Multi-model fusion parallelization reservoir parameter prediction method, system, equipment and medium

Abstract

The invention discloses a multi-model fusion parallelization reservoir parameter prediction method, system, equipment and medium, and the prediction method comprises the following steps: standardizing logging curve data and actually measured reservoir parameter data, constructing a BiGRU-TCN-Transform model, determining an optimal parameter combination, and training to obtain a final model; inputting the training set data into the model for training until the model converges, and obtaining a trained prediction model; predicting the test set data, and outputting a final prediction result; secondly, an improved VAE-Attention model is constructed, the prediction result in the first step serves as input, and a final operation result of the model is output through model processing; the invention further comprises a multi-model fusion parallelization reservoir parameter prediction system, equipment and a medium. The precision is improved, and good stability, reliability and applicability are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas reservoir physical property parameter prediction, and in particular to a multi-model fusion parallel reservoir parameter prediction method, system, equipment and medium. Background Art

[0002] Reservoir evaluation is crucial in the development and management of oil and gas fields. Porosity determines the storage space of the reservoir and directly affects the estimation of oil and gas reserves; permeability evaluates the ability of fluids to pass through rocks, which is related to the recoverability and productivity of oil and gas. The measurement and analysis of these parameters help to fully understand the reservoir characteristics, formulate effective development strategies, optimize the production process, and ultimately improve the economic benefits and development efficiency of oil and gas fields.

[0003] Predicting reservoir parameters through logging curves is of great significance and has significant advantages in oil and gas exploration and development. Compared with drilling coring and experimental analysis, predicting reservoir parameters through logging curves is more economical and efficient, and can obtain a large amount of data in a short time, and reduce the need for expensive and time-consuming experiments. Initially, Mohaghegh et al. (Mohaghegh et al., 1996) used artificial neural networks (ANN) to predict reservoir physical parameters, proving the feasibility and accuracy of neural networks in reservoir parameter prediction, and opened up new ideas for obtaining reservoir parameters. With the development of deep learning, more and more scholars have explored the prediction of reservoir parameters.

[0004] The emergence of recurrent neural networks (RNNs) provides an effective tool for modeling sequence data, but its performance is not ideal when processing long sequence data. It can usually only capture short-term dependencies and cannot effectively remember earlier input information. The introduction of long short-term memory networks (LSTMs) solves this problem well. Long short-term memory networks (LSTMs) introduce a gating mechanism that can well preserve and transmit long-time step information. Gated recurrent units (GRUs) are a simplified version of LSTMs, which retain the ability of LSTMs to capture long-term dependencies while simplifying the network structure, reducing computational complexity, and increasing computational efficiency. However, although LSTMs and GRUs have improved their ability to capture long-term dependencies through gating mechanisms, they still have certain limitations for long-distance dependencies. As the length of the sequence increases, the gradient may gradually become smaller during the back-propagation process, resulting in negligible contributions from the gradients of early time steps, which affects the model's learning of long-term dependencies.

[0005] To solve these problems, Vaswani et al. proposed a Transformer model based on the self-attention mechanism in Vaswani et al., 2017. The self-attention mechanism allows the model to pay attention to the inputs of all other positions in the sequence when calculating the representation of each input, which enables the model to capture long-range dependencies without processing the sequence data step by step. In the prediction task, although the convolutional neural network (CNN) is a deep learning model specifically designed for processing data with grid structure, it also performs well in the task of predicting sequence data. The local receptive field, weight sharing and long-range dependency capture capabilities of CNN enable it to efficiently process and analyze sequence data. However, the standard CNN has no time order constraints during the convolution operation, which may cause future information to leak into the output of the current time step. For this reason, a temporal convolutional network (TCN) designed specifically for processing sequence data has emerged. Through designs such as causal convolution, dilated convolution and residual connection, TCN can effectively capture patterns and long-range dependencies in time series, maintain time order, and remain stable during training.

[0006] For the reservoir physical property parameter prediction task, the patent application with publication number CN115222019A invented a deep Transformer transfer learning reservoir parameter prediction method based on well logging data. This method has shown good results in reservoir prediction tasks, but this method only considers the extraction of global data and still has shortcomings in the extraction of local data. Summary of the invention

[0007] In view of the problems existing in the prior art, the purpose of the present invention is to provide a multi-model fusion parallel reservoir parameter prediction method, system, device and medium, and construct a parallel BiGRU-TCN-Transformer model. First, a bidirectional gated recurrent unit (BiGRU) branch is used to process the bidirectional information of the sequence to ensure the comprehensive extraction of temporal features. At the same time, the dilated convolution is used through the temporal neural network (TCN) branch to extract the local temporal pattern and expand the feature receptive field; then the output results of the BiGRU branch and the TCN branch are added after linear transformation to obtain multi-scale features, and finally the self-attention mechanism of the Transformer module is used to integrate the multi-scale features to enhance the global dependency capture capability. In the model training process, multi-process technology is used to realize multi-parameter parallel prediction, which improves the model operation efficiency. In addition, in order to improve the final prediction accuracy, the VAE model is improved, BiGRU is used as the encoder and decoder, and the Attention mechanism is introduced to dynamically calculate the attention weight, so as to better capture the data distribution characteristics in long-term changes and improve the prediction accuracy of reservoir parameters.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] The multi-model fusion parallel reservoir parameter prediction method includes the following steps:

[0010] Step 1: BiGRU-TCN-Transformer model construction and operation:

[0011] After the well logging curve data and measured reservoir parameter data are standardized, they are divided into training set and test set according to the established ratio. Then, the BiGRU-TCN-Transformer model is constructed, and the model hyperparameters are optimized by genetic algorithm to determine the optimal parameter combination. The optimal parameters are input into the model for training until the model is stable and the final model is obtained. The training set data is input into the model for training until the model converges and the trained prediction model is obtained. Finally, the model is used to predict the test set data and the final prediction results are output.

[0012] Step 2: Improved VAE-Attention model construction and operation:

[0013] An improved VAE-Attention model is constructed, and the measured reservoir parameter data is standardized. The training set and the test set are divided according to the ratio in step one. Gaussian noise is introduced into the training set data to enhance the robustness of the model. The model hyperparameters are then optimized using a genetic algorithm to determine the optimal parameter combination. The optimal parameters are input into the model for training until the model is stable to obtain the final model. Finally, the prediction results in step one are used as input, further processed by the final model, and the final operation results of the model are output.

[0014] The detailed steps in step 1 are as follows:

[0015] (1.1) Data standardization and division

[0016] The collected logging curve data and measured reservoir parameter data are standardized. In the data standardization method, Z-Score standardization is selected. The Z-Score standardization performs data standardization based on the mean and standard deviation of each feature. For each feature, the mean and standard deviation of all samples of the feature are calculated, and then the feature is standardized so that the mean of each feature is 0 and the standard deviation is 1, that is, the converted feature conforms to the standard normal distribution. The expression is:

[0017]

[0018] Among them, x is the original data, μ is the mean of the data, and σ is the standard deviation of the data;

[0019] The standardized data set is divided into training set and test set in a ratio of 9:1;

[0020] (1.2) BiGRU-TCN-Transformer model construction

[0021] The BiGRU-TCN-Transformer model is a multi-branch deep learning framework. The model includes a temporal convolutional network (TCN), a bidirectional GRU (BiGRU), and a Transformer module. It is used to process time series data and make predictions. The TCN and BiGRU are used to extract local and global features of the sequence respectively, and then these features are integrated. The Transformer block is used to capture the long dependencies in the sequence and finally output the prediction results.

[0022] The modeling process of the BiGRU-TCN-Transformer model is as follows:

[0023] If there is well logging data x, its shape is (B, T, D), where B represents the batch size, T is the processing time step, and D represents the number of well logging curves;

[0024] First, for the BiGRU branch, it is composed of bidirectional GRU units, each GRU unit contains an update gate and a reset gate, and the formula of the unidirectional GRU is as follows:

[0025] Update gate z t :

[0026] z t =σ(W z ·[h t-1 ,x t ])

[0027] Reset Gate t :

[0028] r t =σ(W r ·[h t-1 ,x t ])

[0029] Candidate hidden states

[0030]

[0031] Update hidden state h t :

[0032]

[0033] Among them, x t represents the well logging data of the tth time step of the input sequence, ht-1 represents the hidden state of the previous time step, W r , W z and W h Respectively represent the weight parameters of the reset gate, update gate and candidate hidden state, σ is the sigmoid function, ⊙ represents element-by-element multiplication; its bidirectional structure is:

[0034]

[0035]

[0036]

[0037] in, is represented as the forward hidden state, is the backward hidden state, and its output shape is (B,T,2H), where H is the hidden state dimension of GRU;

[0038] The TCN branch consists of a series of one-dimensional convolutional layers, and the calculation process of each layer is as follows:

[0039] First, causal convolution and dilated convolution are performed on the logging data, and the corresponding convolution formula is as follows:

[0040] Causal convolution formula:

[0041]

[0042] Dilated convolution formula:

[0043]

[0044] Among them, w i represents the convolution kernel parameter, k represents the size of the convolution kernel, x t represents the hidden state after BiGRU processing, d represents the expansion rate of the extended convolution, and y t is the output of the TCN layer;

[0045] TCN uses residual connections between each layer, and the corresponding formula is:

[0046] y t =x t +Conv1D(x t )

[0047] Among them, Conv1D represents a one-dimensional convolution operation, namely causal convolution; its output shape is (B, T, C), where C is the number of output channels of the last TCN module;

[0048] The output data of the BiGRU branch and the TCN branch are linearly transformed and added to convert them into feature representations that conform to the shape of the Transformer input structure. The conversion formula is:

[0049]

[0050] Among them, W TCN , W BiGRU is the weight matrix of the linear transformation between TCN and BiGRU, y t is the output result of the BiGRU branch and the TCN branch. The output shape of y is (B, T, D model ), D model is the hidden dimension of Transformer;

[0051] Adding positional encoding to the output after linear transformation enables the model to utilize the position information of the sequence:

[0052]

[0053]

[0054] Among them, pos represents the position index used to generate the position encoding, and the output shape after position encoding is (B, T, D model );

[0055] The position-encoded data is used as the input of the Transformer layer. Each layer of Transformer contains a multi-head attention mechanism and a feedforward neural network, and captures the global dependencies in the sequence through the self-attention mechanism and the feedforward neural network. The formula of the self-attention mechanism is:

[0056]

[0057] Where Q, K, and V represent the query, key, and value matrices used for the self-attention mechanism, respectively. Softmax is a normalization operation. k is the dimension of the key;

[0058] The formula of the multi-head attention mechanism is:

[0059] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O

[0060]

[0061] in, W O Represented as the weight matrix of the multi-head attention mechanism;

[0062] The formula of its feedforward neural network is:

[0063] FFN(x)=max(0,xW1+b1)W2+b2

[0064] Among them, W1, W2, b1, b2 represent the weight matrix and bias vector of the feedforward neural network

[0065] The complete process of the Transformer layer is:

[0066] output1=MultiHeadAttention(LayerNorm(z))+z

[0067] output2=FFN(LayerNorm(output1))+output1

[0068] Among them, z represents the input data after position encoding, and LayerNorm represents the normalization operation;

[0069] Finally, the output of the last layer of Transformer passes through the linear layer to get the final prediction result:

[0070] y out =Linear(output2);

[0071] (1.3) BiGRU-TCN-Transformer model hyperparameter tuning

[0072] Select the genetic algorithm for hyperparameter tuning, build the genetic algorithm framework, and tune the hyperparameters in the model. The tuning range is shown in Table 1:

[0073] Table 1 Neural network hyperparameter selection

[0074] Hyperparameters Experimental setup Number of TCN channels [16,32,64]-[80,160,320] batch 64-256 Learning Rate <![CDATA[10 -4 -10 -1 ]]> Transformer layer dimensions 64-512 Feedforward layer dimensions 64-512 Dropout rate of TCN layer 0-0.5 Transformer Dropout Rate 0-0.5 Transformer Layer 1-10 BiGRU Layer 1-10 Number of longs 1-10 Iterations 10-200

[0075] (1.4) Parallelization framework construction

[0076] A multi-process parallel framework based on the multiprocessing module is built. This framework realizes parallel processing by creating multiple independent processes. Each process has its own Python interpreter and independent memory space. In the reservoir parameter prediction task, a separate process is assigned to each parameter, thereby realizing parallel processing of the task.

[0077] The detailed steps of step 2 are as follows:

[0078] (2.1) Model construction

[0079] The VAE-Attention model includes an encoder and a decoder;

[0080] (2.1.1) Encoder:

[0081] The encoder (q(z|x)) maps the reservoir parameter input data to the latent space and outputs the mean μ(x) and variance σ of the latent variable z 2 (x), the encoder adopts a BiGRU structure, in which each GRU unit contains an update gate and a reset gate, and the corresponding mathematical expression is as follows:

[0082] Reset Gate t :

[0083] r t =σ(W r ·[h t-1 ,x t ])

[0084] Candidate hidden states

[0085]

[0086] Update gate z t :

[0087] z t =σ(W z ·[h t-1 ,x t ])

[0088] Update hidden state h t :

[0089]

[0090] Among them, x t represents the reservoir permeability data at the tth time step of the input sequence, h t-1 represents the hidden state of the previous time step, W r , W z and W represent the weight parameters of the reset gate, update gate, and candidate hidden state, respectively, σ is the sigmoid function, and ⊙ represents element-by-element multiplication;

[0091] After extracting the features of the permeability data using the BiGRU structure, it is distributed to the corresponding latent space through a linear layer, namely:

[0092] q(z|x)=N(z;μ(x),σ 2 (x)

[0093] In order to enable the gradient to be passed through random variables, the reparameterization technique is used to represent the encoded data of the encoder as a function of the mean and variance, namely:

[0094] z=μ(x)+σ(x)⊙∈

[0095] Among them, ∈~N(0,1), represents the noise of standard normal distribution, ⊙ represents element-by-element multiplication;

[0096] (2.1.2) Decoder

[0097] The decoder maps the low-dimensional latent representation back to the original data space, and the decoding layer is also replaced with a BiGRU structure, and the re-parameterized latent variable z is reconstructed back to the original data through the BiGRU and Attention mechanism. The Attention mechanism acts on the decoder process to enhance the accuracy and expressiveness of the model when generating data. In the decoder, the Attention mechanism helps the decoder to dynamically adjust the attention to the input sequence according to the latent variable z and the current output state, thereby generating a more accurate output;

[0098] First, map the reparameterized latent variable z to the initial hidden state h0:

[0099] h0=f c (z)

[0100] Then use GRU to generate data sequences:

[0101] o t ,h t =GRU(o {t-1} ,h {t-1} )

[0102] Among them, t ,h t Respectively represent the output and hidden state at t time step;

[0103] The current output of the decoder is o t and the hidden state h t Associated with the calculation of the attention weights, an attention layer is then used to combine the current decoder output ot and the previous hidden state h t-1 Connect it up:

[0104] score(o t ,h t-1 )=attention(o t ,h t-1 )

[0105] Then normalize the attention score through the Softmax function to get the attention weight a t:

[0106]

[0107] Then use the attention weight a t The output of the encoder is weighted and summed to obtain the context vector c t :

[0108]

[0109] Among them, i is the output sequence of the encoder, a t,i is the attention weight;

[0110] Finally, reconstruct the output using the context vector c t Combined with the decoder's current hidden state h t , through a fully connected layer f to generate the final reconstruction output

[0111]

[0112] It should be noted that the goal of the VAE model is to maximize the marginal likelihood p(x) of the data, that is:

[0113] log p(x)=log(x|z)p(z)dz

[0114] Using the variational inference method, we introduce an approximate posterior distribution q(z|x), and obtain the following lower bound through Jensen's inequality:

[0115]

[0116] Among them, D KL is the KL divergence, which measures the difference between two distributions;

[0117] The loss function of the improved VAE-Attention model consists of reconstruction loss and KL divergence loss:

[0118]

[0119] Among them, the reconstruction loss is to measure the reconstruction data The difference between the original data x, usually using the mean square error (MSE), KL divergence loss is a measure of the difference between the approximate posterior distribution of the encoder output and the prior distribution, which is usually a standard normal distribution;

[0120] (2.2) Model training

[0121] (2.2.1) Data standardization and division

[0122] Standardize the collected measured reservoir parameter data, select Z-Score standardization in the data standardization method, and divide the standardized data into training data and test data with a ratio of 9:1;

[0123] (2.2.2) Introducing noise

[0124] In order to increase the generalization ability of the model, the introduced noise is Gaussian noise, and its expression is:

[0125] x noisy =x+∈,

[0126] Among them, x is the reservoir permeability data, ∈ is subject to N(0,σ 2 ) random variables;

[0127] (2.2.3) Hyperparameter Tuning

[0128] Select the genetic algorithm for hyperparameter tuning, build the genetic algorithm framework, and tune the hyperparameters in the model. The tuning range is shown in Table 2.

[0129] Table 2 Neural network hyperparameter selection

[0130] Hyperparameters Experimental setup Hidden layer dimension 32-512 Latent space vector dimension 16-256 Learning Rate 0.001-0.01 Batch size 32-512 Noise variance 0.0-0.1 Iterations 50-200

[0131] (2.2.3) Model training and application

[0132] The optimal parameters are input into the model for training until the model is stable and the final model is obtained; finally, the prediction results in step one are used as input, further processed by the final model, and the final operation results of the model are output.

[0133] The multi-model fusion parallel reservoir parameter prediction system is applied in the process of oil and gas field reservoir evaluation to realize the multi-model fusion parallel reservoir parameter prediction method mentioned above, and includes a processing instruction module for executing each step in the method.

[0134] Multi-model fusion parallel reservoir parameter prediction equipment, including:

[0135] Memory: a computer program for storing the multi-model fusion parallel reservoir parameter prediction method, which is a computer-readable device;

[0136] Processor: used to implement the multi-model fusion parallel reservoir parameter prediction method when executing the computer program.

[0137] A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, can implement the multi-model fusion parallel reservoir parameter prediction method.

[0138] Compared with the prior art, the present invention has the following beneficial effects:

[0139] (1) Step 1 of the present invention combines the advantages of BiGRU, TCN and Transformer, and has a strong ability to process sequence data. BiGRU captures the bidirectional dependency of sequence data, TCN efficiently captures long-range dependency through extended convolution and large receptive field, and Transformer realizes global feature extraction through self-attention mechanism and position encoding, thereby achieving the effect of improving prediction accuracy.

[0140] (2) The present invention selects a genetic algorithm in the model hyperparameter tuning method and performs a comprehensive screening of the model's hyperparameters, thereby obtaining the model operation effect under the optimal parameters.

[0141] (3) The present invention also uses an improved VAE-Attention model to reconstruct the prediction results, thereby increasing the accuracy and robustness of the final prediction results.

[0142] In summary, the present invention has the advantages of efficient parallel computing when processing sequence data, and can fully understand and utilize the temporal and global characteristics of the data, thereby improving prediction performance and stability; the improved VAE-Attention model is used to reconstruct the prediction results of BiGRU-TCN-Transformer, further improving accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0143] Figure 1 This is the BiGRU-TCN-Transformer structure diagram.

[0144] Figure 2 This is the improved VAE-Attention structure diagram.

[0145] Figure 3 Apply a flowchart to the model.

[0146] Figure 4 It is the BiGRU-TCN-Transformer training loss diagram; where (a) is porosity and (b) is permeability.

[0147] Figure 5 This is the improved VAE-Attention training loss diagram; where (a) is porosity and (b) is permeability.

[0148] Figure 6 The final prediction effect diagram is shown in Figure 1, where (a) is the permeability and (b) is the porosity. DETAILED DESCRIPTION

[0149] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0150] The purpose of the present invention is to improve the prediction accuracy of reservoir parameters, and a parallelized BiGRU-TCN-Transformer model is constructed. The model first uses a bidirectional gated recurrent unit (BiGRU) branch to process the bidirectional information of the sequence to ensure the comprehensive extraction of temporal features. At the same time, the dilated convolution is used to extract local temporal patterns through the temporal neural network (TCN) branch to expand the feature receptive field; then the output results of the BiGRU branch and the TCN branch are added after linear transformation to obtain multi-scale features, and finally the self-attention mechanism of the Transformer module is used to integrate multi-scale features to enhance the global dependency capture capability. In the model training process, multi-process technology is used to realize multi-parameter parallel prediction, which improves the model operation efficiency. In addition, in order to improve the final prediction accuracy, the VAE model is improved, BiGRU is used as the encoder and decoder, and the Attention mechanism is introduced to dynamically calculate the attention weight, so as to better capture the data distribution characteristics in long-term changes.

[0151] Reference Figure 3 , a multi-model fusion parallel reservoir parameter prediction method includes the following steps:

[0152] Step 1: BiGRU-TCN-Transformer model construction and operation:

[0153] After the well logging curve data and measured reservoir parameter data are standardized, they are divided into training set and test set according to the established ratio. Subsequently, the BiGRU-TCN-Transformer model is constructed, and the model hyperparameters are optimized using genetic algorithm to determine the optimal parameter combination. The optimal parameters are input into the model for training until the model is stable and the final model is obtained. The training set data is input into the model for training until the model converges and the trained prediction model is obtained. Finally, the model is used to predict the test set data and the final prediction results are output.

[0154] Step 2: Improved VAE-Attention model construction and operation:

[0155] An improved VAE-Attention model is constructed, and the measured reservoir parameter data is standardized. The training set and the test set are also divided according to the proportion in step 1. Gaussian noise is introduced into the training set data to enhance the robustness of the model. Then the genetic algorithm is used to optimize the model hyperparameters and determine the optimal parameter combination. The optimal parameters are input into the model for training until the model is stable and the final model is obtained. Finally, the prediction results in step 1 are used as input, further processed by the final model, and the final operation results of the model are output.

[0156] The detailed steps in step 1 are as follows:

[0157] (1.1) Data standardization and division

[0158] The collected logging curve data and measured reservoir parameter data are standardized. In the data standardization method, Z-Score standardization is selected. The Z-Score standardization performs data standardization based on the mean and standard deviation of each feature. For each feature, the mean and standard deviation of all samples of the feature are calculated, and then the feature is standardized so that the mean of each feature is 0 and the standard deviation is 1, that is, the converted feature conforms to the standard normal distribution. The expression is:

[0159]

[0160] Among them, x is the original data, μ is the mean of the data, and σ is the standard deviation of the data.

[0161] The standardized data set is divided into training set and test set in a ratio of 9:1.

[0162] (1.2) BiGRU-TCN-Transformer model construction

[0163] Reference Figure 1 The BiGRU-TCN-Transformer model is a multi-branch deep learning framework. The BiGRU-TCN-Transformer model combines the temporal convolutional network (TCN), the bidirectional GRU (BiGRU), and the Transformer module to process time series data and make predictions. The core idea of ​​the model is to extract the local and global features of the sequence through TCN and BiGRU respectively, and then integrate these features, further capture the long dependencies in the sequence through the Transformer block, and finally output the prediction results.

[0164] The modeling process of the BiGRU-TCN-Transformer model is as follows:

[0165] If there is logging curve data x, its shape is (B, T, D), where B represents the batch size, T is the processing time step, and D represents the number of logging curves.

[0166] First, for the BiGRU branch, it is composed of bidirectional GRU units, each GRU unit contains an update gate and a reset gate, and the formula of the unidirectional GRU is as follows:

[0167] Update gate z t :

[0168] z t =σ(W z ·[h t-1,x t ])

[0169] Reset Gate t :

[0170] r t =σ(W r ·[h t-1 ,x t ])

[0171] Candidate hidden states

[0172]

[0173] Update hidden state h t :

[0174]

[0175] Among them, x t represents the well logging data of the tth time step of the input sequence, h t-1 represents the hidden state of the previous time step, W r , W z and W h They represent the weight parameters of the reset gate, update gate and candidate hidden state respectively, σ is the sigmoid function, and ⊙ represents element-by-element multiplication.

[0176] Its bidirectional structure is:

[0177]

[0178]

[0179]

[0180] in, is represented as the forward hidden state, is the backward hidden state, and its output shape is (B,T,2H), where H is the hidden state dimension of GRU.

[0181] The TCN branch consists of a series of one-dimensional convolutional layers, and the calculation process of each layer is as follows:

[0182] First, causal convolution and dilated convolution are performed on the logging data, and the corresponding convolution formula is as follows:

[0183] Causal convolution formula:

[0184]

[0185] Dilated convolution formula:

[0186]

[0187] Among them, w i represents the convolution kernel parameter, k represents the size of the convolution kernel, x t represents the hidden state after BiGRU processing, d represents the expansion rate of the extended convolution, and y t is the output of the TCN layer.

[0188] TCN uses residual connections between each layer, and the corresponding formula is:

[0189] y t =x t +Conv1D(x t )

[0190] Among them, Conv1D represents a one-dimensional convolution operation, namely causal convolution. Its output shape is (B, T, C), where C is the number of output channels of the last TCN module.

[0191] The output data of the BiGRU branch and the TCN branch are linearly transformed and added to convert them into feature representations that conform to the shape of the Transformer input structure. The conversion formula is:

[0192]

[0193] Among them, W TCN , W BiGRU is the weight matrix of the linear transformation between TCN and BiGRU, y t is the output result of the BiGRU branch and the TCN branch. The output shape of y is (B, T, D model ), D model is the hidden dimension of Transformer.

[0194] Adding positional encoding to the output after linear transformation enables the model to utilize the position information of the sequence:

[0195]

[0196]

[0197] Among them, pos represents the position index used to generate the position encoding, and the output shape after position encoding is (B, T, D model ).

[0198] The position-encoded data is used as the input of the Transformer layer. Each layer of Transformer contains a multi-head attention mechanism and a feedforward neural network, and captures the global dependencies in the sequence through the self-attention mechanism and the feedforward neural network.

[0199] The formula of the self-attention mechanism is:

[0200]

[0201] Among them, Q, K, and V represent the query, key, and value matrices used for the self-attention mechanism, respectively, softmax is a normalization operation, and d k The dimension of the key.

[0202] The formula of the multi-head attention mechanism is:

[0203] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O

[0204]

[0205] in, W O Represented as the weight matrix of the multi-head attention mechanism.

[0206] The formula of its feedforward neural network is:

[0207] FFN(x)=max(0,xW1+b1)W2+b2

[0208] Among them, W1, W2, b1, b2 represent the weight matrix and bias vector of the feedforward neural network

[0209] The complete process of the Transformer layer is:

[0210] output1=MultiHeadAttention(LaverNorm(z))+z

[0211] output2=FFN(LayerNorm(output)1))+output1

[0212] Where z represents the input data after position encoding, and LayerNorm represents the normalization operation.

[0213] Finally, the output of the last layer of Transformer passes through the linear layer to get the final prediction result:

[0214] y out =Linear(output2)

[0215] (1.3) BiGRU-TCN-Transformer model hyperparameter tuning

[0216] Select the genetic algorithm for hyperparameter tuning. Construct the genetic algorithm framework and tune the hyperparameters in the model. The tuning range is shown in Table 1. In the genetic algorithm framework, the population size is set to 20, the number of offspring is set to 5, the crossover probability is set to 0.7, and the mutation probability is set to 0.2. The optimizer used in the model training and parameter tuning process is the Adam optimizer. For different tasks, the optimal parameter tuning results are shown in Table 3.

[0217] Table 3 Optimal parameters of genetic algorithm

[0218] Hyperparameters POR PERM Number of TCN channels [48,96,192] [80,160,320] batch 310 376 Learning Rate 0.003 0.001 Transformer layer dimensions 464 394 Feedforward layer dimensions 471 391 Dropout rate of TCN layer 0.38 0.25 Transformer Dropout Rate 0.07 0.27 Transformer Layer 1 1 BiGRU Layer 2 1 Number of longs 10 2 Iterations 50 150

[0219] (1.4) Principles of parallel framework construction

[0220] In order to train the model as a whole, a multi-process parallel framework based on the multiprocessing module was built. This framework achieves parallel processing by creating multiple independent processes, each of which has its own Python interpreter and independent memory space. In the reservoir parameter prediction task, a separate process is assigned to each parameter, thereby achieving parallel processing of the task.

[0221] (1.5) BiGRU-TCN-Transformer model training and testing

[0222] The tuned hyperparameters are input into the model and all models are trained. The error changes during the training of each model refer to Figure 4 It can be seen that the model has finally converged. Finally, the test set data is predicted and the prediction results are saved.

[0223] The detailed steps of step 2 are as follows:

[0224] (2.1) Model construction

[0225] Reference Figure 2 ,VAE-Attention model consists of four parts: encoder, re-parameterization, decoder, and Attention mechanism, and each part is connected in series.

[0226] (2.1.1) Encoder:

[0227] The encoder (q(z|x)) part is responsible for mapping the input data (reservoir parameters) to the latent space and outputting the mean μ(x) and variance σ of the latent variable z. 2 (x). The encoder in this model adopts the BiGRU structure, and each GRU unit contains an update gate and a reset gate. The corresponding mathematical expression is as follows:

[0228] Reset Gate t :

[0229] r t =σ(W r ·[h t-1 ,x t ])

[0230] Candidate hidden states

[0231]

[0232] Update gate z t :

[0233] z t =σ(W z ·[h t-1 ,x t ])

[0234] Update hidden state h t :

[0235]

[0236] Among them, x t represents the reservoir permeability data at the tth time step of the input sequence, h t-1 represents the hidden state of the previous time step, W r , W z and W represent the weight parameters of the reset gate, update gate, and candidate hidden state, respectively, σ is the sigmoid function, and ⊙ represents element-wise multiplication.

[0237] After extracting the features of the permeability data using the BiGRU structure, it is distributed to the corresponding latent space through the linear layer. That is:

[0238] q(z|x)=N(z;μ(x),σ 2 (x)

[0239] In order to enable the gradient to be passed through random variables, the reparameterization technique is used to represent the encoded data of the encoder as a function of the mean and variance. That is:

[0240] z=μ(x)+σ(x)⊙∈

[0241] Among them, ∈~N(0,1) represents the noise of standard normal distribution. ⊙ represents element-by-element multiplication.

[0242] (2.1.2) Decoder

[0243] The decoder is responsible for mapping the low-dimensional latent representation back to the original data space, thereby achieving data reconstruction. This patent replaces the decoding layer with a BiGRU structure, and reconstructs the re-parameterized latent variable z back to the original data through the BiGRU and Attention mechanism. The Attention mechanism mainly acts on the decoder process to enhance the accuracy and expressiveness of the model when generating data. In the decoder, the Attention mechanism helps the decoder to dynamically adjust the attention to the input sequence based on the latent variable z and the current output state, thereby generating a more accurate output.

[0244] First, map the reparameterized latent variable z to the initial hidden state h0:

[0245] h0=f c (z)

[0246] Then use GRU to generate data sequences:

[0247] o t ,h t =GRU(o {t-1} ,h {t-1} )

[0248] Among them t ,h t They represent the output and hidden state at time step t respectively.

[0249] The current output of the decoder is o t and the hidden state h t Associated with the calculation of attention weights. Then use an attention layer to convert the current decoder output o t and the previous hidden state h t-1 Connect it up:

[0250] score(o t ,h t-1 )=attention(o t ,h t-1 )

[0251] Then normalize the attention score through the Softmax function to get the attention weight a t :

[0252]

[0253] Then use the attention weight a t The output of the encoder is weighted and summed to obtain the context vector c t :

[0254]

[0255] Among them, i is the output sequence of the encoder, a t,i is the attention weight.

[0256] Finally, reconstruct the output using the context vector c t Combined with the decoder's current hidden state h t , through a fully connected layer f to generate the final reconstruction output

[0257]

[0258] It should be noted that the goal of the VAE model is to maximize the marginal likelihood p(x) of the data, that is:

[0259] logp(x)=log(x|z)p(z)dz

[0260] Since it is difficult to directly calculate this integral, we use the variational inference method to introduce an approximate posterior distribution q(z|x). Through Jensen's inequality, we can get the following lower bound:

[0261]

[0262] Among them, D KL is the KL divergence, which measures the difference between two distributions.

[0263] The loss function of the improved VAE-Attention model consists of reconstruction loss and KL divergence loss:

[0264]

[0265] The reconstruction loss is used to measure the reconstruction data The difference between the original data x, usually using the mean square error (MSE). KL divergence loss measures the difference between the approximate posterior distribution of the encoder output and the prior distribution, which is usually a standard normal distribution.

[0266] (2.2) Model training

[0267] (2.2.1) Data standardization and division

[0268] The collected measured reservoir parameter data are standardized, and Z-Score standardization is selected in the data standardization method. The standardized data are divided into training data and test data with a ratio of 9:1.

[0269] (2.2.2) Introducing noise

[0270] The purpose of introducing noise is to better increase the generalization ability of the model. The noise introduced in the present invention is Gaussian noise.

[0271] Then its expression is:

[0272] x noisy =x+∈,

[0273] Among them, x is the reservoir permeability data, ∈ is subject to N(0,σ 2 ) is a random variable.

[0274] (2.2.3) Hyperparameter Tuning

[0275] Select the genetic algorithm for hyperparameter tuning. Construct the genetic algorithm framework and tune the hyperparameters in the model. The tuning range is shown in Table 2. In the genetic algorithm framework, the population size is set to 20, the number of offspring is set to 5, the crossover probability is set to 0.7, and the mutation probability is set to 0.2. The optimizer used in the model training and parameter tuning process is the Adam optimizer. The optimal parameters are shown in Table 4.

[0276] Table 4 Optimal parameters of genetic algorithm

[0277] Offspring POR PERM Hidden layer dimension 302 389 Latent space vector dimension 31 82 Learning Rate 0.008 0.002 Batch size 506 311 Noise variance 0.05 0.01 Iterations 140 180

[0278] The corresponding parameters are input into the improved VAE-Attention model for training to obtain the corresponding improved VAE-Attention model. Then, the prediction results in step 1 are put into the trained model to calculate the corresponding mean square error.

[0279] (2.2.4) Model training and application

[0280] The optimal parameters are input into the model for training until the model is stable and the final model is obtained. Figure 5 As shown, it can be seen that the model has finally converged. Finally, the prediction results in step 1 are used as input, further processed by the final model, and the final running results of the model are output.

[0281] Effect comparison of the model of the present invention:

[0282] The running results of the combination of the BiGRU-TCN-Transformer model and the improved VAE-Attention model are compared and analyzed with the BiGRU, BiLSTM, CNN, TCN, and Transformer models. The detailed steps are as follows:

[0283] Reference Figure 6, draw a comparison image of the prediction effect of BiGRU, BiLSTM, CNN, TCN and Transformer models and the combination of BiGRU-TCN-Transformer model and improved VAE-Attention model, and observe Figure 6 It can be found that the prediction effect of the combination of the BiGRU-TCN-Transformer model and the improved VAE-Attention model is obviously closer to the true value, and shows a better fitting effect than the existing models.

[0284] Analysis of reservoir parameter prediction effect of the present invention:

[0285] Calculate the mean square error between the running results of the combination of the BiGRU-TCN-Transformer model and the improved VAE-Attention model and the true value, and analyze the improvement effect of the running results of the combination of the BiGRU-TCN-Transformer model and the improved VAE-Attention model compared with the BiGRU, BiLSTM, CNN, TCN and Transformer models. Finally, in order to conduct a comprehensive analysis of all models, calculate the R 2 The fitting effect of each model and the improvement effect of the BiGRU-TCN-Transformer model combined with the improved VAE-Attention model compared with other models are analyzed. The steps are as follows:

[0286] To prove the improvement effect of the experiment, the mean square errors of the BiGRU, BiLSTM, CNN, TCN, Transformer models, the BiGRU-TCN-Transformer model, and the improved VAE-Attention model are calculated and quantitatively compared.

[0287] The formula for calculating the mean square error is:

[0288]

[0289] Where n is the number of samples, y i is the true value, is the predicted value of the model. The smaller the mean square error value, the closer it is to the true value. The final calculation results are shown in Table 5.

[0290] Table 5 Mean square error of each model

[0291] Model POR PERM BiGRU 0.7309 1.2778 BiLSTM 0.7326 1.3479 CNN 0.7570 1.8600 TCN 0.7084 0.6793 Transformer 0.6929 0.6159 BiGRU-TCN-Transformer 0.4705 0.2750 Improved VAE-Attention Reconstruction 0.3981 0.2396

[0292] In the porosity prediction task, the effects of the prediction models are slightly different, but the TCN model, Transformer model and BiGRU-TCN-Transformer model show higher prediction accuracy. Among them, the BiGRU-TCN-Transformer model has the highest prediction accuracy, with a mean square error of 0.4705, while the mean square errors of other models are all around 0.7. In the permeability prediction task, the prediction accuracy of the BiGRU-TCN-Transformer model is higher than that of other models, with a mean square error of 0.2750, while the CNN model has the worst prediction effect. The TCN model and Transformer model also show good results in permeability prediction. Finally, the improvement rate of the BiGRU-TCN-Transformer model compared with other models is calculated, and the improvement rate of the BiGRU-TCN-Transformer model after reconstruction using the improved VAE-Attention model is calculated.

[0293] The formula for calculating the improvement rate is:

[0294]

[0295] In the porosity prediction task, the mean square error of the BiGRU-TCN-Transformer model is improved by 35.63%, 35.77%, 37.86%, 33.61% and 32.08% compared with the BiGRU, BiLSTM, CNN, TCN and Transformer models, respectively. Among them, the error after reconstruction of the improved VAE-Attention model is 15.38% higher than that of the BiGRU-TCN-Transformer model. In the permeability prediction task, the mean square error of the BiGRU-TCN-Transformer model is improved by 78.48%, 79.60%, 85.21%, 59.52% and 55.36% compared with the BiGRU, BiLSTM, CNN, TCN and Transformer models, respectively. Among them, the error after reconstruction of the improved VAE-Attention model is 12.86% higher than that of the BiGRU-TCN-Transformer model.

[0296] Finally, in order to analyze the final fitting effect of the model, the R2 value of each model in different prediction tasks was calculated.

[0297] The expression of R2 is:

[0298]

[0299] Among them, SS resis the residual sum of squares, which represents the sum of squares of the differences between the model predicted values ​​and the actual values; SS tot is the total sum of squares, which represents the sum of squares of the difference between the dependent variable and its mean. The larger the R2 value, the better the fitting effect. The final calculation results are shown in Table 6.

[0300] Table 6 R of each model 2 value

[0301] Model POR(%) PERM(%) BiGRU 95.66 74.94 BiLSTM 95.80 81.84 CNN 95.81 82.79 TCN 95.94 90.85 Transformer 96.03 91.70 BiGRU-TCN-Transformer 97.30 96.30 Improved VAE-Attention Reconstruction 97.72 96.77

[0302] From Table 6, we can see that in the porosity prediction task, each model showed a good fitting effect, and the difference between the models was not obvious. Among them, the BiGRU-TCN-Transformer model performed best. 2 The value has also been improved, proving that the improved VAE model can effectively improve the accuracy of the model's final prediction results.

[0303] In the permeability prediction task, the performance of each model is different from that of the porosity prediction, which shows that the permeability has a stronger nonlinear relationship. However, the BiGRU-TCN-Transformer model still shows excellent prediction results, and the reconstructed prediction value is also improved, which further verifies the effectiveness of the model.

[0304] The above describes the specific embodiments of the present invention in detail. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work.

[0305] The multi-model fusion parallel reservoir parameter prediction system is applied in the process of oil and gas field reservoir evaluation to realize the multi-model fusion parallel reservoir parameter prediction method mentioned above, and includes a processing instruction module for executing each step in the method.

[0306] Multi-model fusion parallel reservoir parameter prediction equipment, including:

[0307] Memory: a computer program for storing the multi-model fusion parallel reservoir parameter prediction method, which is a computer-readable device;

[0308] Processor: used to implement the multi-model fusion parallel reservoir parameter prediction method when executing the computer program.

[0309] A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, can implement the multi-model fusion parallel reservoir parameter prediction method.

Claims

1. A multi-model fusion parallel reservoir parameter prediction method, characterized in that: The following steps are involved: Step 1: BiGRU-TCN-Transformer model construction and operation: After the well logging curve data and measured reservoir parameter data are standardized, they are divided into training set and test set according to the established ratio. Then, the BiGRU-TCN-Transformer model is constructed, and the model hyperparameters are optimized by genetic algorithm to determine the optimal parameter combination. The optimal parameters are input into the model for training until the model is stable and the final model is obtained. The training set data is input into the model for training until the model converges and the trained prediction model is obtained. Finally, the model is used to predict the test set data and the final prediction results are output. Step 2: Improved VAE-Attention model construction and operation: An improved VAE-Attention model is constructed, and the measured reservoir parameter data is standardized. The training set and the test set are divided according to the ratio in step one. Gaussian noise is introduced into the training set data to enhance the robustness of the model. The model hyperparameters are then optimized using a genetic algorithm to determine the optimal parameter combination. The optimal parameters are input into the model for training until the model is stable to obtain the final model. Finally, the prediction results in step one are used as input, further processed by the final model, and the final operation results of the model are output.

2. The multi-model fusion parallel reservoir parameter prediction method according to claim 1, characterized in that: The detailed steps in step 1 are as follows: (1.1) Data standardization and division The collected logging curve data and measured reservoir parameter data are standardized. In the data standardization method, Z-Score standardization is selected. The Z-Score standardization performs data standardization based on the mean and standard deviation of each feature. For each feature, the mean and standard deviation of all samples of the feature are calculated, and then the feature is standardized so that the mean of each feature is 0 and the standard deviation is 1, that is, the converted feature conforms to the standard normal distribution. The expression is: Among them, x is the original data, μ is the mean of the data, and σ is the standard deviation of the data; The standardized data set is divided into training set and test set in a ratio of 9:1; (1.2) BiGRU-TCN-Transformer model construction The BiGRU-TCN-Transformer model is a multi-branch deep learning framework. The model includes a temporal convolutional network (TCN), a bidirectional GRU (BiGRU), and a Transformer module. It is used to process time series data and make predictions. The TCN and BiGRU are used to extract local and global features of the sequence respectively, and then these features are integrated. The Transformer block is used to capture the long dependencies in the sequence and finally output the prediction results. The modeling process of the BiGRU-TCN-Transformer model is as follows: If there is well logging data x, its shape is (B, T, D), where B represents the batch size, T is the processing time step, and D represents the number of well logging curves; First, for the BiGRU branch, it is composed of bidirectional GRU units, each GRU unit contains an update gate and a reset gate, and the formula of the unidirectional GRU is as follows: Update gate z t : z t =σ(W z ·[h t-1 ,x t ]) Reset Gate t : r t =σ(W r ·[h t-1 ,x t ]) Candidate hidden states Update hidden state h t : Among them, x t represents the well logging data of the tth time step of the input sequence, h t-1 represents the hidden state of the previous time step, W r , W z and W h Respectively represent the weight parameters of the reset gate, update gate and candidate hidden state, σ is the sigmoid function, ⊙ represents element-by-element multiplication; its bidirectional structure is: in, is represented as the forward hidden state, is the backward hidden state, and its output shape is (B, T, 2H), where H is the hidden state dimension of GRU; The TCN branch consists of a series of one-dimensional convolutional layers. The calculation process of each layer is as follows: First, causal convolution and dilated convolution are performed on the logging data, and the corresponding convolution formula is as follows: Causal convolution formula: Dilated convolution formula: Among them, w i represents the convolution kernel parameter, k represents the size of the convolution kernel, x t represents the hidden state after BiGRU processing, d represents the expansion rate of the dilated convolution, and yt is the output of the TCN layer; TCN uses residual connections between each layer, and the corresponding formula is: y t =x t +Conv1D(x t ) Among them, Conv1D represents a one-dimensional convolution operation, namely causal convolution; its output shape is (B, T, C), where C is the number of output channels of the last TCN module; The output data of the BiGRU branch and the TCN branch are linearly transformed and added to convert them into feature representations that conform to the shape of the Transformer input structure. The conversion formula is: Among them, W TCN , W BiGRU is the weight matrix of the linear transformation between TCN and BiGRU, y t is the output result of the BiGRU branch and the TCN branch. The output shape of y is (B, T, D model ), D model is the hidden dimension of Transformer; Adding positional encoding to the output after linear transformation enables the model to utilize the position information of the sequence: Among them, pos represents the position index used to generate the position encoding, and the output shape after the position encoding is (B, T, D model ); The position-encoded data is used as the input of the Transformer layer. Each layer of Transformer contains a multi-head attention mechanism and a feedforward neural network, and captures the global dependencies in the sequence through the self-attention mechanism and the feedforward neural network. The formula of the self-attention mechanism is: Where Q, K, and V represent the query, key, and value matrices used for the self-attention mechanism, respectively. Softmax is a normalization operation. k is the dimension of the key; The formula of the multi-head attention mechanism is: MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O in, W O Represented as the weight matrix of the multi-head attention mechanism; The formula of its feedforward neural network is: FFN(x)=max(0,xW1+b1)W2+b2 Among them, W1, W2, b1, b2 represent the weight matrix and bias vector of the feedforward neural network The complete process of the Transformer layer is: output1=MultiHeadAttention(LayerNorm(z))+z output2=FFN(LayerNorm(output1))+output1 Where z represents the input data after position encoding, and LayerNorm represents the normalization operation; Finally, the output of the last layer of Transformer passes through the linear layer to get the final prediction result: you t =Linear(output2); (1.3) BiGRU-TCN-Transformer model hyperparameter tuning Select the genetic algorithm for hyperparameter tuning, build the genetic algorithm framework, and tune the hyperparameters in the model. The tuning range is shown in Table 1: Table 1 Neural network hyperparameter selection (1.4) Principles of parallel framework construction A multi-process parallel framework based on the multiprocessing module was built. This framework achieves parallel processing by creating multiple independent processes. Each process has its own Python interpreter and independent memory space. In the reservoir parameter prediction task, a separate process is assigned to each parameter, thereby achieving parallel processing of the task.

3. The multi-model fusion parallel reservoir parameter prediction method according to claim 1, characterized in that: The detailed steps of step 2 are as follows: (2.1) Model construction The VAE-Attention model includes an encoder and a decoder; (2.1.1) Encoder: The encoder (q(z|x)) maps the reservoir parameter input data to the latent space and outputs the mean μ(x) and variance σ of the latent variable z 2 (x), the encoder adopts a BiGRU structure, where each GRU unit contains an update gate and a reset gate, and the corresponding mathematical expression is as follows: Reset Gate t : r t =σ(W r ·[h t-1 ,x t ]) Candidate hidden states Update gate z t : z t =σ(W z ·[h t-1 ,x t ]) Update hidden state h t : Among them, x t represents the reservoir permeability data at the tth time step of the input sequence, h t-1 represents the hidden state of the previous time step, W r , W z and W represent the weight parameters of the reset gate, update gate, and candidate hidden state, respectively, σ is the sigmoid function, and ⊙ represents element-by-element multiplication; After extracting the features of the permeability data using the BiGRU structure, it is distributed to the corresponding latent space through a linear layer, namely: q(z|x)=N(z;μ(x),σ 2 (x)) In order to enable the gradient to be passed through random variables, the reparameterization technique is used to represent the encoded data of the encoder as a function of the mean and variance, namely: z=μ(x)+σ(x)⊙∈ Among them, ∈~N(0,1), represents the noise of standard normal distribution, ⊙ represents element-by-element multiplication; (2.1.2) Decoder The decoder maps the low-dimensional latent representation back to the original data space, and the decoding layer is also replaced with a BiGRU structure, and the re-parameterized latent variable z is reconstructed back to the original data through the BiGRU and Attention mechanism. The Attention mechanism acts on the decoder process to enhance the accuracy and expressiveness of the model when generating data. In the decoder, the Attention mechanism helps the decoder to dynamically adjust the attention to the input sequence according to the latent variable z and the current output state, thereby generating a more accurate output; First, map the reparameterized latent variable z to the initial hidden state h0: h0=f c (With) Then use GRU to generate data sequences: the t ,h t =GRU(o {t-1} ,h {t-1} ) Among them, t ,h t Represent the output and hidden state at t time step respectively; The current output of the decoder is o t and the hidden state h t Associated with the calculation of the attention weight, an attention layer is then used to convert the current decoder output o t and the previous hidden state h t-1 Connect it up: score(o t ,h t-1 )=attention(o t ,h t-1 ) Then normalize the attention score through the Softmax function to get the attention weight a t : Then use the attention weight a t The output of the encoder is weighted and summed to obtain the context vector c t : Among them, i is the output sequence of the encoder, a t，i is the attention weight; Finally, reconstruct the output using the context vector c t Combined with the decoder's current hidden state h t , through a fully connected layer f to generate the final reconstruction output It should be noted that the goal of the VAE model is to maximize the marginal likelihood p(x) of the data, that is: log p(x)=log(x|z)p(z)dz Using the variational inference method, we introduce an approximate posterior distribution q(z|x), and obtain the following lower bound through Jensen's inequality: Among them, D KL is the KL divergence, which measures the difference between two distributions; The loss function of the improved VAE-Attention model consists of reconstruction loss and KL divergence loss: Among them, the reconstruction loss is to measure the reconstruction data The difference between the original data x, usually using the mean square error (MSE), KL divergence loss is a measure of the difference between the approximate posterior distribution of the encoder output and the prior distribution, which is usually a standard normal distribution; (2.2) Model training (2.2.1) Data standardization and division Standardize the collected measured reservoir parameter data, select Z-Score standardization in the data standardization method, and divide the standardized data into training data and test data with a ratio of 9:1; (2.2.2) Introducing noise In order to increase the generalization ability of the model, the introduced noise is Gaussian noise, and its expression is: x noisy =x+∈, Among them, x is the reservoir permeability data, ∈ is subject to N(0, σ 2 ) random variables; (2.2.3) Hyperparameter Tuning Select the genetic algorithm for hyperparameter tuning, build the genetic algorithm framework, and tune the hyperparameters in the model. The tuning range is shown in Table 2. Table 2 Neural network hyperparameter selection (2.2.3) Model training and application The optimal parameters are input into the model for training until the model is stable and the final model is obtained; finally, the prediction results in step one are used as input, further processed by the final model, and the final operation results of the model are output.

4. Multi-model fusion parallel reservoir parameter prediction system, characterized by: The method for predicting reservoir parameters by multi-model fusion and parallelization as described in any of claims 1 to 3 is implemented in the process of applying oil and gas field reservoir evaluation, and includes a processing instruction module for executing each step in the method.

5. Multi-model fusion parallel reservoir parameter prediction equipment, characterized in that: include: Memory: a computer program storing the multi-model fusion parallel reservoir parameter prediction method according to any one of claims 1 to 3, which is a computer-readable device; Processor: used to implement the multi-model fusion parallel reservoir parameter prediction method described in any of claims 1-3 when executing the computer program.

6. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can implement the multi-model fusion parallel reservoir parameter prediction method described in any of claims 1-3.

Images