Reservoir parameter prediction method suitable for low-permeability reservoir

Abstract

The invention relates to the technical field of exploration physics geology, in particular to a reservoir parameter prediction method suitable for a low-permeability reservoir, and the method comprises the steps: preprocessing data to obtain pre-stack gather data with a high signal-to-noise ratio; the method comprises the following steps: dividing a logging curve, combining a regional rock physical law, determining a limit change range of elastic parameters, randomly disturbing a plurality of divided small layers to obtain a curve sequence of elastic parameter change, and convolution the curve sequence and regional wavelets to form a pre-stack gather reflecting the elastic parameter change as learning label data; taking the pre-stack gather as input of a network model, optimizing trainable parameters of the network model by optimizing a loss function, and obtaining a mapping relation between sample data and elastic parameters; the pre-stack gather data are used as input, transfer learning is carried out on the neural network operator, spatial prediction of the reservoir parameters is realized, and the prediction precision of the reservoir parameters is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of exploration physical geophysics, and more specifically, to a reservoir parameter prediction method applicable to low-permeability reservoirs. Background Art

[0002] Reservoir parameters such as porosity, shale content and gas saturation reflect the reservoir's ability to store and migrate fluids and are important indicators for reservoir evaluation. Accurate prediction of reservoir parameters can provide a more intuitive geophysical basis for the exploration and development of low-permeability reservoirs. However, the strong heterogeneity of underground geological conditions reduces the prediction accuracy of reservoir parameters, making it difficult to predict the "sweet spots" of low-permeability reservoirs, restricting production capacity and affecting the economic and efficient development and production of gas fields. Therefore, in order to reduce the development cost of oil and gas fields and increase production capacity, it is urgent to improve the prediction accuracy of reservoir parameters.

[0003] In actual exploration, the prediction methods of reservoir parameters can be divided into three categories: the first is to obtain the real reservoir parameters through logging interpretation or core data. However, logging data only reflects the reservoir characteristics in a small area around the wellbore, and the cost of drilling is high, so it is impossible to obtain the reservoir parameters of the entire area. The second is seismic wave impedance inversion. This method is based on seismic data and uses geological and logging data to mine the elastic characteristics of underground media from seismic data with limited bandwidth. For example, multiple elastic parameters such as longitudinal wave impedance, density, longitudinal and transverse wave velocity ratio, combined with rock physics laws, can achieve indirect acquisition of reservoir parameter characteristics. At present, seismic inversion has become a generally accepted porosity prediction method in actual production. However, this method still faces the following difficulties. First, this method requires complete pre-stack data and partial angle stacking data, and the calculation cost is high; although the use of model-based inversion methods can effectively reduce the calculation cost, the construction of the model requires seismic interpretation, which not only consumes a lot of manpower costs, but is also easily affected by the subjective understanding of the interpreters. Secondly, the forward operator that accurately characterizes the propagation law of seismic waves is the prerequisite for the authenticity and reliability of the inversion results. In order to take into account the computational efficiency, conventional seismic inversion usually adopts the simplified or approximate forward model, which leads to multiple solutions of the inversion results. Finally, there is no direct analytical relationship between reservoir parameters and classical elastic wave dynamics, which cannot be represented by equations, and therefore cannot be directly obtained through inversion. In actual calculations, it is usually assumed that there is a simple linear relationship between elastic parameters and reservoir parameters, and the longitudinal wave impedance obtained by inversion is approximately equivalent to the elastic parameters to estimate reservoir parameters such as porosity. However, the actual geological conditions are complex, and a single deterministic relationship cannot accurately estimate reservoir parameters. Summary of the invention

[0004] In order to overcome the problem in the above-mentioned prior art that the actual geological conditions are complex and a single deterministic relationship cannot accurately estimate reservoir parameters, the present invention provides a reservoir parameter prediction method suitable for low permeability reservoirs, which can better realize reservoir parameter prediction, guide exploration well site deployment and development capacity release, and help optimize drilling deployment.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a reservoir parameter prediction method applicable to low permeability reservoirs, comprising the following steps:

[0006] S1. Data preprocessing: Modify the seismic data to obtain pre-stack gather data with high signal-to-noise ratio;

[0007] S2. Label data construction: The logging curve is divided into several small layers according to porosity, shale content and water saturation. Then, the limit variation range of elastic parameters is clarified in combination with regional rock physics laws. On this basis, several small layers after division are randomly perturbed to obtain a curve sequence containing elastic parameter changes. The curve sequence is convolved with regional wavelets to form a series of pre-stack gathers reflecting the elastic parameter changes. The pre-stack gathers are used as label data for deep learning. The elastic parameters include porosity, shale content and water saturation.

[0008] S3, training the neural network operator: using the pre-stack gathers obtained in step S2 as the input of the network model, optimizing the trainable parameters of the network model by optimizing the loss function, and obtaining the mapping relationship between the sample data and the elastic parameters;

[0009] S4, actual seismic transfer learning: using the pre-stack gather data obtained in step S1 as input, transfer learning is performed on the neural network operator in step S3 to achieve spatial prediction of reservoir parameters.

[0010] Preferably, after step S4, step S5 is further included, i.e., evaluating the prediction results, calculating the mean square error and correlation coefficient between the predicted reservoir parameter results and the measured curves, so as to quantitatively evaluate the predicted reservoir parameter results.

[0011] Preferably, in step S3, the network architecture of the neural network operator includes an encoder and a decoder, wherein the decoder is used to obtain high-level features, and the decoder expands the high-level features in the spatiotemporal dimension to predict reservoir parameters;

[0012] The step S3 specifically comprises the following steps:

[0013] S31: firstly use the first encoder to extract the common reflection point gather features;

[0014] S32: then use the second encoder and the third encoder to perform fusion and optimization respectively;

[0015] S33: outputting the optimized features using the first encoder, the second encoder, and the third encoder to obtain porosity, shale content, and water saturation;

[0016] The first encoder, the second encoder, and the third encoder all contain two-dimensional convolution, a number of samples for training, a LeakyReLU activation function, and a maximum downsampling layer. In the decoder, only the downsampling layer needs to be replaced by an upsampling layer.

[0017] Preferably, in step S3, the loss function optimized during the training process is as follows:

[0018] L total =λ 1 L 1 +λ 2 L 2 +λ 3 L 3

[0019] Among them, 1 ,λ 2 ,λ 3 is the weight coefficient, L 1 , L 2 , L 3 They represent the mean square error loss functions of porosity, shale content and water saturation respectively, and the formula of the mean square error loss function is as follows:

[0020]

[0021] Where j represents different reservoir parameters, B represents the number of samples used for training in each iteration, nt is the number of sampling points in the time dimension, and Z j The formula is as follows:

[0022]

[0023] Among them, Y t and Y p represent the true value and predicted value of reservoir parameters, respectively.

[0024] Preferably, in step S2, training samples are generated by combining rock physics modeling with random perturbations of porosity, shale content and water saturation, which specifically includes the following steps:

[0025] S21. Rock physics modeling: Use the pressure sampling data of the wells drilled in this area or the rock physics measurement data of the laboratory to carry out rock physics law analysis, build a rock physics model suitable for this area, and use the rock physics model to predict shear waves. Verify the reliability of rock physics modeling by the degree of matching with the measured data;

[0026] S22. Limit range analysis: Obtain the limit variation range of porosity, water saturation and shale content from the drilling rules, and provide geological constraints for subsequent porosity, water saturation and shale content disturbances;

[0027] S23, calculating elastic parameter perturbation curve: dividing and combining the logging curve according to the three dimensions of water saturation, mud content and porosity, constructing several reservoir units, and then perturbing one or more of the units with respect to porosity, water saturation and mud content, and obtaining several groups of curve sequences containing changes in porosity, water saturation and mud content;

[0028] S24, constructing data labels: forward modeling the curve sequence to obtain multiple sets of pre-stack synthetic seismic gathers, and using the forward modeling gathers as learning samples for the neural network.

[0029] Preferably, in step S5, the correlation coefficient R, the goodness of fit R 2 , the root mean square error is an evaluation index used to evaluate whether the predicted reservoir parameters are available. The correlation coefficient R represents the degree of correlation between variables, and the value range is [-1,1]. The larger the positive value, the stronger the positive correlation between the two variables. When the value is 0, it means that the correlation between the two variables is poor.

[0030] Preferably, the formula of the correlation coefficient R is as follows:

[0031]

[0032] where Y t and Y p represent the true value and predicted value of the reservoir parameter, respectively. and They represent the average values ​​of the true and predicted values ​​of reservoir parameters, respectively, and i represents different reservoir parameters.

[0033] Preferably, the goodness of fit R 2 The formula is as follows:

[0034]

[0035] Goodness of fit R 2 It is mainly used to evaluate the degree of fit between variables. The goodness of fit determines the closeness of the correlation. 2The closer it is to 1, the higher the reference value of the relevant fitting results; conversely, the closer it is to 0, the lower the reference value.

[0036] Preferably, the formula for the root mean square error is as follows:

[0037]

[0038] The root mean square error is the square root of the ratio of the square of the deviation between the predicted value and the true value to the number of observations n. It measures the deviation between the predicted value and the true value and is more sensitive to outliers in the data. The smaller its value is, the higher the measurement accuracy is.

[0039] Preferably, in step S1, the modification processing of the seismic data includes residual time difference correction of the gather, residual multiple wave attenuation and random noise suppression.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. The reservoir parameter prediction method suitable for low permeability reservoirs proposed in the present invention avoids the problems of small differences in porosity, mud content, and water saturation in low permeability areas, and low sensitivity of reservoir parameters to rock physics, and can realize the spatial distribution prediction of low permeability reservoir parameters.

[0042] 2. The method of the present invention can directly establish a nonlinear mapping relationship between rock physical parameters and reservoir parameters, thereby reducing the prediction error caused by linear assumptions and improving the prediction accuracy of reservoir parameters.

[0043] 3. The method of the present invention can solve the problems of less labeled data and uneven distribution of training data in actual exploration, and obtain high-precision reservoir parameter prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of a reservoir parameter prediction method applicable to low permeability reservoirs of the present invention;

[0045] Figure 2 It is a typical seismic profile in the study area;

[0046] Figure 3 It is the prediction result of porosity in the reservoir parameter prediction method applicable to low permeability reservoirs of the present invention;

[0047] Figure 4 It is the prediction result of porosity in deep learning method based on a small amount of labeled data;

[0048] Figure 5 It is the prediction result of porosity in the traditional seismic inversion method;

[0049] Figure 6It is the prediction result of the mud content in the reservoir parameter prediction method applicable to low permeability reservoirs of the present invention;

[0050] Figure 7 It is the prediction result of mud content in deep learning method based on a small amount of labeled data;

[0051] Figure 8 It is the prediction result of shale content in the traditional seismic inversion method;

[0052] Fig. 9 It is the prediction result of water saturation in the reservoir parameter prediction method applicable to low permeability reservoirs of the present invention;

[0053] Fig.10 It is the prediction result of water saturation using deep learning method based on a small amount of labeled data;

[0054] Fig.11 It is the prediction result of water saturation in the traditional seismic inversion method;

[0055] Fig.12 is the reservoir parameter cross plot of porosity;

[0056] Fig.13 is the reservoir parameter cross plot of water saturation;

[0057] Fig.14 is the reservoir parameter cross plot of shale content;

[0058] Fig.15 It is the intersection diagram of the predicted porosity and the measured porosity of the present invention;

[0059] Fig.16 It is an intersection diagram of the predicted mud content and the measured mud content of the present invention;

[0060] Fig.17 It is an intersection diagram of the predicted water saturation and the measured water saturation of the present invention. DETAILED DESCRIPTION

[0061] The technical solution of the present invention is further described in detail below through specific embodiments and in conjunction with the accompanying drawings:

[0062] Example 1

[0063] like Figure 1 As shown, a reservoir parameter prediction method applicable to low permeability reservoirs comprises the following steps:

[0064] S1. Data preprocessing: Modify the seismic data to obtain pre-stack gather data with high signal-to-noise ratio;

[0065] S2. Label data construction: The logging curves are divided into several small layers according to porosity, shale content and water saturation. The limit variation range of elastic parameters is clarified by combining regional rock physics laws. On this basis, several small layers after division are randomly perturbed to obtain a curve sequence containing elastic parameter changes. The curve sequence is convolved with regional wavelets to form a series of pre-stack gathers reflecting the elastic parameter changes. The pre-stack gathers are used as label data for deep learning. The elastic parameters include porosity, shale content and water saturation.

[0066] S3, training the neural network operator: using the pre-stack gathers obtained in step S2 as the input of the network model, optimizing the trainable parameters of the network model by optimizing the loss function, and obtaining the mapping relationship between the sample data and the elastic parameters;

[0067] S4, actual seismic transfer learning: using the pre-stack gather data obtained in step S1 as input, transfer learning is performed on the neural network operator in step S3 to achieve spatial prediction of reservoir parameters.

[0068] Wherein, in step S3, the network architecture of the neural network operator includes an encoder and a decoder, wherein the decoder is used to obtain high-level features, and the decoder expands the high-level features in the spatiotemporal dimension to predict the reservoir parameters;

[0069] Step S3 specifically includes the following steps:

[0070] S31: firstly use the first encoder to extract the common reflection point gather features;

[0071] S32: then use the second encoder and the third encoder to perform fusion and optimization respectively;

[0072] S33: outputting the optimized features using the first encoder, the second encoder, and the third encoder to obtain porosity, shale content, and water saturation;

[0073] The first encoder, the second encoder, and the third encoder all contain two-dimensional convolution, a number of samples for training, a LeakyReLU activation function, and a maximum downsampling layer. In the decoder, only the downsampling layer needs to be replaced by an upsampling layer.

[0074] In addition, in step S3, the loss function optimized during training is as follows:

[0075] L total =λ 1 L 1 +λ 2 L 2 +λ 3 L 3

[0076] Among them, 1 ,λ 2 ,λ 3 is the weight coefficient, L 1 , L 2 , L 3 They represent the mean square error loss functions of porosity, shale content and water saturation respectively, and the formula of the mean square error loss function is as follows:

[0077]

[0078] Where j represents different reservoir parameters, B represents the number of samples used for training in each iteration, nt is the number of sampling points in the time dimension, and Z j The formula is as follows:

[0079]

[0080] Among them, Y t and Y p represent the true value and predicted value of reservoir parameters, respectively.

[0081] It should be noted that the optimized loss function can obtain the total error between the predicted porosity and the actual porosity, the predicted mud content and the actual mud content, and the predicted saturation and the actual saturation, which is a multi-parameter objective function; in order to reduce the amount of calculation of the direct gradient, the present invention uses local gradient for approximation, that is, the difference between the predicted value and the true value of a subset of the entire training set is used as the gradient.

[0082] In addition, in order to improve the robustness of this method in actual data, the trained neural network operators are transferred and applied to actual gather data, thereby realizing the spatial prediction of reservoir parameters and improving the generalization ability of the network.

[0083] Among them, in step S2, in data-driven deep learning, labels can be used to supervise the training objectives of deep learning, and their quantity and quality are the prerequisites for the authenticity and reliability of reservoir parameter prediction results. Limited by exploration costs, there are only a small amount of well logging label data in the actual work area, which cannot accurately describe the spatial variation law of complex reservoir parameters underground. In order to solve the problem of less labeled data in actual exploration, this paper proposes a method for constructing labeled data based on rock physics guidance. This method generates training samples by combining rock physics modeling with random perturbations of porosity, shale content and water saturation, and specifically includes the following steps:

[0084] S21. Rock physics modeling: Use the pressure sampling data of the wells drilled in this area or the rock physics measurement data of the laboratory to carry out rock physics law analysis, build a rock physics model suitable for this area, and use the rock physics model to predict shear waves. Verify the reliability of rock physics modeling by the degree of matching with the measured data;

[0085] S22. Limit range analysis: Obtain the limit variation range of porosity, water saturation and shale content from the drilling rules, and provide geological constraints for subsequent porosity, water saturation and shale content disturbances;

[0086] S23, calculating elastic parameter perturbation curve: dividing and combining the logging curve according to the three dimensions of water saturation, mud content and porosity, constructing several reservoir units, and then perturbing one or more of the units with respect to porosity, water saturation and mud content, and obtaining several groups of curve sequences containing changes in porosity, water saturation and mud content;

[0087] S24. Construct data labels: forward modeling is performed on the curve sequence to obtain multiple sets of pre-stack synthetic seismic gathers, and the forward modeling gathers are used as learning samples for the neural network.

[0088] By using the label data construction method based on the forward model, the single well sample can be expanded by 5 orders of magnitude, up to a maximum of more than 100,000, achieving an exponential expansion of training samples and laying the foundation for subsequent deep learning.

[0089] In step S1, the seismic data is modified by processing including residual time difference correction, residual multiple wave attenuation and random noise suppression of the gather.

[0090] Example 2

[0091] The difference between this embodiment 2 and embodiment 1 is that, after step S4, it also includes step S5, prediction result evaluation, calculating the mean square error and correlation coefficient between the predicted reservoir parameter results and the measured curve to quantitatively evaluate the predicted reservoir parameter results.

[0092] Among them, in step S5, the method proposed in the present invention is based on a supervised learning framework, that is, it relies on generating labels to establish a nonlinear mapping relationship. In order to quantitatively analyze the prediction error and evaluate the accuracy difference between the proposed method and the conventional method, the present invention calculates the correlation between the true value and the predicted value, and uses the correlation coefficient R, the goodness of fit R 2 , the root mean square error is an evaluation index used to evaluate whether the predicted reservoir parameters are available. The correlation coefficient R represents the degree of correlation between variables, and the value range is [-1,1]. The larger the positive value, the stronger the positive correlation between the two variables. When the value is 0, it means that the correlation between the two variables is poor.

[0093] In addition, the formula for the correlation coefficient R is as follows:

[0094]

[0095] where Y t and Y p represent the true value and predicted value of the reservoir parameter, respectively. and They represent the average values ​​of the true and predicted values ​​of reservoir parameters, respectively, and i represents different reservoir parameters.

[0096] Among them, the goodness of fit R 2 The formula is as follows:

[0097]

[0098] Goodness of fit R 2 It is mainly used to evaluate the degree of fit between variables. The goodness of fit determines the closeness of the correlation. 2 The closer it is to 1, the higher the reference value of the relevant fitting results; conversely, the closer it is to 0, the lower the reference value.

[0099] In addition, the formula for the root mean square error is as follows:

[0100]

[0101] The root mean square error is the square root of the ratio of the square of the deviation between the predicted value and the true value to the number of observations n. It measures the deviation between the predicted value and the true value and is more sensitive to outliers in the data. The smaller its value is, the higher the measurement accuracy is.

[0102] Example 3

[0103] Based on the same inventive concept, this Example 3 verifies the correctness of the results of the Example. The proposed method is applied to the actual three-dimensional seismic data of the DF area. Its typical seismic profile is as follows Figure 2 As shown in the figure, a typical seismic profile of the well in the area was selected to predict reservoir parameters. The survey line contains 450 CDPs, with a vertical time range of 1.85 to 2s and a sampling interval of 2ms. The target sand body has obvious waterway characteristics, and the main area of ​​the waterway is obviously cut down, showing the overall characteristics of low-frequency continuous strong amplitude seismic phase. There are 2 drilling data in this profile. Well 1 is used as a training well. A total of 6 curves including P-wave velocity, S-wave velocity, density, gas saturation, shale content and porosity curves are used to construct and expand the label data. The curve depth range is 2750 to 3250m, the sampling interval is 0.2m, and a total of 15,000 data points. Well 2 is used as a verification well for cross-validation with the prediction results and does not participate in any calculation link. The three reservoir parameters of porosity, shale content, and saturation are predicted respectively. Figure 3This is a cross-sectional diagram of reservoir parameter prediction results using different methods, where from left to right they represent porosity, mud content, and water saturation, respectively; from top to bottom they represent the method proposed in this paper, the deep learning method based on a small amount of labeled data, and the traditional seismic inversion method, respectively.

[0104] like Figures 9 to 11 The conventional seismic inversion results are greatly affected by the earthquake amplitude. The low porosity, high mud content, high water saturation and other erroneous anomalies caused by the seismic side lobe interference inside the waterway will seriously interfere with the interpretation of the prediction results; the prediction results of a small amount of labeled data (such as Figures 6 to 8 As shown in the figure, due to the small number of learning samples, the seismic feature extraction is insufficient, and the prediction network is underfitted, resulting in high porosity surrounding low porosity inside the waterway, and a sharp vertical mutation of mud content. The prediction results have poor vertical continuity and are quite different from geological knowledge. In contrast, the method of this embodiment (as shown in the figure) Figures 3 to 5 The water channel has high porosity and low mud content inside, while the porosity on the sides of the water channel decreases and the mud content increases. The water saturation shows that the gas saturation is higher at the top of the water channel and lower at the bottom, which is consistent with the geological understanding and actual drilling results.

[0105] In order to further verify the accuracy of the reservoir parameter prediction results of the proposed method, cross-validation was carried out on the validation wells. Figure 12 to Figure 14 This is a consistency analysis diagram of the method of the present invention and the measured data. Fig.12 In the figure, black represents the measured porosity data, and green represents the predicted porosity data; Fig.13 In the figure, black represents the measured water saturation data, and green represents the predicted water saturation data; Fig.14 In the figure, black indicates measured data of mud content, and green indicates predicted data of mud content. In the reservoir section of 3020-3240 m, the predicted results are consistent with the measured trend as a whole. The porosity value is slightly lower than the measured data; the predicted mud content at the bottom of the reservoir is higher than the measured data; and there is a certain error in the prediction of water saturation in the low value area. There are different degrees of errors in the porosity, mud content and water saturation of the non-reservoir section, but they are also within the tolerance range. Figures 15 to 17 The cross plot of this method and the actual reservoir parameters shows that the prediction results are highly correlated with the measured results, and the quantitative evaluation indicators R, R2 and RMSE are 0.9461, 0.9344 and 0.9118 respectively. The prediction errors are mainly concentrated in the high-value area of ​​shale content and the relatively high-value area of ​​porosity. In summary, the proposed method can also obtain accurate reservoir parameter prediction results in low permeability areas.

[0106] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A reservoir parameter prediction method suitable for low permeability reservoirs, characterized in that: The following steps are involved: S1. Data preprocessing: Modify the seismic data to obtain pre-stack gather data with high signal-to-noise ratio; S2. Label data construction: The logging curve is divided into several small layers according to porosity, shale content and water saturation. Then, the limit variation range of elastic parameters is clarified in combination with regional rock physics laws. On this basis, several small layers after division are randomly perturbed to obtain a curve sequence containing elastic parameter changes. The curve sequence is convolved with regional wavelets to form a series of pre-stack gathers reflecting the elastic parameter changes. The pre-stack gathers are used as label data for deep learning. The elastic parameters include porosity, shale content and water saturation. S3, training the neural network operator: using the pre-stack gathers obtained in step S2 as the input of the network model, optimizing the trainable parameters of the network model by optimizing the loss function, and obtaining the mapping relationship between the sample data and the elastic parameters; S4, actual seismic transfer learning: using the pre-stack gather data obtained in step S1 as input, transfer learning is performed on the neural network operator in step S3 to achieve spatial prediction of reservoir parameters.

2. The reservoir parameter prediction method applicable to low permeability reservoirs according to claim 1, characterized in that: After step S4, the method further includes step S5, prediction result evaluation, which calculates the mean square error and correlation coefficient between the predicted reservoir parameter result and the measured curve to quantitatively evaluate the predicted reservoir parameter result.

3. The reservoir parameter prediction method applicable to low permeability reservoirs according to claim 1, characterized in that: In step S3, the network architecture of the neural network operator includes an encoder and a decoder, wherein the decoder is used to obtain high-level features, and the decoder expands the high-level features in the spatiotemporal dimension to predict reservoir parameters; The step S3 specifically comprises the following steps: S31: firstly use the first encoder to extract the common reflection point gather features; S32: then use the second encoder and the third encoder to perform fusion and optimization respectively; S33: outputting the optimized features using the first encoder, the second encoder, and the third encoder to obtain porosity, shale content, and water saturation; The first encoder, the second encoder, and the third encoder all contain two-dimensional convolution, a number of samples for training, a LeakyReLU activation function, and a maximum downsampling layer. In the decoder, only the downsampling layer needs to be replaced by an upsampling layer.

4. The reservoir parameter prediction method applicable to low permeability reservoirs according to claim 1, characterized in that: In step S3, the loss function optimized during the training process is as follows: L total =λ1L1+λ2L2+λ3L3 Where, λ1, λ2, λ3 are weight coefficients, L1, L2, L3 represent the mean square error loss functions of porosity, shale content and water saturation, respectively, and the formula of the mean square error loss function is as follows: Where j represents different reservoir parameters, B represents the number of samples used for training in each iteration, nt is the number of sampling points in the time dimension, and Z j The formula is as follows: Among them, Y t and Y p represent the true value and predicted value of reservoir parameters, respectively.

5. The reservoir parameter prediction method applicable to low permeability reservoirs according to claim 1, characterized in that: In step S2, training samples are generated by combining rock physics modeling with random perturbations of porosity, shale content, and water saturation, specifically including the following steps: S21. Rock physics modeling: Use the pressure sampling data of the wells drilled in this area or the rock physics measurement data of the laboratory to carry out rock physics law analysis, build a rock physics model suitable for this area, and use the rock physics model to predict shear waves. Verify the reliability of rock physics modeling by the degree of matching with the measured data; S22. Limit range analysis: Obtain the limit variation range of porosity, water saturation and shale content from the drilling rules, and provide geological constraints for subsequent porosity, water saturation and shale content disturbances; S23, calculating elastic parameter perturbation curve: dividing and combining the logging curve according to the three dimensions of water saturation, mud content and porosity, constructing several reservoir units, and then perturbing one or more of the units with respect to porosity, water saturation and mud content, and obtaining several groups of curve sequences containing changes in porosity, water saturation and mud content; S24, constructing data labels: forward modeling the curve sequence to obtain multiple sets of pre-stack synthetic seismic gathers, and using the forward modeling gathers as learning samples for the neural network.

6. The reservoir parameter prediction method applicable to low permeability reservoirs according to claim 2, characterized in that: In step S5, the correlation coefficient R and the goodness of fit R are used. 2 , the root mean square error is an evaluation index used to evaluate whether the predicted reservoir parameters are available. The correlation coefficient R represents the degree of correlation between variables, and its value range is [-1,1]. The larger the positive value, the stronger the positive correlation between the two variables. When the value is 0, it means that the correlation between the two variables is poor.

7. The reservoir parameter prediction method applicable to low permeability reservoirs according to claim 6, characterized in that: The formula for the correlation coefficient R is as follows: where Y t and Y p represent the true value and predicted value of the reservoir parameter, respectively. and They represent the average values ​​of the true and predicted values ​​of reservoir parameters, respectively, and i represents different reservoir parameters.

8. The reservoir parameter prediction method applicable to low permeability reservoirs according to claim 7, characterized in that: Goodness of fit R 2 The formula is as follows: Goodness of fit R 2 It is mainly used to evaluate the degree of fit between variables. The goodness of fit determines the closeness of the correlation. 2 The closer it is to 1, the higher the reference value of the relevant fitting results; On the contrary, the closer it is to 0, the lower the reference value.

9. The reservoir parameter prediction method applicable to low permeability reservoirs according to claim 8, characterized in that: The formula for the root mean square error is as follows: The root mean square error is the square root of the ratio of the square of the deviation between the predicted value and the true value to the number of observations n. It measures the deviation between the predicted value and the true value and is more sensitive to outliers in the data. The smaller its value is, the higher the measurement accuracy is.

10. The reservoir parameter prediction method applicable to low permeability reservoirs according to any one of claims 1 to 9, characterized in that: In step S1, the seismic data is modified by processing including residual time difference correction, residual multiple wave attenuation and random noise suppression of the gather.

Images