Porosity prediction method based on time-frequency characteristics

Abstract

The invention discloses a porosity prediction method based on time-frequency characteristics, and belongs to the field of geological exploration. The method comprises the following steps: firstly, preprocessing each kind of logging data to generate a logging data curve; extracting spatial domain features and frequency domain features of the logging data curve and fusing the spatial domain features and the frequency domain features to obtain a fused feature vector; a three-dimensional matrix is constructed according to the fusion feature vectors of the multiple kinds of logging data, and then the three-dimensional matrix is expanded to obtain a two-dimensional matrix; and finally, respectively inputting the three-dimensional matrix and the two-dimensional matrix into two input ends of a porosity prediction model so as to obtain the porosity. According to the method, by fusing the time-frequency features and combining a specific model input mode, the porosity prediction precision is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of geological exploration, and in particular to a porosity prediction method based on time-frequency characteristics. Background Art

[0002] Traditional methods of measuring porosity are usually based on single data features or simple data processing models, which makes it difficult to fully capture the complex information contained in logging data.

[0003] With the expanding application of deep learning in geophysics, researchers are exploring more advanced porosity measurement methods. Existing research proposes a spatiotemporal network design embedded with an attention mechanism for porosity prediction. This method first extracts spatial features from well logging data (such as spontaneous potential and acoustic transit time) using a CNN. A bidirectional LSTM (BiLSTM) then integrates temporal dependencies between surrounding rocks above and below the formation. Causal convolution is then used to generate the query and key values ​​for the attention mechanism, incorporating local information into the attention weights. Finally, a Monte Carlo dropout technique is introduced to generate 95% confidence intervals (86.93% coverage) using 1000 random samplings to quantify prediction uncertainty.

[0004] However, the existing technology only obtains the spatial characteristics of logging data, and does not fully mine the features in the logging data, resulting in low porosity prediction accuracy. Summary of the Invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a porosity prediction method based on time-frequency characteristics to solve the problem of low porosity prediction accuracy in the prior art.

[0006] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a porosity prediction method based on time-frequency characteristics, comprising the following steps: S1. Preprocess each type of well logging data, generate a data curve based on each type of well logging data after preprocessing, and obtain a well logging data curve; S2, extracting spatial domain features and frequency domain features from the logging data curve, and fusing them to obtain a fused feature vector; S3, converting the fused feature vectors of the various logging data into an initial three-dimensional matrix, shearing and flattening the initial three-dimensional matrix to obtain a two-dimensional matrix; S4. Input the initial three-dimensional matrix into the first input terminal of the porosity prediction model, and input the two-dimensional matrix into the second input terminal of the porosity prediction model to obtain porosity.

[0007] Furthermore, the logging data in S1 include: permeability PERM, total organic carbon TOC, kerogen content VKER, gas content QALL, gamma GR, water saturation SW, resistivity RTHU and uranium U.

[0008] Furthermore, the preprocessing in S1 includes: missing value filling, outlier removal and standardization.

[0009] Furthermore, S2 includes the following sub-steps: S21, sliding a sliding window on the well logging data curve, and extracting the spatial characteristics of the data under the sliding window at each sliding; S22, extracting frequency domain features from the logging data curve using continuous wavelet transform; S23, performing linear transformation on the spatial domain features and the frequency domain features respectively to obtain dimensionally adjusted spatial domain features and dimensionally adjusted frequency domain features; S24, multiplying the dimensionally adjusted spatial domain features and the dimensionally adjusted frequency domain features element-wise to obtain an initial fusion vector; S25, reconstructing the initial fusion vector into an initial fusion matrix; S26. Sum the elements of each row of each initial fusion matrix to obtain a fusion feature vector.

[0010] Furthermore, the spatial domain features in S21 include: mean, standard deviation, range, kurtosis and skewness; the frequency domain features in S22 include: frequency, phase and amplitude.

[0011] Furthermore, S3 includes the following sub-steps: S31, cutting the fused feature vectors of the multiple logging data to obtain multiple fused feature vectors with the same length; S32, splicing each fused feature vector after cutting into a feature fusion matrix; S33, converting the feature fusion matrix into an initial three-dimensional matrix according to the time step, wherein the three dimensions of the initial three-dimensional matrix include: number of samples, time step, and number of features; S34. Crop the information of the last multiple time steps from the initial three-dimensional matrix in the time dimension to obtain a cropped three-dimensional matrix, wherein the three dimensions of the cropped three-dimensional matrix include: the number of samples, N, and the number of features, where N is a positive integer greater than or equal to 2; S35. Flatten the two dimensions of the cropped three-dimensional matrix to obtain a two-dimensional matrix. The two dimensions of the two-dimensional matrix include: the number of samples and N×the number of features.

[0012] Furthermore, the porosity prediction model in S4 includes: a neural network sub-model, an autoregressive sub-model, and a fully connected layer; The input of the neural network sub-model serves as the first input of the porosity prediction model; The input of the autoregressive sub-model serves as the second input of the porosity prediction model; The input end of the fully connected layer is connected to the output end of the neural network sub-model and the output end of the autoregressive sub-model respectively, and its output end serves as the output end of the porosity prediction model.

[0013] Furthermore, the neural network sub-model includes: the first convolutional layer, residual unit, recurrent layer, recurrent skip layer, Concat layer, multi-head attention unit, linear transformation unit and Sigmoid unit; The input end of the first convolutional layer serves as the input end of the neural network sub-model, and its output end is connected to the input end of the residual unit; the output end of the residual unit is connected to the input end of the recurrent layer and the input end of the recurrent skip layer respectively; the input end of the Concat layer is connected to the output end of the recurrent layer and the output end of the recurrent skip layer respectively, and its output end is connected to the input end of the multi-head attention unit; the input end of the linear transformation unit is connected to the output end of the multi-head attention unit, and its output end is connected to the input end of the Sigmoid unit; the output end of the Sigmoid unit serves as the output end of the neural network sub-model.

[0014] Furthermore, the residual unit includes: a second convolutional layer, a BN layer, a ReLU layer and an adder; The input of the second convolutional layer is connected to the first input of the adder and serves as the input of the residual unit; the input of the BN layer is connected to the output of the second convolutional layer, and its output is connected to the input of the ReLU layer; the second input of the adder is connected to the output of the ReLU layer, and its output serves as the output of the residual unit.

[0015] Furthermore, the first convolutional layer is configured with 48 convolution kernels, the kernel size is 6, and the stride is 1.

[0016] The beneficial effects of the present invention are: In the feature extraction link, the present invention extracts spatial and frequency domain features from the logging data curve and fuses them, changing the limitation of the existing technology of only obtaining spatial features and fully mining the complex information of multiple dimensions in the frequency and spatial domains of the logging data; the initial three-dimensional matrix and the two-dimensional matrix are respectively input into different input ends of the porosity prediction model, and the input data of the two aspects are processed respectively. The feature mapping of the two aspects is integrated to improve the accuracy of porosity prediction, effectively solving the problem of low porosity prediction accuracy in the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a porosity prediction method based on time-frequency characteristics; Figure 2 Schematic diagram of the porosity prediction model; Figure 3 It is a structural diagram of the neural network sub-model; Figure 4Schematic diagram of the structure of the residual unit; Figure 5 This is a comparison chart of the evaluation of the porosity prediction model; Figure 6 This is a comparison chart of the structural effects of the porosity prediction model; Figure 7 This is a comparison chart of the predicted value and the true value of the model of the present invention. DETAILED DESCRIPTION

[0018] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0019] like Figure 1 As shown, a porosity prediction method based on time-frequency characteristics includes the following steps: S1. Preprocess each type of well logging data, generate a data curve based on each type of well logging data after preprocessing, and obtain a well logging data curve; S2, extracting spatial domain features and frequency domain features from the logging data curve, and fusing them to obtain a fused feature vector; S3, converting the fused feature vectors of the various logging data into an initial three-dimensional matrix, shearing and flattening the initial three-dimensional matrix to obtain a two-dimensional matrix; S4. Input the initial three-dimensional matrix into the first input terminal of the porosity prediction model, and input the two-dimensional matrix into the second input terminal of the porosity prediction model to obtain porosity.

[0020] In this embodiment, the logging data in S1 include: permeability PERM, total organic carbon TOC, kerogen content VKER, gas content QALL, gamma GR, water saturation SW, resistivity RTHU and uranium U.

[0021] In this embodiment, the preprocessing in S1 includes: missing value filling, outlier removal and standardization.

[0022] In this embodiment, 8 types of well logging data are collected, and each type of well logging data constitutes a data set.

[0023] In this embodiment, missing value filling is performed as follows: samples at the upper and lower edge values ​​of the data set lack practical significance for porosity prediction, so deletion is adopted for processing. After deletion, missing values ​​exist at these positions. For missing values ​​in the middle of the data set, considering that the logging data presents a sequential relationship with the change of well depth, the average value of the two adjacent values ​​is selected for filling.

[0024] In this embodiment, missing value processing can also adopt the following methods: KNN interpolation method: filling in the mean of K nearest neighbors based on similar samples; EM algorithm: iterative estimation of missing values ​​through expectation maximization (applicable to non-random missing data); multiple interpolation (MICE): generating multiple complete data sets and aggregating the results (improving statistical reliability).

[0025] Outlier Removal: Box plots provide an intuitive way to visualize data distribution characteristics by displaying the median, quartiles (Q1, Q3), and extreme values ​​(minimum and maximum). Any data point below the lower bound (i.e., less than Q1 - 1.5 × IQR) is considered an outlier. Any data point above the upper bound (i.e., greater than Q3 + 1.5 × IQR) is also considered an outlier.

[0026] In this embodiment, outlier processing may also adopt the following methods: Isolation Forest: efficiently identifying outliers in high-dimensional data; Local Outlier Factor (LOF): detecting local anomalies based on density (adaptive to non-uniformly distributed data).

[0027] The results of the parameters obtained after using the box plot test are as follows: Most of the outliers are distributed around the upper and lower boundaries, and their number is relatively large. According to the geological characteristics of shale gas, these changes in data are likely to be a direct reflection of stratum changes, so they should be regarded as normal geological changes rather than anomalies. Data far beyond the boundaries are considered outliers. Outliers can be treated as missing values. Outliers at the beginning and end of the data set will be deleted to ensure data consistency and accuracy, while outliers in the middle are replaced by filling methods to maintain data integrity while correcting these potential error points. This treatment takes into account the characteristics of shale gas geology and ensures the rigor and accuracy of data analysis.

[0028] Data normalization: Robust normalization is used to scale the data by subtracting the median and dividing by the interquartile range, which is more suitable for datasets with outliers.

[0029] In this embodiment, feature standardization can also adopt the following methods: Box-Cox transformation: first perform power transformation to normalize the data, and then standardize (to solve the skewed distribution problem); quantile normalization: force the data to obey a specific distribution (such as standard normal distribution).

[0030] In this embodiment, S2 includes the following sub-steps: S21, sliding a sliding window on the well logging data curve, and extracting the spatial characteristics of the data under the sliding window at each sliding; S22, extracting frequency domain features from the logging data curve using continuous wavelet transform; S23, performing linear transformation on the spatial domain features and the frequency domain features respectively to obtain dimensionally adjusted spatial domain features and dimensionally adjusted frequency domain features; S24, multiplying the dimensionally adjusted spatial domain features and the dimensionally adjusted frequency domain features element-wise to obtain an initial fusion vector; S25, reconstructing the initial fusion vector into an initial fusion matrix; S26. Sum the elements of each row of each initial fusion matrix to obtain a fusion feature vector.

[0031] In this embodiment, the spatial domain features in S21 include: mean, standard deviation, range, kurtosis, and skewness. The mean reflects the average trend of the curve; the standard deviation reflects the degree of fluctuation of the curve at a certain depth; the range reflects the maximum amplitude of the curve at a certain depth; the kurtosis reflects the steepness of the data distribution of the curve at a certain depth; and the skewness reflects the asymmetry of the data distribution of the curve at a certain depth. To preserve micro-fluctuation information, the sliding window size is 8, and the actual depth corresponding to the sliding window is 1 meter.

[0032] In this embodiment, S22 specifically uses a continuous wavelet transform to extract frequency domain features. The wavelet transform can adjust frequency analysis based on the signal's different time scales, allowing for better capture of the signal's local characteristics. It is more suitable for non-stationary signals, while the Fourier transform is more suitable for stationary signals. Therefore, the wavelet transform is chosen to extract frequency domain features. The frequency domain features in S22 include frequency, phase, and amplitude.

[0033] S23 is specifically: the formula for linear transformation of spatial domain features and frequency domain features is: ,in, is the spatial feature of dimension adjustment, is the frequency domain feature with dimension adjustment, U is the weight matrix of the linear transformation of the spatial domain feature x, V is the weight matrix of the linear transformation of the frequency domain feature y, m is the dimension of the spatial domain feature x, n is the dimension of the frequency domain feature y, dk is the shared dimension after transformation, and R is a real number set; S24 is specifically: , where z is the initial fusion vector.

[0034] S25 is specifically as follows: reconstructing the initial fusion vector z of dimension dk into a d×k matrix, i.e., the initial fusion matrix; S26 is specifically as follows: sum each sub-block in the initial fusion matrix to obtain the fusion feature vector: ,in, is the i-th element in the fusion feature vector, z i,jis the element in the i-th row and j-th column of the reconstructed d×k matrix, where i and j are positive integers, d is the number of rows of the reconstructed d×k matrix, and k is the number of columns of the reconstructed d×k matrix.

[0035] In this embodiment, S3 includes the following sub-steps: S31, cutting the fused feature vectors of the multiple logging data to obtain multiple fused feature vectors with the same length; S32, splicing each fused feature vector after cutting into a feature fusion matrix; S33, converting the feature fusion matrix into an initial three-dimensional matrix according to the time step, wherein the three dimensions of the initial three-dimensional matrix include: number of samples, time step, and number of features; S34. Crop the information of the last multiple time steps from the initial three-dimensional matrix in the time dimension to obtain a cropped three-dimensional matrix, wherein the three dimensions of the cropped three-dimensional matrix include: the number of samples, N, and the number of features, where N is a positive integer greater than or equal to 2; S35. Flatten the two dimensions of the cropped three-dimensional matrix to obtain a two-dimensional matrix. The two dimensions of the two-dimensional matrix include: the number of samples and N×the number of features.

[0036] N is the number of time steps. In this embodiment, the value of N is 3.

[0037] like Figure 2 As shown, the porosity prediction model in S4 includes: a neural network sub-model, an autoregressive sub-model, and a fully connected layer; The input of the neural network sub-model serves as the first input of the porosity prediction model; The input of the autoregressive sub-model serves as the second input of the porosity prediction model; The input end of the fully connected layer is connected to the output end of the neural network sub-model and the output end of the autoregressive sub-model respectively, and its output end serves as the output end of the porosity prediction model.

[0038] In this embodiment, the autoregressive sub-model is implemented using a fully connected layer. This fully connected layer multiplies the flattened input tensor with the weight matrix and adds a bias term to obtain a new tensor as output. This process is equivalent to a linear transformation of the input, thereby realizing the AR autoregressive process.

[0039] like Figure 3 As shown in the figure, the neural network sub-model includes: the first convolutional layer, residual unit, recurrent layer, recurrent skip layer, Concat layer, multi-head attention unit, linear transformation unit and Sigmoid unit; The input end of the first convolutional layer serves as the input end of the neural network sub-model, and its output end is connected to the input end of the residual unit; the output end of the residual unit is connected to the input end of the recurrent layer and the input end of the recurrent skip layer respectively; the input end of the Concat layer is connected to the output end of the recurrent layer and the output end of the recurrent skip layer respectively, and its output end is connected to the input end of the multi-head attention unit; the input end of the linear transformation unit is connected to the output end of the multi-head attention unit, and its output end is connected to the input end of the Sigmoid unit; the output end of the Sigmoid unit serves as the output end of the neural network sub-model.

[0040] The residual unit solves the gradient vanishing problem that may occur as the network depth increases through the residual connection mechanism, ensuring that the model can effectively learn the deep features of the data. Its output serves as the input of the recurrent layer and the recurrent skip layer.

[0041] The recurrent layer and recurrent skip layer operate in parallel to capture sequential dependencies in well logging data. The recurrent layer handles short-term dependencies, while the recurrent skip layer, through a skip connection mechanism, directly connects distant data points, thereby enhancing the model's ability to capture long-term dependencies. The processed feature information is fed into the attention unit. The application of a multi-head attention mechanism enables the model to simultaneously focus on key information from different locations, ensuring full utilization of effective information.

[0042] Finally, the autoregressive sub-model uses the AR model for regression analysis, capturing temporal dependencies and sequential patterns in the historical porosity data. The output of the autoregressive sub-model is integrated with the output of the neural network through a fully connected layer, which weights and sums information from different sources to generate the final forecast.

[0043] In this embodiment, the residual unit can be replaced by a dilated convolution or a separable convolution. The advantage of the dilated convolution is that it can expand the receptive field to capture long-term dependencies, and the advantage of the separable convolution is that it can reduce the number of parameters and improve computational efficiency. The GRU recurrent layer can be replaced by a temporal convolutional network. The advantage of the temporal convolutional network is that it can process long sequences in parallel and avoid gradients. The multi-head attention mechanism can be replaced by sparse attention or linear attention. The advantage of sparse attention is that it can reduce computational complexity and adapt to ultra-long sequences. The advantage of linear attention is that it can reduce complexity from O(N²) to O(N). The autoregressive sub-model can adopt a state-space model or a Bayesian structure time series. The advantage of the state-space model is that it can be combined with Kalman filtering to process noisy data. The advantage of the Bayesian structure time series is that it can introduce uncertainty quantification.

[0044] like Figure 4 As shown in the figure, the residual unit includes: the second convolutional layer, the BN layer, the ReLU layer and the adder; The input of the second convolutional layer is connected to the first input of the adder and serves as the input of the residual unit; the input of the BN layer is connected to the output of the second convolutional layer, and its output is connected to the input of the ReLU layer; the second input of the adder is connected to the output of the ReLU layer, and its output serves as the output of the residual unit.

[0045] The first convolutional layer is configured with 48 convolution kernels, with a convolution kernel size of 6 and a stride of 1. It performs a one-dimensional convolution operation on the initial three-dimensional matrix, and then performs residual processing. It first performs a one-dimensional convolution operation, then performs batch normalization on the convolution result, and finally activates the normalized result using the ReLU function to enhance the nonlinear processing capability; the output after residual unit processing is input into the recurrent layer and the recurrent skip layer respectively. The recurrent layer further learns the features of the input by using the gated recurrent unit GRU. The number of units is 64. The recurrent skip layer selectively inputs each recurrent neural network unit by adjusting the input The multi-head attention mechanism converts the tensor into three input matrices through three different mapping operations, performs self-attention calculation on each matrix, and then integrates each result and activates it through linear change and Sigmoid function to obtain the output tensor of the nonlinear part; finally, the output tensor of the nonlinear part is integrated with the output tensor of the linear part, and passes through the fully connected layer to obtain the final prediction result.

[0046] Model training: Initial three-dimensional and two-dimensional matrices are used as training samples. These training samples are then fed into the porosity prediction model. After convolution, a convolutional layer augmented with residual units is used to deeply extract the complex features of the well logging data. Second, a multi-head attention mechanism is used to optimize the recurrent and recurrent skip layers to accurately capture long-term and short-term sequence dependencies. Finally, an autoregressive submodel is introduced to address the neural network's sensitivity to scale changes. This method not only overcomes the limitations of traditional linear prediction models but also does not rely on specific physical models or statistical assumptions, demonstrating strong robustness and excellent scalability in porosity prediction.

[0047] Evaluation of porosity prediction model: Root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R 2 ) Three indicators are used to evaluate the trained porosity prediction model, such as Figure 5 As shown in the figure, the present invention compares the evaluation results of existing models MLP, CNN, RestNet, LSTM, BiGRU and the model of the present invention, and the present invention is significantly better than the existing models.

[0048] This paper verifies the effectiveness of residual units and multi-head attention units in the porosity prediction model. Figure 6 shown.

[0049] In this embodiment, when training the porosity prediction model, the loss functions that can be used include: MSE loss function, quantile loss function and Huber loss function.

[0050] like Figure 7 As shown in the figure, the predicted value of the model of the present invention is close to the true value.

[0051] When training the porosity prediction model, the optimizers used include: Lion optimizer, Adam optimizer and AdaHessian optimizer.

[0052] In the feature extraction link, the present invention extracts spatial and frequency domain features from the logging data curve and fuses them, changing the limitation of the existing technology of only obtaining spatial features and fully mining the complex information of multiple dimensions in the frequency and spatial domains of the logging data; the initial three-dimensional matrix and the two-dimensional matrix are respectively input into different input ends of the porosity prediction model, and the input data of the two aspects are processed respectively. The feature mapping of the two aspects is integrated to improve the accuracy of porosity prediction, effectively solving the problem of low porosity prediction accuracy in the existing technology.

[0053] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A porosity prediction method based on time-frequency characteristics, characterized in that: The following steps are involved: S1. Preprocess each type of well logging data, generate a data curve based on each type of well logging data after preprocessing, and obtain a well logging data curve; S2, extracting spatial domain features and frequency domain features from the logging data curve, and fusing them to obtain a fused feature vector; S3, converting the fused feature vectors of the various logging data into an initial three-dimensional matrix, shearing and flattening the initial three-dimensional matrix to obtain a two-dimensional matrix; S4. Input the initial three-dimensional matrix into the first input terminal of the porosity prediction model, and input the two-dimensional matrix into the second input terminal of the porosity prediction model to obtain porosity.

2. The porosity prediction method based on time-frequency characteristics according to claim 1 is characterized in that: The logging data in S1 include: permeability PERM, total organic carbon TOC, kerogen content VKER, gas content QALL, gamma GR, water saturation SW, resistivity RTHU and uranium U.

3. The porosity prediction method based on time-frequency characteristics according to claim 1, characterized in that: The preprocessing in S1 includes: missing value filling, outlier removal and standardization.

4. The porosity prediction method based on time-frequency characteristics according to claim 1, characterized in that: The S2 comprises the following sub-steps: S21, sliding a sliding window on the logging data curve, and extracting the spatial characteristics of the data under the sliding window at each sliding; S22, extracting frequency domain features from the logging data curve using continuous wavelet transform; S23, performing linear transformation on the spatial domain features and the frequency domain features respectively to obtain dimensionally adjusted spatial domain features and dimensionally adjusted frequency domain features; S24, multiplying the dimensionally adjusted spatial domain features and the dimensionally adjusted frequency domain features element-wise to obtain an initial fusion vector; S25, reconstructing the initial fusion vector into an initial fusion matrix; S26. Sum the elements of each row of each initial fusion matrix to obtain a fusion feature vector.

5. The porosity prediction method based on time-frequency characteristics according to claim 4 is characterized in that: The spatial domain features in S21 include: mean, standard deviation, range, kurtosis and skewness; The frequency domain features in S22 include: frequency, phase and amplitude.

6. The porosity prediction method based on time-frequency characteristics according to claim 1, characterized in that: The S3 includes the following sub-steps: S31, cutting the fused feature vectors of the multiple well logging data to obtain multiple fused feature vectors with the same length; S32, splicing each fused feature vector after cutting into a feature fusion matrix; S33, converting the feature fusion matrix into an initial three-dimensional matrix according to the time step, wherein the three dimensions of the initial three-dimensional matrix include: number of samples, time step, and number of features; S34. Crop the information of the last multiple time steps from the initial three-dimensional matrix in the time dimension to obtain a cropped three-dimensional matrix, wherein the three dimensions of the cropped three-dimensional matrix include: the number of samples, N, and the number of features, where N is a positive integer greater than or equal to 2; S35. Flatten the two dimensions of the cropped three-dimensional matrix to obtain a two-dimensional matrix. The two dimensions of the two-dimensional matrix include: the number of samples and N×the number of features.

7. The porosity prediction method based on time-frequency characteristics according to claim 1, characterized in that: The porosity prediction model in S4 includes: a neural network sub-model, an autoregressive sub-model and a fully connected layer; The input end of the neural network sub-model serves as the first input end of the porosity prediction model; The input end of the autoregressive sub-model serves as the second input end of the porosity prediction model; The input end of the fully connected layer is connected to the output end of the neural network sub-model and the output end of the autoregressive sub-model respectively, and the output end thereof serves as the output end of the porosity prediction model.

8. The porosity prediction method based on time-frequency characteristics according to claim 7, characterized in that: The neural network sub-model includes: a first convolutional layer, a residual unit, a recurrent layer, a recurrent skip layer, a Concat layer, a multi-head attention unit, a linear transformation unit, and a Sigmoid unit; The input end of the first convolutional layer serves as the input end of the neural network sub-model, and its output end is connected to the input end of the residual unit; the output end of the residual unit is respectively connected to the input end of the recurrent layer and the input end of the recurrent skip layer; the input end of the Concat layer is respectively connected to the output end of the recurrent layer and the output end of the recurrent skip layer, and its output end is connected to the input end of the multi-head attention unit; the input end of the linear transformation unit is connected to the output end of the multi-head attention unit, and its output end is connected to the input end of the Sigmoid unit; the output end of the Sigmoid unit serves as the output end of the neural network sub-model.

9. The porosity prediction method based on time-frequency characteristics according to claim 8, characterized in that: The residual unit includes: a second convolutional layer, a BN layer, a ReLU layer and an adder; The input end of the second convolutional layer is connected to the first input end of the adder and serves as the input end of the residual unit; the input end of the BN layer is connected to the output end of the second convolutional layer, and its output end is connected to the input end of the ReLU layer; the second input end of the adder is connected to the output end of the ReLU layer, and its output end serves as the output end of the residual unit.

10. The porosity prediction method based on time-frequency characteristics according to claim 8, characterized in that: The first convolutional layer is configured with 48 convolution kernels, the convolution kernel size is 6, and the stride is 1.

Images