Unet-based Prediction Method for the Top Surface of River Channel Sand Bodies

Abstract

The present invention relates to the technical field of oilfield development, and specifically relates to a method for predicting the top surface of channel sand bodies based on Unet. The method includes the following steps: reading post-stack seismic data and the expert interpretation results of the top surface of channel sand bodies; generating sample feature values and labels; designing Unet and completing model training; applying the model to the actual work area to predict the distribution of the top surface of channel sand bodies. By setting the expert interpretation results as labels, the method of the present invention deeply analyzes the data-driven relationship between seismic data and channel labels, so as to be unaffected by the strength of seismic amplitude energy. Whether it is a channel with a strong amplitude form or a weak amplitude form, it can be accurately identified.

Description

Technical Field

[0001] The present invention relates to the technical field of oilfield development, and particularly relates to a method for predicting the top surface of channel sand bodies based on Unet. Background Technique

[0002] In the oilfields developed in the Mesozoic and Cenozoic hydrocarbon-bearing basins in the eastern part of China, the petroleum geological reserves of fluvial facies reservoirs account for more than 40%, which is an important reservoir type for increasing reserves and production. Due to the influence of environmental factors such as the development of faults, the change of structural undulations, and the frequent channel avulsion, the integrity of the channels is damaged. These situations increase the difficulty of reservoir prediction work for channel sand bodies.

[0003] From the literature research, it can be seen that the reservoir prediction technologies for channel sand bodies are mainly divided into single-attribute technologies, multi-attribute fusion technologies, spectral analysis and imaging technologies, inversion technologies, and neural network technologies, etc.

[0004] Using a single attribute to predict the distribution of channel sand bodies mainly focuses on attributes such as amplitude, coherence, and texture. When there are obvious differences in velocity between mudstone and channel sand, the amplitude attribute volume and coherence attribute volume can be extracted, and then combined with means such as slicing and 3D visualization for comprehensive analysis of channel sand. The texture attribute can also be used to depict underwater distributary channels. From the perspective of image recognition, the degree of difference between pixels in the image is measured, and the channel sand morphology is highlighted by enhancing the amplitude intensity of channel sand and weakening the amplitude intensity of mudstone. With the in-depth research, more and more researchers have found that there is strong multi-solution in predicting channels only from a single attribute. With the development of seismic attribute technologies, attribute fusion technology has gradually become a new attribute analysis method, which can effectively improve the multi-solution problem of single attributes.

[0005] For relatively thin channel sand reservoirs, due to the limited resolution of seismic data, thin-layer channel sand cannot be characterized from the amplitude attribute. Spectral analysis and imaging technologies detect thin-layer channel sand based on the tuning principle of thin-layer reflection, study attributes such as amplitude, phase, and coherence at different frequencies, and achieve good results in the study of thin layers.

[0006] Single-attribute technology, attribute fusion technology, and spectral analysis technology all start from seismic data to predict the distribution of channel sand bodies, and it is difficult to meet the increasingly high requirements for the prediction accuracy of channel sand reservoirs. Therefore, inversion technology has emerged. This type of technology combines seismic data and logging data, and the resulting prediction accuracy is higher. Inversion technology is divided into deterministic inversion and stochastic inversion. The deterministic inversion technology depends heavily on seismic data and has a slightly higher resolution than seismic data. It obtains more accurate inter-well prediction results by continuously correcting the inversion results using logging data. The stochastic inversion technology has a higher trust in logging data, and its vertical resolution is much greater than that of the deterministic inversion, but its inter-well reliability is lower than that of the deterministic inversion. Geostatistics is an important principle of stochastic inversion. In the field of geophysics, sequential Gaussian simulation and sequential indicator simulation are common methods of two-point geostatistics. Later, SNESIM algorithm, SIMPAT algorithm, FILTERSIM algorithm, and distance algorithm, etc. are all multi-point geostatistics stochastic simulation technologies. Compared with two-point geostatistics, multi-point geostatistics has a stronger ability to describe the spatial variation characteristics of variables.

[0007] Chinese invention patent CN109541685B discloses a method for identifying channel sand bodies, which includes: performing seismic acquisition on a target work area containing channel sand bodies to obtain seismic data, and logging in the target work area to obtain logging data. Interpreting seismic horizons using the seismic data, and interpreting sand body thickness, sedimentary facies, and each development horizon using the logging data. Performing well-seismic calibration using the seismic horizons, sand body thickness, sedimentary facies, and each development horizon to obtain well-seismic calibration results. According to the well-seismic calibration results, establish the spatial horizons of sandstone groups, sub-layers, and sedimentary units in the time domain. Obtain the types of sedimentary facies, and perform vertical combination on the spatial horizons of the above three according to the types of sedimentary facies to obtain a framework model. Use the framework model to complete the identification of the distribution characteristics of channel sand bodies in the target work area.

[0008] Chinese invention patent CN105372703B discloses a method for fine identification of channel sand bodies. First, determine the seismic response characteristics corresponding to different sand body deposition patterns to identify channel characteristics; then determine the phase relationship of sand layers on the time section through desanding experiments, and identify the envelope surface of channel sand bodies for the seismic response characteristics of channels; cut the stratigraphic slices at equal time intervals from top to bottom to obtain the superimposed relationship of channels in different periods; carry out clustering analysis of five basic attributes and other unconventional attributes to determine the boundary of channel sand bodies; carry out fine stratigraphic correlation to determine the internal spatial superimposed relationship of sand bodies, accurately implement the oil-bearing area and reserves for single sand bodies, and determine the deployment range of development well patterns. Through the above process, the invention solves the technical problems such as the complex planar distribution of reservoirs in multi-source intersection areas, strong heterogeneity of sand bodies, and the resolution of seismic data not meeting the requirements for predicting thin reservoirs, greatly improves the drilling success rate, and has broad market application prospects.

[0009] The above technologies have made great contributions to the prediction of channel sand reservoirs and had a positive impact on increasing reserves and production in major oilfields. However, with the deepening of the research on channel sands, researchers have found that the velocity of channel sands is not necessarily higher than that of mudstones in some work areas, and is even comparable to or lower than that of mudstones. Therefore, the multi-solution of using the amplitude strength to judge the channel boundary is relatively strong. It can be judged from this that for all prediction results obtained from seismic data and logging data, regardless of which technology is used, a certain definite value cannot be used to divide sand and mud, because in a certain definite work area, the boundary values of sand and mud are different at different positions. Currently, all prediction results obtained by technologies use a definite boundary value to divide sand and mud, which is an unreasonable situation.

[0010] In view of this problem, a prediction method for the top surface of channel sand bodies is needed that is not affected by the strength of seismic amplitude energy and can accurately identify channels in both strong and weak amplitude forms. Summary of the Invention

[0011] The main object of the present invention is to provide a prediction method for the top surface of channel sand bodies based on Unet. The method of the present invention sets the expert interpretation results as labels and deeply analyzes the data-driven relationship between seismic data and channel labels, so as to be not affected by the strength of seismic amplitude energy and can accurately identify channels in both strong and weak amplitude forms.

[0012] To achieve the above object, the present invention adopts the following technical solutions:

[0013] The present invention provides a prediction method for the top surface of channel sand bodies based on Unet, including the following steps:

[0014] Step 1. Read the post-stack seismic data and the expert interpretation results of the top surface of channel sand bodies; in this step, the expert interpretation results refer to: based on the geological understanding and actual drilling conditions, on the basis of fine well-seismic calibration, it is clear which seismic reflection axis represents the top surface of the channel sand, and the expert interprets this seismic reflection axis completely and correctly. The layer data interpreted by the expert is the expert interpretation result.

[0015] Step 2. Generate sample feature values and labels;

[0016] Step 3. Design Unet and complete model training;

[0017] Step 4. Apply the model to the actual work area to predict the distribution of the top surface of channel sand bodies.

[0018] Furthermore, in Step 2, the post-stack seismic data and the expert interpretation results of the top surface of channel sand bodies are calibrated to have the same spatial range; the spatial range includes the line number and the time value.

[0019] Furthermore, traverse all points within the spatial range and access the seismic data within this range as sample feature values.

[0020] Furthermore, set the points with expert interpretation results to 1 and the remaining points to 0.

[0021] Further, in step 3, design Unet, including multi-layer downsampling CNN convolutions, multiple feature concatenations, and multi-layer transposed convolutions.

[0022] Furthermore, during downsampling CNN convolutions, each convolutional kernel receives all the outputs of the previous layer as input, and the outputs after CNN convolutions are all used as the input for the next layer; during upsampling, transposed convolutions are used. By padding a large number of 0s around the small-sized input data, the originally reduced-sized input data is restored to the size of the previous layer through the convolution process.

[0023] Furthermore, concatenate the original features obtained from downsampling and the segmented features restored by upsampling in the channel dimension to obtain new combined features and send them as input to the double convolution and the subsequent transposed convolution.

[0024] Further, in step 3, extract training batches from the eigenvalue data volume and the label data volume around the river channel top surface to participate in model training; through continuous training, make the loss function as small as possible.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] The method of the present invention applies the Unet network to identify the top surface of river channel sand bodies from seismic data. By setting the expert interpretation results as labels, this method deeply analyzes the data-driven relationship between seismic data and river channel labels, thus being unaffected by the strength of seismic amplitude energy. Whether it is a river channel with a strong amplitude form or a weak amplitude form, it can be accurately identified, thereby overcoming the deficiencies of existing prediction methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings forming a part of this invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0028] Figure 1 It is a flowchart of the method for predicting the top surface of river channel sand bodies based on Unet according to a specific embodiment of the present invention;

[0029] Figure 2 It is a structural diagram of the method for predicting the top surface of river channel sand bodies based on Unet according to a specific embodiment of the present invention; Figure 2On the left in is the downsampling process, including 4 layers of downsampling CNN convolutions, and on the right is the upsampling process, including 4 times of feature splicing and 4 layers of transposed convolutions; Figure 2 In, arrow 1 represents the CNN convolution, arrow 2 represents the max pooling, arrow 3 represents the upsampling transposed convolution, and arrow 4 represents the feature moment splicing;

[0030] Figure 3 This is the actual data prediction result of the method for predicting the top surface of channel sand bodies based on Unet according to a specific embodiment of the present invention. Detailed implementation manners

[0031] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0032] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, and / or combinations thereof.

[0033] In order to enable those skilled in the art to more clearly understand the technical solution of the present invention, the technical solution of the present invention will be described in detail below in conjunction with specific embodiments.

[0034] Embodiment 1

[0035] As Figure 1 shown, the method for predicting the top surface of channel sand bodies based on Unet includes the following steps:

[0036] Step 1, read the post-stack seismic data and the expert interpretation results of the top surface of the channel sand bodies.

[0037] Step 2, calibrate the post-stack seismic data and the expert interpretation results of the top surface of the channel sand bodies so that they have the same spatial range, including the line number and the time value. Traverse all points within the spatial range, and store the seismic data within this range as sample feature values. Traverse all points within the spatial range, set the points with expert interpretation results as 1, and the remaining points as 0.

[0038] In step 3, design Unet, and the Unet is as Figure 2As shown, it includes 4 layers of downsampling CNN convolutions, 4 times of feature concatenation, and 4 layers of transposed convolutions. During downsampling CNN convolution, each convolution kernel receives all the outputs of the previous layer as input, and the outputs after CNN convolution are all used as the input of the next layer. When upsampling, transposed convolution is used. By padding a large number of 0s around the small-sized input data, the originally reduced-sized input data is restored to the size of the previous layer through the convolution process. The original features obtained by downsampling and the segmented features restored by upsampling are concatenated in the channel dimension to obtain a new combined feature, which is sent as input to the double convolution and the subsequent transposed convolution. Training batches are extracted around the top surface of the river channel from the eigenvalue data volume and the label data volume to participate in model training. By continuously adjusting parameters such as the optimizer and the learning rate, the model is trained to make the loss function as small as possible.

[0039] In step 4, the model is applied to predict the top surface distribution of the channel sand body in the actual work area, and the prediction results of the actual data are as Figure 3 shown, and it can be seen from Figure 3 that the channel morphology in the figure is depicted relatively clearly.

[0040] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A method for predicting the top surface of river channel sand bodies based on Unet, characterized in that, It includes the following steps: Step 1. Read the post-stack seismic data and the expert interpretation results of the top surface of the channel sand body; the expert interpretation results refer to: combining geological understanding and actual drilling conditions, on the basis of fine well-seismic calibration, clarifying which seismic reflection axis represents the top surface of the channel sand, and having the expert interpret this seismic reflection axis completely and correctly, and the horizon data interpreted by the expert; Step 2. Generate sample feature values and labels; Step 3. Design Unet and complete model training; Step 4. Apply the model to the actual work area to predict the distribution of the top surface of the channel sand body; In Step 2, calibrate the post-stack seismic data and the expert interpretation results of the top surface of the channel sand body so that they have the same spatial range; the spatial range includes line numbers and time values; traverse all points within the spatial range, and store the seismic data within this range as sample feature values; set the points with expert interpretation results to 1, and the remaining points to 0; In Step 3, design Unet, including multiple layers of downsampling CNN convolutions, multiple feature splicings, and multiple layers of transposed convolutions; when performing downsampling CNN convolutions, each convolutional kernel receives all the outputs of the previous layer as input, and the outputs after CNN convolutions are all used as the input of the next layer; when performing upsampling, use transposed convolutions to restore the originally reduced-size input data to the size of the previous layer through the convolution process by padding a large number of 0s around the small-size input data; further, splice the original features obtained by downsampling and the segmented features restored by upsampling in the channel dimension to obtain new combined features and send them as input to the double convolution and the subsequent transposed convolution; extract training batches around the channel top from the feature value data volume and the label data volume to participate in model training; through continuous training, make the loss function as small as possible.

Images