Earthquake sand body prediction method based on dynamic progressive search algorithm

Abstract

The invention relates to the technical field of sand body content prediction, in particular to an earthquake sand body prediction method based on a dynamic progressive search algorithm. Comprises: collecting data; establishing a neural network model; training a neural network model; obtaining a local optimal connection weight correction; setting a threshold range, and obtaining a first weight correction; when the first weight correction amount is not within the threshold range, carrying out second iterative calculation on the first weight correction amount; obtaining a global optimal connection weight correction amount; obtaining a trained neural network model; and predicting sand body parameters. According to the method, the connection weight is corrected by adopting the error, the threshold range is set for the first weight correction of the connection weight, and when the first weight correction exceeds the set threshold range, the second iterative calculation is performed on the first weight correction, so that the network input is prevented from falling into the saturation region of the activation function, and the network performance is improved. The training stability of the neural network model is improved, and the precision of earthquake prediction sand body is improved.

Description

Technical Field

[0001] The invention relates to the technical field of sand body content prediction, and more specifically, to a seismic sand body prediction method based on a dynamic progressive search algorithm. Background Art

[0002] A neural network is a computational model consisting of a large number of interconnected nodes. Each node represents a specific output function, called an activation function. The connection between each two nodes represents a weighted value for the signal passing through the connection, called a weight, also known as a connection weight, which is equivalent to the memory of an artificial neural network. The output of the network varies depending on the network's connection method, connection weights, and activation functions. The network itself is usually an approximation of a natural algorithm or function, or it may be an expression of a logical strategy.

[0003] With the in-depth research of big data and deep learning in the fields of science and engineering, neural network technology has the characteristics of strong generalization ability and high robustness, and can effectively establish a multi-dimensional nonlinear relationship between seismic data and sand body parameters. It has been widely used in the fields of oil and gas reservoir sand body prediction. However, the traditional correction algorithm uses the gradient descent method to adjust the connection weights of the nodes of the neural network model. Since the neural network model is nonlinear, if the connection weight correction is too large, the network input will fall into the saturation area of ​​the activation function, resulting in an extremely small derivative of the objective function, which makes the weight correction process in the denominator of the formula impossible to calculate, resulting in low stability of the neural network model training. Summary of the invention

[0004] The purpose of the present invention is to provide a seismic sand body prediction method based on a dynamic progressive search algorithm, which effectively improves the training stability of a neural network model by correcting the connection weights, thereby achieving improved accuracy in seismic sand body prediction.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] A seismic sand body prediction method based on a dynamic progressive search algorithm comprises:

[0007] Acquiring data: acquiring data on seismic attributes and sand body parameters in the seismic work area, wherein the data on seismic attributes include data on seismic attributes for training and data on seismic attributes for prediction; the seismic attributes include root mean square amplitude, arc length, instantaneous frequency, half energy, average amplitude and energy;

[0008] Establishing a neural network model: taking the seismic attributes as input and the sand body parameters as output, and establishing an objective function;

[0009] Training the neural network model: inputting the seismic attribute data and sand body parameter data used for training into the neural network model to train the neural network model; the training of the neural network model includes the following steps:

[0010] Obtaining a local optimal connection weight correction amount: setting a threshold range, performing a first correction on the connection weight based on the objective function, and obtaining a first weight correction amount; when the first weight correction amount is within the threshold range, the first weight correction amount is the local optimal connection weight correction amount; when the first weight correction amount is not within the threshold range, performing a second iterative calculation on the first weight correction amount to obtain a second weight correction amount, and the second weight correction amount is the local optimal connection weight correction amount;

[0011] Obtaining a global optimal connection weight correction amount: correcting the weight correction amounts of adjacent nodes based on the local optimal connection weight correction amount to obtain a third weight correction amount, wherein the third weight correction amount is a global optimal connection weight correction amount;

[0012] Obtaining a trained neural network model: obtaining an optimal connection weight through the global optimal connection weight correction amount to obtain a trained neural network model;

[0013] Predict sand body parameters: Based on the trained neural network model, input the data of seismic attributes used for prediction and predict the data of sand body parameters.

[0014] According to the above technical means, the present invention provides a seismic sand body prediction method based on a dynamic progressive search algorithm, which establishes a connection relationship between seismic attributes and sand body parameters through a neural network model, and can effectively establish a multidimensional nonlinear relationship between seismic attributes and sand body parameters. The connection weight is corrected by using an objective function, and a threshold range is set for the first weight correction amount of the connection weight. When the first weight correction amount is not within the set threshold range, the first weight correction amount is iteratively calculated for the second time to avoid the network input falling into the saturation zone of the activation function, improve the training stability of the neural network model, and then achieve the purpose of improving the accuracy of seismic sand body prediction. After obtaining the local optimal connection weight correction amount, the weight correction amount of the adjacent nodes is corrected to obtain the global optimal connection weight correction amount, so as to avoid the prediction result falling into the local optimum and further improve the accuracy of seismic sand body prediction.

[0015] Furthermore, in the process of obtaining the global optimal connection weight correction, a deviation range is set, and correction is stopped when the adjacent connection weight calculation correction is zero; when correction is stopped, if the deviation between the third weight correction and the first weight correction is within the deviation range, the third weight correction is the global optimal connection weight correction; when correction is stopped, if the deviation between the third weight correction and the first weight correction is not within the deviation range, the third weight correction is iteratively calculated for the third time to obtain a fourth weight correction, which is the global optimal connection weight correction. When correcting the weight correction of adjacent nodes, if the third weight correction deviates greatly from the first weight correction, it will still cause the network input to fall in the saturation area of ​​the activation function. By adding the third iterative calculation, the large deviation between the third weight correction and the first weight correction is avoided, and the stability of the neural network model training is improved, thereby achieving the improvement of the accuracy of earthquake prediction sand bodies.

[0016] Furthermore, the neural network model includes an input layer, an intermediate layer and an output layer, and adjacent layers are connected by nodes; obtaining the local optimal connection weight correction amount and obtaining the global optimal connection weight correction amount are both to correct and iterate the weight correction amount of the input layer and the intermediate layer, and the connection weight of the output layer and the intermediate layer is calculated according to the optimal connection weight of the input layer and the intermediate layer. By adding an intermediate layer to the neural network model, the accuracy of seismic prediction of sand bodies can be improved. The connection weight of the output layer and the intermediate layer is calculated by the optimal connection weight of the input layer and the intermediate layer, which can simplify the algorithm.

[0017] Furthermore, the objective function is established according to the principle of minimizing the error between the forward expected output value of the neural network model and the known actual value, and specifically adopts the following formula:

[0018]

[0019] Among them, E represents the global error, Y represents the known actual value, and x i represents the earthquake attribute of the ith node in the input layer, w ij represents the connection weight between the i-th node in the input layer and the j-th node in the middle layer, v j represents the connection weight between the output layer node and the jth node in the middle layer, α i Indicates the threshold value.

[0020] Furthermore, in the process of obtaining the local optimal connection weight correction amount,

[0021] The connection weight is first corrected based on the objective function, and the objective function is solved by the gradient descent method to obtain the first weight correction amount, which is specifically adopted by the following formula:

[0022]

[0023] Among them, Δw ij represents the initial correction value corresponding to the connection weight between the i-th node in the input layer and the j-th node in the middle layer, w ij represents the connection weight between the i-th node in the input layer and the j-th node in the middle layer, x i represents the earthquake attribute of the i-th node in the input layer, α i represents the threshold value, v j represents the connection weight between the output layer node and the jth node in the middle layer, and η represents the learning rate, which ranges from [0,1].

[0024] In the second iterative calculation of the first weight correction value, a maximum number of iterations is set to obtain a second weight correction value, specifically using the following formula:

[0025]

[0026] in, It represents the value of the connection weight correction between the i-th node in the input layer and the j-th node in the middle layer at the t-th iteration. represents the value of the connection weight correction between the i-th node in the input layer and the j-th node in the middle layer at the t+1th iteration, i represents the i-th node in the input layer, β represents a random number ranging from [-1,1], T max Indicates the maximum number of iterations.

[0027] Furthermore, in the process of obtaining the global optimal connection weight correction amount,

[0028] The third weight correction amount is specifically formulated using the following formula:

[0029]

[0030] in, It means that after p iterations, the local optimal connection weight correction between the i-th node in the input layer and the j-th node in the middle layer is obtained, Δw i,j+1 represents the weight correction of the adjacent nodes in the middle layer, D represents a 1×n matrix, n corresponds to the total number of nodes in the middle layer, each element in the matrix is ​​randomly assigned 1 or -1, and D + =D T (DD T ) -1 .

[0031] The fourth weight correction amount is specifically expressed by the following formula:

[0032]

[0033] in, It means that after t iterations, the weight correction of the i-th node in the input layer and the j+1-th node in the middle layer is obtained. It means that after t+1 iterations, the weight correction of the i-th node in the input layer and the j+1-th node in the middle layer is obtained. It indicates that the global optimal weight correction of the i-th node in the input layer and the j-th node in the middle layer is obtained after B iterations, and β indicates a random number in the range of [-1,1].

[0034] Furthermore, the connection weight between the output layer and the intermediate layer is calculated according to the optimal connection weight between the input layer and the intermediate layer, specifically using the following formula:

[0035]

[0036] Among them, w ij represents the connection weight between the i-th node in the input layer and the j-th node in the middle layer, x i represents the earthquake attribute of the i-th node in the input layer, α i represents the threshold value, v j represents the connection weight between the output layer node and the jth node in the middle layer, η represents the learning rate in the range [0,1], and Y represents the known actual value.

[0037] Furthermore, the collected data is specifically:

[0038] Extract the data of seismic attributes of each seismic survey line in the three-dimensional seismic area and the two-dimensional seismic area corresponding to the target layer respectively, wherein the data includes three columns of numbers, namely, X coordinate, Y coordinate and seismic attribute value; extract the data of sand body parameters of the seismic area where the seismic attribute data used for training is located;

[0039] Performing standard deviation standardization algorithm processing on the extracted seismic attribute data respectively;

[0040] The two-dimensional and three-dimensional seismic attribute data processed by the standard deviation normalization algorithm are combined. For data points with the same coordinates, they are replaced according to the attribute average value to obtain fused seismic attribute data; by fusing the two-dimensional and three-dimensional seismic attributes, the distribution morphology of sand bodies in a larger area can be analyzed.

[0041] The fused seismic attribute data used for training in the fused seismic attribute data are subjected to correlation analysis with the data of sand body parameters, and the seismic attributes with statistical significance after the correlation analysis are selected as the preferred seismic attributes; in the process of establishing the neural network model, the preferred seismic attributes are used as input. The accuracy of sand body parameter prediction can be improved by selecting the seismic attributes with statistical significance as the preferred seismic attributes through correlation analysis.

[0042] Furthermore, the extracted seismic attribute data are processed by a standard deviation standardization algorithm, specifically using the following formula:

[0043] a * =(a-mean) / std

[0044] Among them, a * represents the seismic attribute data after standardization, a represents the original seismic attribute data, mean represents the mean of the seismic attribute data, and std represents the standard deviation of the seismic attribute data.

[0045] Furthermore, the prediction of sand body parameters is specifically as follows: inputting the fused seismic attribute data used for prediction in the fused seismic attribute data of the study area into the trained neural network model to obtain the simulation output of the sand body parameters; gridding the simulation output of the sand body parameters to obtain the plane distribution map of the sand body parameters in the study area, and completing the prediction of the sand body distribution. Through gridding, the prediction of the sand body distribution can be seen more clearly and intuitively.

[0046] Compared with the prior art, the beneficial effects are:

[0047] 1. The present invention provides a seismic sand body prediction method based on a dynamic progressive search algorithm. When the connection weight is corrected by using an objective function, a threshold range is set for the first weight correction amount of the connection weight; when the first weight correction amount is not within the set threshold range, a second iterative calculation is performed on the first weight correction amount to avoid the network input falling in the saturation area of ​​the activation function, thereby improving the stability of the neural network model training and improving the accuracy of seismic sand body prediction.

[0048] 2. The present invention provides a seismic sand body prediction method based on a dynamic progressive search algorithm, which avoids a large deviation between the third weight correction amount and the first weight correction amount by adding a third iterative calculation, thereby improving the stability of neural network model training and further improving the accuracy of seismic sand body prediction.

[0049] 3. The present invention provides a seismic sand body prediction method based on a dynamic progressive search algorithm, which can analyze the distribution morphology of sand bodies in a larger area by fusing two-dimensional and three-dimensional seismic attributes. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of the seismic sand body prediction method based on the dynamic progressive search algorithm of the present invention;

[0051] Figure 2 It is a flow chart of the neural network model training of the present invention;

[0052] Figure 3is a cross-plot of seismic attributes and sand content of the present invention, wherein Figure a is a cross-plot of root mean square amplitude and sand content, Figure b is a cross-plot of arc length and sand content, Figure c is a cross-plot of instantaneous frequency and sand content, Figure d is a cross-plot of half energy and sand content, Figure e is a cross-plot of average amplitude and sand content, and Figure f is a cross-plot of energy and sand content;

[0053] Figure 4 is a correlation coefficient diagram of seismic attributes and sand body content of the present invention;

[0054] Figure 5 It is a comparison diagram of the prediction of sand body content by the neural network of the present invention;

[0055] Figure 6 This is a comparison chart of sand content prediction using traditional neural network;

[0056] Figure 7 This is a prediction map of sandstone content in the first section of the Huangliu Formation in a certain basin;

[0057] Figure 8 It is a comprehensive logging diagram of drilling, among which Figure a is the comprehensive logging diagram of drilling LD-E, Figure b is the comprehensive logging diagram of drilling LD-I, and Figure c is the comprehensive logging diagram of drilling LT-F. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. The present invention is described in one of the embodiments in combination with the specific implementation methods. Among them, the drawings are only used for exemplary descriptions, and only schematic diagrams are shown, not physical drawings, and cannot be understood as limitations on the present invention; in order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0059] In the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right" and the like indicate an orientation or positional relationship based on the orientation or positional relationship shown in the drawings, it is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the drawings are only used for exemplary explanations and cannot be understood as limiting the present invention. For ordinary technicians in the field, the specific meanings of the above terms can be understood according to specific circumstances. In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only used for descriptive purposes, and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" can explicitly or implicitly include at least one of the features. In addition, the meaning of "and / or" appearing in the full text is to include three parallel schemes. Taking "A and / or B" as an example, it includes scheme A, or scheme B, or schemes that satisfy both A and B.

[0060] Example 1

[0061] like Figure 1 and Figure 2 As shown, this embodiment proposes a seismic sand body prediction method based on a dynamic progressive search algorithm, including:

[0062] Collect data: collect data on seismic attributes and sand body parameters in the seismic work area; the seismic attributes include root mean square amplitude, arc length, instantaneous frequency, half energy, average amplitude and energy, or other seismic attributes. Sand body parameters refer to sand body content, etc. The specific sand body parameters are set according to the parameters to be predicted. The seismic attribute data include data on seismic attributes for training and data on seismic attributes for prediction, wherein the data on seismic attributes for training are seismic attribute data for the entire work area, the data on seismic attributes for prediction are seismic attribute data for part of the work area, and the data on sand body parameters are data on sand body parameters corresponding to the part of the work area. The sand body parameters of the part of the work area and the seismic attribute data of the part of the work area corresponding to the sand body parameters are used as standard sample sets to train the neural network model.

[0063] Establishing a neural network model: taking the seismic attributes as input and the sand body parameters as output, establishing an objective function; the neural network model may have only an input layer and an output layer, and the number of intermediate layers is not limited here.

[0064] Training the neural network model: inputting the seismic attribute data and sand body parameter data used for training into the neural network model to train the neural network model; the training of the neural network model includes the following steps:

[0065] Obtaining the local optimal connection weight correction: setting a threshold range, where the threshold range is set according to actual needs, the purpose is to prevent the network input from falling in the saturation zone of the activation function, resulting in an extremely small derivative of the objective function, and thus making the weight correction process in the denominator of the formula impossible to calculate, and performing a first correction on the connection weight based on the objective function to obtain a first weight correction; when the first weight correction is within the threshold range, the first weight correction is the local optimal connection weight correction; when the first weight correction is not within the threshold range, performing a second iterative calculation on the first weight correction to obtain a second weight correction, which is the local optimal connection weight correction; that is, it is possible to prevent the network input from falling in the saturation zone of the activation function, thereby improving the stability of the neural network model training, and thereby achieving the goal of improving the accuracy of seismic sand body prediction.

[0066] Obtaining a global optimal connection weight correction amount: based on the local optimal connection weight correction amount, correcting the weight correction amount of the adjacent nodes to obtain a third weight correction amount, wherein the third weight correction amount is a global optimal connection weight correction amount, which can avoid causing the prediction result to fall into a local optimum and further improve the accuracy of seismic sand body prediction;

[0067] Obtaining a trained neural network model: obtaining an optimal connection weight through the global optimal connection weight correction amount, that is, performing a simple calculation between the third weight correction amount and the initially set connection weight to obtain a trained neural network model;

[0068] Predict sand body parameters: Based on the trained neural network model, input the data of seismic attributes used for prediction, and predict the data of sand body parameters. The seismic attributes input here are the seismic attribute data of the entire work area required to be predicted in the data acquisition step. The sand body parameters of the entire work area are predicted by the trained neural network model.

[0069] According to the above technical means, the present invention provides a seismic sand body prediction method based on a dynamic progressive search algorithm, which adopts a dynamic progressive search method to correct the connection weights in the network model, improve the training stability of the neural network model, and thus improve the accuracy of seismic sand body prediction.

[0070] Example 2

[0071] This embodiment proposes a seismic sand body prediction method based on a dynamic progressive search algorithm. On the basis of Embodiment 1, in this embodiment, Figure 2As shown, the neural network model includes an input layer, an intermediate layer and an output layer, and adjacent layers are connected by nodes; the number of nodes in the input layer is consistent with the number of preferred seismic attributes, the number of nodes in the output layer is set to one, and the number of nodes in the intermediate layer is set to twice the sum of the number of nodes in the input layer and the output layer. By adding an intermediate layer to the neural network model, the accuracy of earthquake prediction of sand bodies can be improved.

[0072] According to the principle of minimizing the error between the forward expected output value of the neural network model and the known actual value, the objective function is established, and the following formula is used specifically:

[0073]

[0074] Among them, E represents the global error, Y represents the known actual value, and x i represents the earthquake attribute of the ith node in the input layer, w ij represents the connection weight between the i-th node in the input layer and the j-th node in the middle layer, v j represents the connection weight between the output layer node and the jth node in the middle layer, α i Indicates the threshold value.

[0075] Obtaining the local optimal connection weight correction amount and obtaining the global optimal connection weight correction amount are both to iterate the correction of the weight correction amount of the input layer and the middle layer, and the connection weight of the output layer and the middle layer is calculated according to the optimal connection weight of the input layer and the middle layer. Calculating the connection weight of the output layer and the middle layer by the optimal connection weight of the input layer and the middle layer can simplify the algorithm.

[0076] The neural network model is trained by a given standard sample set. The neural network model training process includes forward propagation and back propagation. In the forward propagation process, the seismic attributes are transmitted to the output layer through the input layer and the intermediate layer. If the sand body parameters obtained by the output layer are too different from the actual sand body parameters, back propagation begins. The error propagates from the output layer through the intermediate layer to the input layer, and the connection weights of the nodes between the layers are adjusted along the way to continuously reduce the error. After repeated iterative learning and training, the global error of the network model tends to the set minimum value.

[0077] First, obtain the local optimal connection weight correction. Set the first weight correction threshold range. The threshold range is set according to actual needs. The purpose is to prevent the network input from falling into the saturation area of ​​the activation function, resulting in an extremely small derivative of the objective function, and thus making the weight correction process in the denominator of the formula impossible to calculate. Based on the objective function, the connection weights of the input layer and the middle layer are first corrected, and the gradient descent method is used to solve the objective function until the global error meets the requirements. The first weight correction is obtained. When the first weight correction is within the threshold range, the first weight correction is the local optimal connection weight correction, and the first weight correction is basically close to the accurate value; the first weight correction is specifically formulated in the following formula:

[0078]

[0079] Among them, Δw ij represents the initial correction value corresponding to the connection weight between the i-th node in the input layer and the j-th node in the middle layer, w ij represents the connection weight between the i-th node in the input layer and the j-th node in the middle layer, x i represents the earthquake attribute of the i-th node in the input layer, α i represents the threshold value, v j represents the connection weight between the output layer node and the jth node in the middle layer, and η represents the learning rate, which ranges from [0,1].

[0080] When the first weight correction value is not within the threshold range, the first weight correction value is iterated for the second time, and the maximum number of iterations is set to obtain the second weight correction value. The second weight correction value is specifically calculated using the following formula:

[0081]

[0082] in, It represents the value of the connection weight correction between the i-th node in the input layer and the j-th node in the middle layer at the t-th iteration. represents the value of the connection weight correction between the i-th node in the input layer and the j-th node in the middle layer at the t+1th iteration, i represents the i-th node in the input layer, β represents a random number ranging from [-1,1], T max Indicates the maximum number of iterations.

[0083] After p iterations, after obtaining the local optimal connection weight correction, the weight correction of the adjacent nodes is corrected based on the local optimal connection weight correction, and the deviation range is set. The deviation range is set according to actual needs to prevent the network input from falling into the saturation area of ​​the activation function, resulting in an extremely small derivative of the objective function, and thus making the weight correction process in the denominator of the formula impossible to calculate. When the adjacent connection weight calculation correction is zero, the correction is stopped to obtain the third weight correction; the distance between adjacent nodes does not exceed half of the total number of nodes in the middle layer of the network. The third weight correction is specifically calculated using the following formula:

[0084]

[0085] in, It means that after p iterations, the local optimal connection weight correction between the i-th node in the input layer and the j-th node in the middle layer is obtained, Δw i,j+1 represents the weight correction of the adjacent nodes in the middle layer, D represents a 1×n matrix, n corresponds to the total number of nodes in the middle layer, each element in the matrix is ​​randomly assigned 1 or -1, and D + =D T (DD T ) -1 .

[0086] When the correction is stopped, if the deviation between the third weight correction amount and the first weight correction amount is within a deviation range, the third weight correction amount is the global optimal connection weight correction amount.

[0087] When the correction is stopped, if the deviation between the third weight correction amount and the first weight correction amount is not within the deviation range, the third weight correction amount is iteratively calculated for the third time to obtain the fourth weight correction amount, and the fourth weight correction amount is the global optimal connection weight correction amount. When correcting the weight correction amount of adjacent nodes, if the third weight correction amount deviates greatly from the first weight correction amount, it will still cause the network input to fall in the saturation area of ​​the activation function. By adding the third iterative calculation, the large deviation between the third weight correction amount and the first weight correction amount can be avoided, and the training stability of the neural network model can be improved, thereby achieving the improvement of the accuracy of earthquake prediction sand bodies. The fourth weight correction amount is specifically calculated using the following formula:

[0088]

[0089] in, It means that after t iterations, the weight correction of the i-th node in the input layer and the j+1-th node in the middle layer is obtained. It means that after t+1 iterations, the weight correction of the i-th node in the input layer and the j+1-th node in the middle layer is obtained. It indicates that the global optimal weight correction of the i-th node in the input layer and the j-th node in the middle layer is obtained after B iterations, and β indicates a random number in the range of [-1,1].

[0090] The optimal connection weight is obtained through the fourth weight correction amount. Through these four steps, the connection weights of the neural network model are repeatedly iterated and corrected, so that all connection weights of the input layer and the middle layer in the neural network model reach the global optimum.

[0091] The connection weight between the output layer and the middle layer is calculated based on the optimal connection weight between the input layer and the middle layer, using the following formula:

[0092]

[0093] Among them, w ij represents the connection weight between the i-th node in the input layer and the j-th node in the middle layer, x i represents the earthquake attribute of the i-th node in the input layer, α i represents the threshold value, v j represents the connection weight between the output layer node and the jth node in the middle layer, η represents the learning rate in the range [0,1], and Y represents the known actual value.

[0094] After obtaining the optimal connection weights between the output layer and the middle layer and between the input layer and the middle layer, the neural network training task can be completed, and all the connection weights of the input layer, middle layer and output layer in the neural network model can reach the global optimum, thus obtaining a trained neural network model.

[0095] Example 3

[0096] This embodiment proposes a method for predicting seismic sand bodies based on a dynamic progressive search algorithm. Based on Embodiment 1 or Embodiment 2, in this embodiment, the data acquisition is specifically as follows: extracting the seismic attribute data of each seismic survey line in the three-dimensional seismic area and the two-dimensional seismic area corresponding to the target layer, and extracting as many types of seismic attributes as possible from the perspectives of amplitude, frequency, energy, etc. The data contains three columns, namely, the X coordinate, the Y coordinate, and the seismic attribute value; by extracting the seismic attribute data of the three-dimensional seismic area and the two-dimensional seismic area, the scope of sand body prediction can be expanded, so that the sand body prediction scope is not limited to only the three-dimensional seismic area or only the two-dimensional seismic area. And extract the data of the sand body parameters of the seismic area where the data of the seismic attributes used for training are located.

[0097] The extracted seismic attribute data are processed by standard deviation standardization algorithm; the specific formula is as follows:

[0098] a * =(a-mean) / std

[0099] Among them, a * represents the seismic attribute data after standardization, a represents the original seismic attribute data, mean represents the mean of the seismic attribute data, and std represents the standard deviation of the seismic attribute data.

[0100] The extracted seismic attribute data are processed by the standard deviation normalization algorithm. The data after standardization conforms to the standard normal distribution. The peak value of the data distribution is set to zero. The seismic attribute data only retains the data in the range of [-3, 3], which can avoid the interference of outlier attribute abnormal data.

[0101] Merge the two-dimensional and three-dimensional seismic attribute data processed by the standard deviation normalization algorithm, and replace the data points with the same coordinates according to the attribute average value to obtain the fused seismic attribute data. By fusing the two-dimensional and three-dimensional seismic attributes, the plane range of the seismic data mapping is expanded, and the sand body distribution morphology in a larger area can be analyzed.

[0102] The fused seismic attribute data used for training in the fused seismic attribute data is subjected to correlation analysis with the data of sand body parameters. The correlation analysis can be performed by existing software, correlation coefficient, cross plot and other methods, and the seismic attributes with statistical significance after the correlation analysis are selected as the preferred seismic attributes; the seismic attributes with statistical significance means that the correlation coefficient between the seismic attributes and the sand body parameters is greater than 0.2. The number of seismic attributes selected here is not limited, and only the seismic attributes with the largest correlation coefficient can be selected, or several or all of them with statistical significance can be selected. In the process of establishing the neural network model, the preferred seismic attributes are used as input. The accuracy of sand body parameter prediction can be improved by selecting the seismic attributes with statistical significance as the preferred seismic attributes through correlation analysis.

[0103] The predicted sand body parameters are specifically as follows: inputting the fused seismic attribute data used for prediction in the fused seismic attribute data of the study area into the trained neural network model, obtaining the sand body parameter data consistent with the fused seismic attribute data range through the forward output of the neural network model; gridding the sand body parameter data into a map, obtaining the plane distribution map of the sand body parameters in the study area, and completing the sand body distribution prediction. By gridding into a map, the sand body distribution prediction can be seen more clearly and intuitively.

[0104] Example 4

[0105] This embodiment applies a seismic sand body prediction method based on a dynamic progressive search algorithm to predict the sand body content of the first section of the Huangliu Formation in a certain basin. On the basis of Example 1, Example 2 and Example 3, in this embodiment, the first section of the Huangliu Formation in a certain basin is taken as the study area, and the study area includes two-dimensional and three-dimensional seismic work areas.

[0106] First, data is collected and fused into two-dimensional and three-dimensional seismic attribute data. Due to the different dimensions between the attributes of the three-dimensional and two-dimensional seismic work areas, especially the two-dimensional seismic survey lines, the study area has been collecting data in batches since 1979, and the magnitude of seismic values ​​varies greatly, resulting in local anomalies often being submerged in the regional background, and the existence of some outlier abnormal values. Therefore, the seismic attributes must be standardized and preprocessed before two-dimensional and three-dimensional data fusion can be performed.

[0107] Considering that there are a large number of outliers and extreme values ​​in seismic data, the global maximum and minimum values ​​of the data change very quickly. The traditional normalization processing method relies too much on the maximum and minimum values. The abnormal seismic line data of individual data will cause the overall trend to be unclear. The standard deviation normalization processing method is suitable for the situation where there are outliers or the data is out of range. This method is especially suitable for complex two-dimensional seismic lines. The two-dimensional seismic and three-dimensional seismic attributes corresponding to the first section of Huang in the study area are extracted, and the seismic attributes corresponding to each two-dimensional line and three-dimensional data body are standardized by standard deviation. The two-dimensional and three-dimensional attribute fusion processing process is completed according to the above specific methods.

[0108] The second step is to conduct correlation analysis. The sandstone content parameters of the first section of Huang in the study area are statistically analyzed, and the fused seismic attribute data corresponding to the well bypass are extracted. The sensitivity of various seismic attributes to sand bodies is analyzed using the cross plot and correlation coefficient method. The arc length, average amplitude, energy, root mean square and half energy of the first section of Huang in the study area are extracted for sand body sensitivity analysis. Through the cross plot of various seismic attributes and sand body content, such as Figure 3 As shown in Figure 2, combined with the correlation coefficients between various seismic attributes and sand body content, such as Figure 4 As shown in the figure, it can be seen that the correlation between a single seismic attribute and sand content is not very high, and the best correlation is the root mean square seismic attribute, with a correlation coefficient of only 0.545. Therefore, it is very necessary to extract the sand body information contained in multiple seismic attributes (seismic waves) and use a combination of multiple seismic attributes to predict sand bodies. After analysis, it was finally confirmed that the seismic attributes with higher sensitivity to sandstone in Huangliu Group 1 are root mean square, arc length and half energy, and these three seismic attribute combinations are used as the most sensitive seismic attribute combination of the sand body in Huangyi Section 1.

[0109] The third step is to build a neural network model. The neural network model input layer is set to 3 nodes, the output layer is set to 1 node, and the number of nodes in the middle layer is set to 8.

[0110] The fourth step is to modify the connection weights. The drilling statistical sand content parameters and well side channel seismic attributes are combined into a standard sample set to train the neural network model.

[0111] In order to evaluate the prediction effect of the neural network adopted in the present invention, the root mean square error (RMSE), mean relative error (MPE), mean absolute error (MAE) and mean absolute percentage error (MAPE) are used as evaluation indicators of the neural network model and compared with the prediction results of the traditional neural network. As shown in Table 1, the method for correcting the connection weights of the neural network based on dynamic progressive search proposed in the present invention has smaller error values ​​than those of the traditional neural network.

[0112] Table 1 Neural network prediction evaluation indicators

[0113]

[0114] In order to test whether the neural network model is over-trained, resulting in over-fitting of the predicted output, the data of three wells in the study area are retained and not involved in the network training, which is used to test the training effect of the neural network. The actual sandstone content data of the drilling is compared with the sandstone content data predicted by the neural network, such as Figure 5 and Figure 6 As shown, the blue curve is the real data of drilling sandstone content, and the brown curve is the neural network prediction data. The LD-E, HK-A and LG-A on the far right of the figure are verification wells that do not participate in training. It can be seen that although a small amount of drilling sandstone content data is different from the predicted value, the vast majority of the actual drilling sandstone content data is basically consistent with the neural network prediction results. In addition, the present invention adopts a neural network weight correction method, especially when the sandstone content value in the training sample set fluctuates greatly, it can maintain the correction when the output information mutates, effectively avoiding the situation where the prediction deviation is large when the sandstone content value mutates, such as Figure 5 As shown in Figure 2, the accuracy of the final sand body prediction result is guaranteed. However, when the training set contains data with large deviations, the traditional neural network algorithm is often prone to fall into the local optimum and it is difficult to ensure the true prediction of the data. Figure 6 shown.

[0115] The fifth step is to predict sand body parameters. The three seismic attributes sensitive to sand bodies, namely, root mean square, arc length and half energy, are selected as the input data of the neural network model. After the forward output of the neural network, the simulated data of sand body content consistent with the distribution range of seismic attribute data is obtained. After gridding into a map, the plane distribution prediction of sand bodies in the study area is finally realized. Figure 7 As shown. In the figure, red and yellow correspond to areas with high sandstone content, and blue corresponds to areas with low sandstone content. The sand bodies in the first section of the Huangliu Formation in the study area are relatively developed as a whole. On the east and west sides of the basin, the sand bodies are distributed in a fan-shaped distribution. In the Dongfang District and Ledong District, the sand bodies are mostly elliptical or nearly circular, and the sand body distribution is significantly different. The sand body content in many areas of the basin is more than 70%, the average sand body content in the Dongfang District is about 40%, and the sand body content in the Ledong District is relatively high, basically around 60%.

[0116] Combination Figure 7 , that is, the comprehensive logging diagram of the three wells, is analyzed. Among them, the LD-E well is located in the center of Ledong District, with relatively developed sandstone. The sandstone content of drilling statistics is about 60%, and the corresponding sandstone content of the LD-E well in the sandstone content prediction diagram is between 50% and 75%; the LD-I well is located in the eastern part of the diapir zone, with interbedded sand and mudstone, and the sandstone content is about 49%, and the corresponding sandstone content of the LD-I well in the sandstone content prediction diagram is between 40% and 60%; LT-F is located in the southern part of the eastern slope zone. The well has a complete set of sandstones in the Huangliu Formation, with a sandstone content of about 100%, and the corresponding sandstone content of the LT-F well in the sandstone content prediction diagram is between 90% and 100%. Figure 8 shown.

[0117] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0118] 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 seismic sand body prediction method based on a dynamic progressive search algorithm, characterized in that: The prediction method comprises: Acquiring data: Acquiring data on seismic attributes and sand body parameters in the seismic work area, wherein the seismic attribute data includes data on seismic attributes used for training and data on seismic attributes used for prediction; the seismic attributes include root mean square amplitude, arc length, instantaneous frequency, half energy, average amplitude and energy; Establishing a neural network model: taking the seismic attributes as input and the sand body parameters as output, and establishing an objective function; Training the neural network model: inputting the seismic attribute data and sand body parameter data used for training into the neural network model to train the neural network model; the training of the neural network model includes the following steps: Obtaining a locally optimal connection weight correction value: setting a threshold range, performing a first correction on the connection weight based on the objective function, and obtaining a first weight correction value; when the first weight correction value is within the threshold range, the first weight correction value is the locally optimal connection weight correction value; when the first weight correction value is not within the threshold range, performing a second iterative calculation on the first weight correction value to obtain a second weight correction value, and the second weight correction value is the locally optimal connection weight correction value; Obtaining a global optimal connection weight correction amount: correcting the weight correction amounts of adjacent nodes based on the local optimal connection weight correction amount to obtain a third weight correction amount, wherein the third weight correction amount is the global optimal connection weight correction amount; Obtaining a trained neural network model: obtaining an optimal connection weight by using the global optimal connection weight correction amount to obtain a trained neural network model; Predict sand body parameters: Based on the trained neural network model, the seismic attribute data used for prediction is input to predict the sand body parameter data.

2. The seismic sand body prediction method based on the dynamic progressive search algorithm according to claim 1 is characterized in that: In the process of obtaining the global optimal connection weight correction amount, a deviation range is set, and the weight correction amounts of adjacent nodes are corrected based on the local optimal connection weight correction amount. The correction is stopped when the calculated correction amount of the adjacent connection weight is zero; when the correction is stopped, if the deviation between the third weight correction amount and the first weight correction amount is within the deviation range, the third weight correction amount is the global optimal connection weight correction amount; when the correction is stopped, if the deviation between the third weight correction amount and the first weight correction amount is not within the deviation range, the third weight correction amount is iteratively calculated for the third time to obtain a fourth weight correction amount, and the fourth weight correction amount is the global optimal connection weight correction amount.

3. The seismic sand body prediction method based on the dynamic progressive search algorithm according to claim 2 is characterized in that: The neural network model includes an input layer, an intermediate layer and an output layer, and adjacent layers are connected by nodes; obtaining the local optimal connection weight correction amount and obtaining the global optimal connection weight correction amount are both performed by iteratively correcting the weight correction amounts of the input layer and the intermediate layer, and the connection weight of the output layer and the intermediate layer is calculated based on the optimal connection weight of the input layer and the intermediate layer.

4. The seismic sand body prediction method based on the dynamic progressive search algorithm according to claim 3 is characterized in that: The objective function is established according to the principle of minimizing the error between the forward expected output value of the neural network model and the known actual value, and specifically adopts the following formula: Among them, E represents the global error, Y represents the known actual value, and x i represents the earthquake attribute of the i-th node in the input layer, w ij represents the connection weight between the i-th node in the input layer and the j-th node in the middle layer, v j represents the connection weight between the output layer node and the jth node in the middle layer, α i Indicates the threshold value.

5. The seismic sand body prediction method based on dynamic progressive search algorithm according to claim 4 is characterized in that: In the process of obtaining the local optimal connection weight correction amount, The connection weights are first corrected based on the objective function, and the objective function is solved by the gradient descent method to obtain the first weight correction amount, specifically using the following formula: Where Δw ij represents the initial correction value corresponding to the connection weight between the i-th node in the input layer and the j-th node in the middle layer, w ij represents the connection weight between the i-th node in the input layer and the j-th node in the middle layer, x i represents the earthquake attribute of the i-th node in the input layer, α i represents the threshold, v j represents the connection weight between the output layer node and the jth node in the middle layer, η represents the learning rate, which ranges from [0,1]; In the second iterative calculation of the first weight correction value, a maximum number of iterations is set to obtain a second weight correction value, specifically using the following formula: in, It represents the value of the connection weight correction between the i-th node in the input layer and the j-th node in the middle layer at the t-th iteration. represents the value of the connection weight correction between the i-th node in the input layer and the j-th node in the middle layer at the t+1th iteration, i represents the i-th node in the input layer, β represents a random number ranging from [-1,1], T max Indicates the maximum number of iterations.

6. The seismic sand body prediction method based on dynamic progressive search algorithm according to claim 3 is characterized in that: In the process of obtaining the global optimal connection weight correction amount, The third weight correction amount is specifically expressed by the following formula: in, It means that after p iterations, the local optimal connection weight correction between the i-th node in the input layer and the j-th node in the middle layer is obtained, Δw i,j+1 represents the weight correction of the adjacent nodes in the middle layer, D represents a 1×n matrix, n corresponds to the total number of nodes in the middle layer, each element in the matrix is ​​randomly assigned 1 or -1, and D + =D T (DD T ) -1 ; The fourth weight correction amount is specifically expressed by the following formula: in, It means that after t iterations, the weight correction of the i-th node in the input layer and the j+1-th node in the middle layer is obtained. It means that after t+1 iterations, the weight correction of the i-th node in the input layer and the j+1-th node in the middle layer is obtained. It represents the global optimal weight correction of the i-th node in the input layer and the j-th node in the middle layer obtained after B iterations, and β represents a random number in the range of [-1,1].

7. The seismic sand body prediction method based on dynamic progressive search algorithm according to claim 4 is characterized in that: The connection weight between the output layer and the middle layer is calculated according to the optimal connection weight between the input layer and the middle layer, specifically using the following formula: Among them, w ij represents the connection weight between the i-th node in the input layer and the j-th node in the middle layer, x i represents the earthquake attribute of the i-th node in the input layer, α i represents the threshold, v j represents the connection weight between the output layer node and the j-th node in the middle layer, η represents the learning rate in the range [0,1], and Y represents the known actual value.

8. The seismic sand body prediction method based on a dynamic progressive search algorithm according to any one of claims 1 to 9, characterized in that: The collected data specifically includes: Extracting seismic attribute data of each seismic line in the three-dimensional seismic area and the two-dimensional seismic area corresponding to the target layer, respectively, wherein the data includes three columns of numbers, namely, X coordinate, Y coordinate and seismic attribute value; extracting sand body parameter data of the seismic area where the seismic attribute data for training is located; Performing standard deviation normalization algorithm processing on the extracted seismic attribute data; Merge the two-dimensional and three-dimensional seismic attribute data processed by the standard deviation normalization algorithm. For data points with the same coordinates, replace them according to the attribute average value to obtain fused seismic attribute data. performing correlation analysis on the fused seismic attribute data used for training and the data of sand body parameters in the fused seismic attribute data, and selecting seismic attributes with statistical significance after the correlation analysis as preferred seismic attributes; In the process of establishing the neural network model, the preferred earthquake attributes are used as input.

9. The seismic sand body prediction method based on dynamic progressive search algorithm according to claim 8, characterized in that: The extracted seismic attribute data are processed by a standard deviation normalization algorithm, specifically using the following formula: a * =(a-mean) / std Among them, a * represents the seismic attribute data after standardization, a represents the original seismic attribute data, mean represents the mean of the seismic attribute data, and std represents the standard deviation of the seismic attribute data.

10. The seismic sand body prediction method based on dynamic progressive search algorithm according to claim 9, characterized in that: The predicted sand body parameters are specifically: Inputting the fused seismic attribute data used for prediction in the fused seismic attribute data of the study area into the trained neural network model to obtain a simulation output of sand body parameters; The simulated output of sand body parameters is gridded into a map to obtain the plane distribution map of sand body parameters in the study area and complete the sand body distribution prediction.

Images