A seismic reservoir prediction method based on ResNet
Through the reservoir prediction method based on ResNet, a reservoir characteristic model is established using seismic amplitude data, which solves the problem of low longitudinal wave impedance distinction accuracy and overfitting of BP neural networks in traditional methods, achieving higher reservoir prediction accuracy and better Low exploration and development risks.
Patent Information
- Application Number
- CN202110004196.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-01-04
- Publication Date
- 2024-12-10
- Estimated Expiration
- 2041-01-04
AI Technical Summary
Traditional seismic reservoir prediction methods have low accuracy when the longitudinal wave impedance cannot distinguish lithologies. BP neural networks are prone to fall into local extremely small points and the structure is too deep, resulting in overfitting, resulting in low prediction accuracy.
Using the ResNet-based reservoir prediction method, the pre-trained ResNet residual network model is used to establish the sand body thickness and porosity model using seismic amplitude data, establish the nonlinear relationship between the seismic amplitude data and the reservoir characteristics, and improve the reservoir. Prediction accuracy.
Through deep learning, the slight differences between reservoirs and non-reservoirs can be fully explored, the accuracy of reservoir prediction will be improved, the risks of exploration and development will be reduced, and more accurate reservoir prediction will be achieved.
Smart Images
Figure CN114721041B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a seismic reservoir prediction method based on ResNet, and belongs to the technical field of seismic exploration reservoir prediction. Background Art
[0002] As the requirements for seismic exploration and development become higher and higher, the traditional reservoir prediction methods alone can no longer meet the geological needs, especially when the difference between seismic reservoirs and non-reservoirs in seismic waveforms is too small, resulting in the inability to distinguish reservoirs by longitudinal wave impedance. Traditional seismic attributes and seismic inversion methods cannot make effective reservoir predictions. Among them, for the case where longitudinal wave impedance cannot directly distinguish lithology, the post-stack theory is almost invalid. The curve constructed based on the idea of curve reconstruction, which can distinguish lithology and has a high correlation with longitudinal wave impedance, is essentially an extension of longitudinal wave impedance. The random inversion method using logging data as hard data has improved the vertical resolution, but the changes in lateral distribution and formation occurrence are quite different from the actual seismic profile. Although the simultaneous inversion of three parameters based on prestack theory can directly obtain data that can distinguish lithology, this method generally approximates the isotropic Zoeppritz equation based on the convolution model theory into the Aki-Richars equation, which has certain approximations and assumptions, and the shear wave is often missing, which is basically obtained by calculation. The elastic parameters such as P-wave velocity, S-wave velocity and density obtained by inversion have deficiencies in theoretical basis and inversion accuracy.
[0003] Currently, there are methods for seismic reservoir prediction based on BP neural networks, but traditional BP neural networks have two problems: first, they are easily trapped in local minima; second, when the network structure is shallow, the learning ability is weak, but when the network structure is deep, it will produce greater redundancy and lead to overfitting. These problems lead to low prediction accuracy for seismic reservoir prediction based on BP neural networks. Summary of the invention
[0004] The purpose of the present invention is to provide a seismic reservoir prediction method based on ResNet, which can improve the accuracy of reservoir prediction.
[0005] In order to achieve the above object, the present invention provides a seismic reservoir prediction method based on ResNet, which comprises the following steps:
[0006] (1) acquiring seismic data of a target layer section in a work area, and extracting seismic amplitude data of the target layer section from the seismic data;
[0007] (2) inputting the seismic amplitude data of the target layer into a pre-trained ResNet residual network model to predict the sand body thickness and / or porosity of the target layer to complete reservoir prediction;
[0008] The pre-trained ResNet residual network model includes: an amplitude-sand body thickness model and / or an amplitude-porosity model, and the amplitude-sand body thickness model is obtained by the following steps:
[0009] Obtain logging data and well bypass seismic data of multiple known wells in the target layer section of the drilling area;
[0010] The thickness of the sand body at the target layer section of each known well is calculated by using the well logging data of each known well, and the seismic amplitude data at the target layer section of each known well is extracted from the well bypass seismic data of each known well;
[0011] Taking the seismic amplitude data at each known well target layer section as input and the sand body thickness at the corresponding known well target layer section as output, thereby obtaining a first training sample set from the seismic amplitude data and sand body thickness at multiple known well target layer sections, and using the first training sample set to train the ResNet residual network model to obtain an amplitude-sand body thickness model;
[0012] The amplitude-porosity model is obtained by the following steps:
[0013] Obtain logging data and well bypass seismic data of multiple known wells in the target layer section of the drilling area;
[0014] The porosity of the target layer section of each known well is calculated by using the well logging data of each known well, and the seismic amplitude data of the target layer section of each known well is extracted from the well bypass seismic data of each known well;
[0015] The seismic amplitude data at each known well target layer segment is taken as input, and the porosity at the corresponding known well target layer segment is taken as output, so as to obtain a second training sample set from the seismic amplitude data and porosity at multiple known well target layer segments. The ResNet residual network model is trained using the second training sample set to obtain an amplitude-porosity model.
[0016] The beneficial effects of the present invention are as follows: the present invention uses a pre-trained ResNet residual network model (including an amplitude-sand body thickness model and / or an amplitude-porosity model) to realize reservoir prediction, wherein the amplitude-sand body thickness model uses the seismic amplitude data and sand body thickness of multiple known wells drilled in the target layer section in the work area to train the ResNet residual network model, and can establish a nonlinear relationship between seismic amplitude data and sand body thickness; the amplitude-porosity model uses the seismic amplitude data and porosity of multiple known wells drilled in the target layer section in the work area to train the ResNet residual network model, and can establish a nonlinear relationship between seismic amplitude data and porosity. Since the present invention can fully exploit the slight differences between reservoirs and non-reservoirs through deep learning, compared with traditional geophysical theory, the present invention can improve the accuracy of reservoir prediction, effectively reduce the risk of exploration and development, and has good application prospects in reservoir prediction.
[0017] Furthermore, in the above method, when the pre-trained ResNet residual network model includes an amplitude-sand body thickness model, the method also includes a normalization step, normalizing the seismic amplitude data to the interval [-1,1], and normalizing the sand body thickness to the interval [0,1], using the normalized seismic amplitude data and sand body thickness to train the ResNet residual network model to obtain the amplitude-sand body thickness model, and inputting the normalized seismic amplitude data into the amplitude-sand body thickness model to predict the sand body thickness of the target layer.
[0018] The beneficial effect of this is that by normalizing the seismic amplitude data and the sand body thickness and using the normalized data for convolution processing, the speed of the convolution processing can be greatly improved.
[0019] Furthermore, in the above method, the seismic amplitude data at the target layer of the known well is obtained by the following steps: First, the seismic line number coordinates (a, b) at the wellhead of the known well are calculated by combining the wellhead coordinates (x, y) of the known well and the formulas b=b0+round[(x-x0) / m] and a=a0+round[(y-y0) / m], where (x0, y0) are the coordinates of the initial point in the work area, (a0, b0) are the seismic line number coordinates corresponding to the initial point, m is the track spacing, and round represents Round to the nearest integer; then, combine the seismic track number coordinates at the wellhead of the known well to search for the wellside track seismic data at the target layer segment of the known well from the seismic data of the work area, or first search for the wellside track seismic data of the known well from the seismic data of the work area in combination with the seismic track number coordinates at the wellhead of the known well, and then combine the position of the target layer segment in the known well to obtain the wellside track seismic data at the target layer segment of the known well; finally, extract the seismic amplitude data at the target layer segment of the known well from the wellside track seismic data at the target layer segment of the known well.
[0020] Furthermore, in the above method, the network structure of the ResNet residual network model is ResNet18, ResNet34, ResNet50, ResNet101 or ResNet152.
[0021] Furthermore, in the above method, the known well is a vertical well. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 is a flow chart of a seismic reservoir prediction method based on ResNet in Example 1 of the method of the present invention;
[0023] Figure 2 1 is a diagram of the training process of the ResNet residual network model in Embodiment 1 of the method of the present invention;
[0024] Figure 3-1 is a graph showing the change of the root mean square with the number of iterations during the ResNet18 training process in Example 1 of the method of the present invention;
[0025] Figure 3-2 : is a graph showing the change of the loss function with the number of iterations during the ResNet18 training process in Example 1 of the method of the present invention;
[0026] Figure 4 This is a graph showing the sand body thickness prediction results in Example 1 of the method of the present invention. DETAILED DESCRIPTION
[0027] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.
[0028] Method Example 1:
[0029] like Figure 1 The seismic reservoir prediction method based on this embodiment (hereinafter referred to as the method of this embodiment) includes the following steps:
[0030] (1) obtaining seismic data of a target layer section in the work area, and extracting seismic amplitude data of the target layer section from the seismic data of the target layer section;
[0031] (2) The seismic amplitude data of the target layer is normalized to the interval [-1, 1], and the normalized seismic amplitude data of the target layer is input into the pre-trained amplitude-sand body thickness model to predict the sand body thickness of the target layer and complete the reservoir prediction.
[0032] The training process of the amplitude-sand body thickness model is as follows: Figure 2 As shown, the details are as follows:
[0033] (1) Obtaining well logging data and well bypass seismic data of multiple known wells in the target layer section of the drilling area;
[0034] The basis for selecting well logging data is that the well logging data is more evenly distributed, and vertical well logging data is preferred, which is easier to match with seismic data. The horizontal well logging data is only used as verification data.
[0035] The basis for selecting the well-side channel seismic data is that the well-side channel seismic data has a high signal-to-noise ratio and the seismic isotropic axis is relatively continuous.
[0036] (2) using the well logging data (e.g., gamma logging curve) of each known well to calculate the sand body thickness at the target layer of each known well, and extracting the seismic amplitude data at the target layer of each known well from the well bypass seismic data of each known well;
[0037] Among them, the seismic amplitude data at the target layer section of the known well is obtained by the following steps: first, combining the wellhead coordinates (x, y) of the known well and the formulas b=b0+round[(x-x0) / m] and a=a0+round[(y-y0) / m], the seismic track number coordinates (a, b) at the wellhead of the known well are calculated, where (x0, y0) are the coordinates of the initial point in the work area, (a0, b0) are the seismic track number coordinates corresponding to the initial point, m is the track spacing, and round means rounding off; then, combining the seismic track number coordinates at the wellhead of the known well, the wellside track seismic data of the known well is searched from the seismic data of the work area, and then the wellside track seismic data of the target layer section in the known well is obtained by combining the position of the target layer section in the known well, and finally, the seismic amplitude data of the target layer section of the known well is extracted from the wellside track seismic data at the target layer section of the known well.
[0038] As another implementation, the seismic data of the well side track at the target layer section of the known well may also be directly searched from the seismic data of the target layer section of the work area in combination with the seismic line track number coordinates at the wellhead of the known well.
[0039] (3) Normalizing the seismic amplitude data to the interval [-1, 1] and the sand body thickness to the interval [0, 1] to obtain the normalized seismic amplitude data and sand body thickness at the target layer of each known well;
[0040] Since the seismic amplitude data has both positive and negative values, the seismic amplitude data is normalized to the interval [-1,1]; since the sand body thickness has only positive values but no negative values, the sand body thickness is normalized to the interval [0,1], and the larger the value, the thicker the sand body.
[0041] Normalization processing is based on the fact that the numerical variation range of seismic amplitude data and sand body thickness is large, and the amount of calculation required for direct convolution processing is large. By normalizing the seismic amplitude data and sand body thickness and using the normalized data for convolution processing, the speed of convolution processing can be greatly improved.
[0042] (4) The normalized seismic amplitude data at each known well target layer segment is taken as input, and the normalized sand body thickness at the corresponding known well target layer segment is taken as output, so as to obtain a training sample set from the normalized seismic amplitude data and sand body thickness at multiple known well target layer segments, and the ResNet residual network model is trained using the training sample set to obtain a pre-trained amplitude-sand body thickness model.
[0043] ResNet residual network is a deep residual convolutional neural network. Unlike traditional convolutional neural networks, ResNet residual network learns the difference between output and input (i.e. residual). When the residual is 0, it is equivalent to that the stacking layer only performs an identity transformation, and the performance of the network structure will not decrease. In fact, the residual will not be 0, which enables the stacking layer to learn new features based on the input features. Therefore, ResNet residual network has better performance than traditional convolutional neural networks.
[0044] Among them, the commonly used network structures of the ResNet residual network model are ResNet18, ResNet34, ResNet50, ResNet101 and ResNet152. In practical applications, the network structure of the ResNet residual network model can be selected according to actual conditions.
[0045] The network structure used by the ResNet residual network model in this embodiment is ResNet18, which consists of 17 convolutional layers and 1 fully connected layer, with a total of 4 rounds of residual learning. The parameter values in the ResNet residual network model should be filled in according to the actual training data, and the network parameter values should be fine-tuned according to the training data.
[0046] In this embodiment, the normalized seismic amplitude data and sand body thickness are used to train the ResNet residual network model to obtain a pre-trained amplitude-sand body thickness model, and the normalized seismic amplitude data is input into the pre-trained amplitude-sand body thickness model to complete the reservoir prediction. It is easy to know that the sand body thickness predicted at this time is also the normalized sand body thickness; as other implementation methods, the normalization step can also be omitted, and the seismic amplitude data at the known well target layer section is directly used as input, and the sand body thickness at the corresponding known well target layer section is used as output, and the training sample set is obtained by the seismic amplitude data and sand body thickness at multiple known well target layers, that is: during training, the seismic amplitude data and sand body thickness are directly used to train the ResNet residual network model to obtain a pre-trained amplitude-sand body thickness model; during prediction, the seismic amplitude data is directly input into the pre-trained amplitude-sand body thickness model to complete the reservoir prediction. The effectiveness of the method of this embodiment is verified by examples below.
[0047] Affected by the small number of actual wells drilled in the field of seismic reservoir prediction, in this embodiment, the logging data and well-side seismic data of 16 exploratory wells (all vertical wells) that encountered the target layer in the work area were used to train the ResNet18 residual network model. In actual application, the more data involved in the training, the better the reservoir prediction effect.
[0048] Before performing reservoir prediction, the ResNet18 residual network model is trained first. The training process is as follows:
[0049] (1) Construct a training sample set.
[0050] For each of the 16 exploration wells, the gamma logging curve of each well is used to calculate the sand body thickness at the target layer of each well, and the sand body thickness is normalized to the interval [0,1]. Combined with the wellhead coordinates (x, y) of each well and the formulas b=b0+round[(x-x0) / m] and a=a0+round[(y-y0) / m], the seismic line number coordinates (a, b) at the wellhead of each well are calculated, and combined with the seismic line number coordinates at the wellhead of each well, the seismic data of the wellside channel at the target layer of each well are searched from the seismic data of the target layer of the work area, and the seismic amplitude data of the wellside channel seismic data at the target layer of each well are extracted, and the seismic amplitude data are normalized to the interval [-1,1].
[0051] In this embodiment, seismic data with good lateral continuity in the target layer section of the work area are selected, and the time layer is used as the time window. The time window range is a geological layer. 15 seismic data sampling points are extracted from each seismic channel (which can be continuous sampling or random sampling), and the sampling interval is 1ms (the number of sampling points and the sampling interval can be adjusted according to actual needs). Assuming that there are X channels in the horizontal survey line direction and Y channels in the vertical survey line direction in the work area, the seismic data of the target layer section of the work area contains a total of X*Y*15 seismic sampling data bodies. Therefore, the data volume of the wellside channel seismic data at the target layer section of these 16 exploration wells is 16*15, and the data volume of the seismic amplitude data is also 16*15.
[0052] The normalized seismic amplitude data and the normalized sand body thickness are integrated into a 16*15*1*1 four-dimensional array as a training sample set, which is equivalent to 16 pictures. Each picture is a 15*1 pixel point and each picture has a feature value.
[0053] (2) Train the ResNet18 residual network model.
[0054] The ResNet18 residual network model is trained using the training sample set. The training process is as follows: Figure 3-1 and Figure 3-2 As shown in the figure, with the increase of the number of iterations, the root mean square (RMSE) and the loss function (Loss) gradually tend to 0, and the performance of the ResNet18 residual network model is getting better and better. After the training, the weight of each layer in the trained ResNet18 residual network model can characterize the nonlinear relationship between the seismic amplitude data and the sand body thickness.
[0055] (3) Use the trained ResNet18 residual network model for reservoir prediction.
[0056] The sand body thickness of the 21 verification wells in the work area was predicted by inputting the seismic amplitude data of the side channels of the wells into the trained ResNet18 residual network model. The prediction results of the sand body thickness of the 21 verification wells (see Figure 4 ).
[0057] Figure 4 Among them, the predicted results of sand body thickness of these 21 verification wells are consistent with the actual conditions of these 21 verification wells in 16 (as shown in the ellipse in the figure), with a consistency of 76.2%.
[0058] In summary, the method of this embodiment uses the seismic amplitude data and sand body thickness of multiple known wells that have encountered the target layer in the work area to train the ResNet residual network model, fully explores the slight differences between reservoirs and non-reservoirs through deep learning, and uses the trained amplitude-sand body thickness model to establish a nonlinear relationship between seismic amplitude data and sand body thickness, and then uses the trained amplitude-sand body thickness model to achieve reservoir prediction, and verifies the effectiveness of the method of this embodiment with 21 verification wells in the work area. Compared with traditional geophysical theory, the method of this embodiment can improve the accuracy of reservoir prediction, effectively reduce the risk of exploration and development, and has good application prospects in reservoir prediction.
[0059] Method Example 2:
[0060] The ResNet-based seismic reservoir prediction method of this embodiment is different from the ResNet-based seismic reservoir prediction method in method embodiment 1 only in that in this embodiment, the porosity of the target layer segment is predicted by inputting the seismic amplitude data of the target layer segment into a pre-trained amplitude-porosity model to complete the reservoir prediction.
[0061] The amplitude-porosity model is obtained by the following steps:
[0062] Obtain logging data and well bypass seismic data of multiple known wells in the target layer section of the drilling area;
[0063] The porosity of the target layer section of each known well is calculated by using the well logging data of each known well, and the seismic amplitude data of the target layer section of each known well is extracted from the well bypass seismic data of each known well;
[0064] The seismic amplitude data at each known well target layer section is taken as input, and the porosity at the corresponding known well target layer section is taken as output, so as to obtain a training sample set from the seismic amplitude data and porosity at multiple known well target layers. The training sample set is used to train the ResNet residual network model to obtain an amplitude-porosity model.
[0065] Method embodiment 3:
[0066] The ResNet-based seismic reservoir prediction method of this embodiment is different from the ResNet-based seismic reservoir prediction method in method embodiment 1 only in that: in this embodiment, the sand body thickness of the target layer segment is predicted by inputting the seismic amplitude data of the target layer segment into a pre-trained amplitude-sand body thickness model, and the porosity of the target layer segment is predicted by inputting the seismic amplitude data of the target layer segment into a pre-trained amplitude-porosity model, and the reservoir prediction is completed by combining the sand body thickness and porosity of the target layer segment.
[0067] The training process of the amplitude-sand body thickness model is shown in method embodiment 1, and the training process of the amplitude-porosity model is shown in method embodiment 2, which will not be repeated here.
Claims
1. A seismic reservoir prediction method based on ResNet, characterized in that: The method comprises the following steps: (1) acquiring seismic data of a target layer section in a work area, and extracting seismic amplitude data of the target layer section from the seismic data; (2) inputting the seismic amplitude data of the target layer into a pre-trained ResNet residual network model to predict the sand body thickness and / or porosity of the target layer to complete reservoir prediction; The pre-trained ResNet residual network model includes: an amplitude-sand body thickness model and / or an amplitude-porosity model, and the amplitude-sand body thickness model is obtained by the following steps: Obtain logging data and well bypass seismic data of multiple known wells in the target layer section of the drilling area; The thickness of the sand body at the target layer section of each known well is calculated by using the well logging data of each known well, and the seismic amplitude data at the target layer section of each known well is extracted from the well bypass seismic data of each known well; Taking the seismic amplitude data at each known well target layer section as input and the sand body thickness at the corresponding known well target layer section as output, thereby obtaining a first training sample set from the seismic amplitude data and sand body thickness at multiple known well target layer sections, and using the first training sample set to train the ResNet residual network model to obtain an amplitude-sand body thickness model; The amplitude-porosity model is obtained by the following steps: Obtain logging data and well bypass seismic data of multiple known wells in the target layer section of the drilling area; The porosity of the target layer section of each known well is calculated by using the well logging data of each known well, and the seismic amplitude data of the target layer section of each known well is extracted from the well bypass seismic data of each known well; The seismic amplitude data at each known well target layer segment is taken as input, and the porosity at the corresponding known well target layer segment is taken as output, so as to obtain a second training sample set from the seismic amplitude data and porosity at multiple known well target layer segments. The ResNet residual network model is trained using the second training sample set to obtain an amplitude-porosity model.
2. The seismic reservoir prediction method based on ResNet according to claim 1, characterized in that: When the pre-trained ResNet residual network model includes an amplitude-sand body thickness model, the method also includes a normalization step, normalizing the seismic amplitude data to the interval [-1, 1], normalizing the sand body thickness to the interval [0, 1], using the normalized seismic amplitude data and sand body thickness to train the ResNet residual network model to obtain the amplitude-sand body thickness model, and inputting the normalized seismic amplitude data into the amplitude-sand body thickness model to predict the sand body thickness of the target layer segment.
3. The seismic reservoir prediction method based on ResNet according to claim 1 or 2, characterized in that: The seismic amplitude data at the target layer of a known well are obtained by the following steps: First, the wellhead coordinates (x, y) of the known well are combined with the formulas b=b0+round[(x-x0) / m] and a=a0+round[(y-y0) / m] to calculate the seismic line number coordinates (a, b) at the wellhead of the known well, where (x0, y0) are the coordinates of the initial point in the work area, (a0, b0) are the seismic line number coordinates corresponding to the initial point, m is the track spacing, and round means rounding to the nearest integer. ; Then, the wellside channel seismic data of the target layer segment of the known well is searched from the seismic data of the target layer segment of the work area in combination with the seismic track number coordinates at the wellhead of the known well, or the wellside channel seismic data of the known well is first searched from the seismic data of the work area in combination with the seismic track number coordinates at the wellhead of the known well, and then the wellside channel seismic data of the target layer segment of the known well is obtained in combination with the position of the target layer segment in the known well; finally, the seismic amplitude data of the target layer segment of the known well is extracted from the wellside channel seismic data at the target layer segment of the known well.
4. The seismic reservoir prediction method based on ResNet according to claim 3, characterized in that: The network structure of the ResNet residual network model is ResNet18, ResNet34, ResNet50, ResNet101 or ResNet152.
5. The seismic reservoir prediction method based on ResNet according to claim 3, characterized in that: The known well is a vertical well.