lithologic oil and gas reservoir, and the precision and the efficiency of a conventional reservoir prediction method are difficult to meet the prediction requirements of the oil and gas reservoirs, particularly under the deep condition. Therefore, the method for predicting the seismic reservoir by pertinently developing along with a new thought has great significance for oil and gas exploration and development.

The development of the artificial intelligence deep learning technology provides a new idea and method for accurate inversion prediction of the seismic reservoir. Deep learning is an algorithm for automatically analyzing and obtaining rules from data by using a computer and predicting unknown data by using the obtained rules, is suitable for nonlinear complex tasks, and achieves breakthrough application results in many scientific fields. The essence of deep learning is to construct a nested mapping matrix called deep network, and the functions are equivalent to the variable mapping relations in the numerical analysis model. In the complex problem that the variable relation can not be determined, deep learning is the best means for constructing a mapping relation model at the present stage. The problem we are now faced with is that of: the relation between the reservoir type and the seismic data is quite complex, and the reservoir type and the seismic data are influenced by various factors such as the elastic modulus of a solid framework, a pore structure, the size of pores, the connectivity of pores, the composition of pore fluids and the like, so that the reservoir type and the seismic data show strong nonlinear characteristics. In this case, using deep learning to build a deep network mapping model may be the best choice.

Therefore, the method is inspired by an integrated learning idea, two special network convolution neural networks and a bidirectional circulation neural network in deep learning are integrated based on the reservoir geological deposition rule and the characteristics of the reservoir geological deposition rule on seismic response, and the seismic reservoir prediction method based on the deep hybrid neural network deep learning model is invented, so that the intrinsic characteristics between input and output data are mined to the maximum degree by fully utilizing the advantages of the two networks, and the direct inversion prediction from the known to the unknown seismic reservoir is realized.

Disclosure of Invention

The invention aims to provide a reservoir prediction method based on a deep hybrid neural network, which takes logging data and seismic data as drive, trains a deep learning seismic inversion model by using reservoir classification data of well points as training labels and preferably seismic attribute data as training input, establishes a mapping complex mapping relation between the seismic data and the reservoir, and realizes direct inversion prediction from known to unknown seismic reservoirs. The predicted earthquake oil and gas reservoir boundary is clearer, and the prediction result is basically consistent with the actual situation.

In order to achieve the purpose, the invention adopts the following technical scheme:

a reservoir prediction method based on a deep hybrid neural network comprises the following steps:

step 1, accurately calibrating a horizon according to exploratory well information and seismic information to obtain seismic data and well data;

step 2, calculating a plurality of seismic attributes according to the interpretation horizon of the study interval, and extracting the classification information of the drilled reservoir;

step 3, the seismic attributes extracted in the step 2 are optimized through a clustering analysis technology, seismic attributes sensitive to an oil-gas reservoir are selected, and influences of irrelevant seismic attributes are reduced;

step 4, based on the step 2 and the step 3, training a deep hybrid neural network deep learning seismic inversion model by using the logging data and the seismic data as driving, using reservoir classification data of well points as training labels and using preferred seismic attribute data as training input;

and 5, taking seismic attribute data of the non-well area as input, and performing inversion prediction from a known seismic reservoir to an unknown seismic reservoir by using the deep learning model trained in the step 4.

In the present invention, the reservoir type data and seismic data are common knowledge in the art.

The invention discloses a reservoir prediction method based on a deep hybrid neural network, which has the core problems that logging data and seismic data are used as drive, reservoir classification data of well points are used as training labels, optimal seismic attribute data are used as training input, a deep hybrid neural network deep learning seismic inversion model is trained, and a mapping complex mapping relation between the seismic data and reservoir type information is established.

The reservoir prediction method based on the deep hybrid neural network has the following characteristics that:

(1) on the basis of accurate calibration of a target layer, the layer position of seismic interpretation is utilized for constraint, and seismic attributes of the target layer section, including seismic attributes such as amplitude, frequency and phase, are extracted; on the basis, establishing a corresponding relation between the seismic attributes and the reservoir classification; and extracting various seismic attribute data of the well points according to the well point coordinates, and establishing a well point position reservoir classification and various seismic attribute database.

(2) The deep hybrid neural network deep learning model is formed by stacking a convolutional neural network and a bidirectional cyclic neural network, and the intrinsic characteristics between input and output data are mined to the greatest extent by fully utilizing the advantages of the two networks; the convolutional neural network is used as a feature extractor for extracting seismic data information, the feature information extracted by the convolutional neural network is used as the input of the bidirectional cyclic neural network to consider the time sequence relation of stratum sedimentation, the stratum sedimentation is converted into internal hidden layer information, and then the hidden information is converted into reservoir type data through a full connecting layer.

(3) The deep hybrid neural network prediction method mainly comprises three parts of data standardization, model establishment and prediction, analysis and evaluation. The model building and predicting part is a core module of the model, mainly comprises network building and data prediction, the complex mapping relation between seismic attribute data and reservoir type information is trained and approximated by a training data set through a network, and the seismic attribute data of a well-free area is used as input to realize the prediction of the known to unknown seismic reservoir.

The invention has the beneficial effects that: the method provided by the invention is driven by logging data and seismic data, integrates well point data and seismic attribute data, can well mine effective information contained in the original seismic data, avoids influences caused by transverse and longitudinal heterogeneity of a reservoir to the greatest extent, realizes direct inversion prediction from a known reservoir to an unknown seismic reservoir, and has the advantages that the predicted seismic oil-gas reservoir boundary is clearer, and the prediction result is basically consistent with the actual situation.

Drawings

FIG. 1 is a flow chart of an embodiment of a method for reservoir prediction based on a deep hybrid neural network according to the present invention;

FIG. 2 is a diagram of a deep hybrid neural network in accordance with an embodiment of the present invention;

FIG. 3 is a sectional view of a well-connected seismic section of a study area in accordance with an embodiment of the present invention;

FIG. 4 is a cross-sectional view of a well log of predicted results of a conventional method in accordance with an embodiment of the present invention;

FIG. 5 is a cross-sectional view of a prediction result well based on a deep hybrid neural network method in an embodiment of the present invention.

Detailed Description

In order to make those skilled in the art better understand the technical solutions in the present specification, the technical solutions in one or more embodiments of the present specification will be clearly and completely described below with reference to the drawings in one or more embodiments of the present specification, and it is obvious that the described embodiments are only a part of the embodiments of the specification, and not all embodiments.

As shown in fig. 1, fig. 1 is a flow chart of a reservoir prediction method based on a deep hybrid neural network according to the present invention.

Step 1, accurately calibrating a horizon according to exploratory well information and seismic information to obtain seismic data and well data;

step 2, calculating a plurality of seismic attributes according to the interpretation horizon of the study interval, and extracting the classification information of the drilled reservoir;

step 3, the seismic attributes extracted in the step 2 are optimized through a clustering analysis technology, seismic attributes sensitive to an oil-gas reservoir are selected, and influences of irrelevant seismic attributes are reduced;

step 4, based on the step 2 and the step 3, training a deep hybrid neural network deep learning seismic reservoir inversion model by using the logging data and the seismic data as driving, using reservoir classification data of well points as training labels and using preferred seismic attribute data as training input;

and 5, taking seismic attribute data of the non-well area as input, and performing inversion prediction from a known seismic reservoir to an unknown seismic reservoir by using the deep learning model trained in the step 4.

The seismic data volume in step 2 comprises original seismic data and derivative data volumes, including original seismic data, amplitude data, fluid detection data, frequency data, phase data, construction data and the like.

And 2, obtaining reservoir and non-reservoir according to lithological data and oily data of each layer obtained by exploratory well.

4, the deep hybrid neural network deep learning model is formed by stacking a convolutional neural network and a bidirectional cyclic neural network, and the intrinsic characteristics between input and output data are mined to the greatest extent by fully utilizing the advantages of the two networks; the convolutional neural network is used as a feature extractor for extracting seismic data information, the feature information extracted by the convolutional neural network is used as the input of the bidirectional cyclic neural network to consider the time sequence relation of stratum sedimentation, the stratum sedimentation is converted into internal hidden layer information, and then the hidden information is converted into reservoir type data through a full connecting layer.

And 4, the deep hybrid neural network prediction method mainly comprises three parts of data standardization, model establishment and prediction, analysis and evaluation. Data normalization: in order to reduce errors caused by large magnitude difference between input and output data, a mean-variance normalization method is adopted to carry out standardization processing on the input data so as to ensure that all parameters are in a reasonable distribution range, and a network can learn the internal relation between the data more easily; model building and prediction: the part is a core module of the model and mainly comprises network construction and data prediction, and the reservoir prediction without a well area is directly carried out by utilizing seismic data through training a complex mapping relation between network approximation input and target output by a training data set; analysis and evaluation: and evaluating the accuracy of the model prediction result according to different evaluation indexes.

FIG. 2 is a diagram of a deep hybrid neural network structure of the method of the present invention, based on reservoir geological deposition rules and their characteristics in seismic response, and based on the complex relationship between seismic data and reservoir type information, inspired by the idea of ensemble learning, in order to fully utilize the characteristics of two special networks, namely convolutional neural network and bidirectional cyclic network, and to maximally mine the intrinsic characteristics between seismic attribute data and reservoir type information, the present invention proposes a model combining the two networks to process data with complex characteristics, which is called a deep hybrid neural network model. The main thought of the method has two aspects. Firstly, as a special network model, the convolutional neural network can better learn data characteristics, and the calculation amount is greatly reduced through weight sharing. The convolutional neural network can not cause a large amount of characteristic information loss due to the extraction of deep-level characteristics of the data, and the input data of the one-dimensional convolutional neural network model is considered to be a numerical vector, and the seismic data can be generated to meet the input requirement of the one-dimensional convolutional neural network model after being properly processed. Therefore, the one-dimensional convolution neural network is selected as a feature extractor for extracting seismic data information. Secondly, from the perspective of reservoir deposition continuity, seismic data can be regarded as time series data with connection in the longitudinal direction, and time series characteristics are key information for reservoir prediction. The time sequence characteristics of time sequence data can be efficiently extracted by the bidirectional circulation network due to strong memory capacity, so that the characteristic information extracted by the convolutional neural network is used as the input of the bidirectional circulation network, the trend information of the seismic data along the depth direction is fully learned by the bidirectional circulation network, the time sequence relation of stratum deposition is considered, and the uncertainty of a prediction result is reduced to a certain extent.

FIG. 3 is a sectional view of a seismic profile of a well-connected area of interest from a deep carbonate reservoir in West, Chunxi, which is a seismic profile through three wells, according to an embodiment of the present invention. The T1 layer from top to bottom is the top of the weathering crust at the top of the Lei Kong slope group, the bottom of the saddle pond group, the T2 layer is the top of the four sections of the Lei Kong slope group, the T3 layer is the bottom of the four sections of the Lei Kong slope group, the T4 layer is the bottom of the three sections of the Lei Kong slope group, and the T5 layer is the bottom of the one section of the Lei Kong slope group. In terms of the current drilling result, the main reservoir gas-containing target horizon is a gas-containing horizon drilled by three wells in a Leikou slope group top weathering crust reservoir (T1), wherein the gas-containing property of the wells 1 and 2 in the gas-containing horizon is better; and drilling meeting gas layers of the four sections of tops (T2) of the mine opening slope group, namely a drilled well 2 and a drilled well 3, wherein the well 1 is a water layer; the top of the three sections (T4) of the Leifang slope set is only provided with a well 2 drilling gas layer, and the rest is a water layer.

Fig. 4 is a sectional view of a well-connecting predicted result of a classic AVO method, and it can be seen from the predicted result of fig. 3 that the overall accuracy of the AVO predicted result is low and the goodness of fit with the drilling result is low. Well 1 is a water well, although the prediction result shows that the well does not contain a gas well, the gas containing characteristic of the top of the mine opening slope group is not shown; the well 2 is a high-yield period well, and the prediction result shows that the well does not contain a gas well; well 3 is a low producing gas well and the gas bearing prediction result is a strong gas bearing zone, which does not match the drilling result. The prediction result based on the AVO inversion analysis method has larger difference with the actual drilling, which indicates that the method is not suitable for deep conditions, so that new means and methods are needed for reservoir prediction of the carbonate rock of the deep Leifou slope group.

And FIG. 5 is a well-connecting section diagram of a prediction result based on a depth hybrid neural network method, and as can be seen from the prediction result in FIG. 5, the reservoir prediction accuracy is effectively improved compared with the original data, and the overall gas content prediction result has high goodness of fit with the drilling result and the geological rule. The weathering crust (the bottom T1 of the saddle pond group) and the top (T2) of the four sections of the leigou slope group are all embodied, and the top (T3) of the three sections of the leigou slope group is also embodied. The actual drilling of the posterior well 1 shows that except the weathering crust at the top, the water layers and the non-reservoir layers are all water layers and non-reservoir layers, the predicted result has high goodness of fit relative to the original result, the actual drilling of the well 2 shows that the predicted result is a high-yield gas layer, the predicted result is at the top of the Leikou slope group, the top of three sections of the Leikou slope group, the middle and lower subsections represent the gas content, and the gas content of the well 3 at the top of the Leikou slope group is also reflected in the predicted result. The result shows that the goodness of fit of the whole gas-bearing result and the actual drilling result is high, and the reservoir of the gas-bearing reservoir of the weathering crust on the top of the Raujin slope group is all characterized.

The comparative analysis of the prediction results of the classical AVO method and the deep hybrid neural network prediction method in the figures 4 and 5 shows that the method integrates well point data and seismic data, can better mine effective information contained in the original seismic data, avoids influences caused by transverse and longitudinal heterogeneity of a reservoir to the maximum extent, is successfully applied to a carbonate rock area with a complex gas-water relationship in a Chunxi deep Lekou slope group, the predicted boundary of the seismic oil-gas reservoir is clearer, and the prediction result is basically consistent with the actual situation.

The foregoing shows and describes the general principles of the present patent, with the primary features and characteristics of the present patent. It will be understood by those skilled in the art that the invention is not limited to the embodiments described above, which are described in the specification and illustrated only by the principles of the invention, but that various changes and modifications can be made therein without departing from the spirit and scope of the invention as defined by the appended claims.