A Well-Interconnectivity Analysis Method and Device Based on LAE

Abstract

The present invention discloses a method and device for analyzing inter-well connectivity based on LAE. For the historical injection and production data of an oilfield well group, LSTM-AE is used to encode the water injection volume sequence, construct a vector representation of the water injection volume sequence, and input this vector into a fully connected layer to predict the liquid production volume of production wells. The water injection volumes of each water well in the well group are used as input variables for sensitivity analysis, and the global sensitivity analysis method EFAST is applied to analyze the sensitivity of the input variables of the liquid production volume prediction model based on LSTM-AE. By decomposing the variance of the model prediction results, the influence of the water injection volumes of each water well and the coupling effect between the water injection volumes of water wells on the model results is obtained to calculate the parameter sensitivity index, identify the key water injection wells that have a significant impact on the liquid production volume of production wells, and clarify the inter-well connectivity status. The present invention can accurately and effectively analyze the inter-well connectivity.

Description

Technical Field

[0001] The present invention relates to a method and device for analyzing inter-well connectivity, in particular to a method and device for analyzing inter-well connectivity based on LAE. Background Art

[0002] In China, the reservoir heterogeneity is strong, and the contradictions between layers and within layers are serious. After the previous water flooding development with strong injection and production, the oil production in China is currently in the high water cut stage. Problems such as ineffective circulation of injected water, small swept area, and low water flooding efficiency are becoming increasingly prominent. The study of inter-well connectivity is of great significance for the determination and adjustment of oilfield development plans and the improvement of oil recovery.

[0003] Currently, the study of inter-well connectivity generally starts from two aspects: static connectivity and dynamic connectivity. The static connectivity research methods mainly include stratigraphic correlation methods using seismic, logging, geochemical and other data. However, static data cannot well reflect the dynamic changes of the reservoir and cannot be used as the sole basis for judging connectivity. Dynamic connectivity analysis refers to the analysis of the connectivity of fluids between wells. The traditional dynamic analysis methods mainly include: geochemical methods, inter-well tracer technology analysis methods, well testing analysis methods, and reservoir numerical simulation methods. Chemical, well testing and inter-well tracer analysis methods can accurately reflect the inter-well connectivity, but they are complex in operation, time-consuming and laborious, and often have an impact on the normal production of the oilfield, so they are not suitable for large-scale application in the oilfield. Reservoir numerical simulation methods are generally combined with tracer methods, which can quantitatively evaluate the reservoir connectivity and its dynamic changes, but establishing a geological model requires a large amount of material and human resources, and the calculation process is very complex. Due to the deficiencies of traditional methods, the production dynamic data inversion method for analyzing inter-well connectivity using a large amount of injection-production data has been widely used. This method has the advantages of convenient data collection and rich available data. The production dynamic data inversion method mainly includes methods such as the Spearman rank correlation method, multiple linear regression method, and capacitance model analysis method.

[0004] The Spearman rank correlation method is used to establish the injection-production model for single-well single-production. By calculating the rank correlation of the injection-production sequence, the inter-well connectivity is analyzed, and the correlation degree of the injection-production fluid volume sequence is used to characterize the inter-well connectivity. However, this method can only analyze the relationship between a single oil well and a single water well. In actual production, the influence on the oil well is often the superposition of multiple water wells, and the rank correlation degree often shows negative results, which cannot accurately characterize the inter-well connectivity status. The linear regression method assumes that the production of the oil well is linearly related to the injection volume of the surrounding water wells, and the regression coefficient is used as a measure of connectivity. The larger the regression coefficient, the stronger the connectivity between the well pairs. However, the multiple linear regression analysis method needs to meet conditions such as high diffusivity, constant permeability, and non-water flooding. The commissioning of new wells and the shut-in of old wells have a great impact on the results. The capacitance model analysis method is a widely used method for estimating inter-well connectivity, which can infer inter-well connectivity based on historical injection-production data and pressure data in water injection wells. However, the establishment and derivation process of the capacitance model are relatively complex, and the solution of model parameters takes a long time and is difficult to solve. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and device for analyzing inter-well connectivity based on LAE, which can provide reference for the determination and adjustment of oilfield development plans and improve oil recovery.

[0006] To solve the above technical problem, the technical solution adopted by the present invention is: a method for analyzing inter-well connectivity based on LAE, including the following steps:

[0007] A. Perform data preprocessing operations on the historical injection-production data of the well group, including data cleaning, dimensionless normalization, time series division, and training set and test set division.

[0008] B. Based on the preprocessed data, use the LSTM encoder to extract features from the input injection volume sequence, and the LSTM decoder reconstructs the injection volume sequence according to the feature mapping; use the time-based backpropagation algorithm to train the LSTM-AE model with minimizing the reconstruction error as the training objective, and continuously optimize the model parameters to achieve the purpose of feature extraction, and the training is completed.

[0009] C. After the training is completed, only keep the LSTM encoder to encode the input injection volume sequence, extract features, input the fully connected layer, and train the weight parameters of the fully connected layer using the error between the true value and the predicted value of the liquid production volume to realize the prediction of the liquid production volume.

[0010] D. Take the water injection volumes of each water well in the well group as the input variables of the sensitivity analysis, apply the global sensitivity analysis method EFAST to analyze the sensitivity of the input variables of the liquid production prediction model, obtain the coupling effect between each water well injection volume and the water well injection volumes through the decomposition of the model variance to calculate the parameter sensitivity index, identify the key water injection wells that have a significant impact on the liquid production of the production well, and clarify the well - to - well connectivity status.

[0011] In step A, the data cleaning operation includes removing null values, duplicate values, and outliers from the dataset. The method used for dimensionless normalization is normalization, and the formula is:

[0012]

[0013] where \(x\) represents the parameter to be normalized, \(x_{min}\) min represents the minimum value of the feature, and \(x_{max}\) max represents the maximum value of the feature; the data is segmented using the sliding window method. The time window length is set to \(l\) and the step size is 1, dividing the dataset into \(N\) time series; 70% of the dataset is divided into the training set, and the remaining 30% is used as the test set.

[0014] In step B, LSTM - AE refers to the LSTM auto - encoder. The working process is as follows. The LSTM encoder is used to extract the feature vector from the input water injection volume sequence \(X\) t =\(\{x_1\) t , \(x_2\) t+1 , …, \(x_n\) t+l-1 \}. The water injection volume \(x_t\) t at the \(t\) - th moment is concatenated with the hidden state \(h_{t - 1}\) of the LSTM neuron at the \((t - 1)\) - th moment and input into the LSTM neuron corresponding to the \(t\) - th moment. After encoding, the hidden state \(h_t\) of the LSTM is obtained. \(c\) is the number of neurons in the LSTM hidden layer. The initial state of the first LSTM neuron is \(c_0\) t -1 =\(h_0\) t-1 = 0; the hidden state \(h_{t + l - 1}\) of the \((t + l - 1)\) - th LSTM neuron in the encoder is used as the initial hidden state of the decoder. Through the linear layer operation of the LSTM neuron corresponding to the \((t + l - 1)\) - th moment in the decoder, the reconstructed water injection volume vector \(x'_{t + l - 1}\) t+l-1 is obtained; \(x'_{t + l - 1}\) t+l-1 and \(h_{t + l - 2}\) are used as the input of the LSTM neuron at the \((t + l - 2)\) - th moment, and the hidden state \(h_{t + l - 2}\) of the LSTM neuron corresponding to the \((t + l - 2)\) - th moment in the decoder is obtained and the reconstructed vector \(x'_{t + l - 2}\) t+l-2 is output. The decoding process continues until the reconstruction is completed. The reconstructed water injection volume sequence by the decoder is \(X'\)t ′ = {x′ t+l-1 , x′ t+l-2 , …, x′ t}。

[0015] In step D, the EFAST method is the Extended Fourier Amplitude Sensitivity Test method, which is one of the methods for global sensitivity analysis. The algorithm believes that the variance of the model is caused by the input parameters and the interactions between the parameters. By performing variance decomposition, the influence degree of each variable and the coupling effect between variables on the model is obtained to calculate the parameter sensitivity index.

[0016] An inter-well connectivity analysis device based on LAE includes the following modules:

[0017] Data preprocessing module: used to perform preprocessing operations such as data cleaning, dimensionless transformation, time series division, and training set and test set division on the historical injection and production data of the well group;

[0018] Liquid production prediction module based on LSTM-AE: uses the LSTM encoder to extract features from the input water injection volume sequence, and the LSTM decoder reconstructs the water injection volume sequence according to the feature mapping; uses the time-based backpropagation algorithm to train the LSTM-AE model with minimizing the reconstruction error as the training objective, and continuously optimizes the model parameters to achieve the purpose of feature extraction; after training, only retain the LSTM encoder part to encode the input water injection volume sequence, extract features, input to the fully connected layer, and use the error between the true value and the predicted value of the liquid production volume to train the weight parameters of the fully connected layer to achieve the prediction of the liquid production volume;

[0019] Inter-well connectivity analysis module: used to take the water injection volume of each water well in the well group as the input variable of sensitivity analysis, apply the global sensitivity analysis method EFAST to analyze the sensitivity of the input variables of the prediction model, obtain the water injection volume of each water well and the coupling effect between the water injection volumes of the water wells through the decomposition of the model variance to calculate the parameter sensitivity index, identify the key injection wells that have a significant impact on the liquid production volume of the production well, and clarify the inter-well connectivity status.

[0020] The data preprocessed by the data preprocessing module is input into the liquid production prediction module based on LSTM-AE, and the water injection volume of each injection well obtained by the liquid production prediction module based on LSTM-AE is used as the input variable of the inter-well connectivity analysis module, and the inter-well connectivity status is obtained after processing.

[0021] The beneficial effects of the present invention are as follows: For the historical injection-production data of the oilfield, LSTM-AE is applied to encode the water injection volume sequence, construct a vector representation of the water injection volume sequence, input the feature vector into the fully connected layer, and predict the liquid production volume of the production well. The water injection volumes of each water well in the well group are used as the input variables for sensitivity analysis. The global sensitivity analysis method EFAST is applied to analyze the sensitivity of the input variables of the prediction model. The coupling effect between each water well injection volume and the water well injection volumes is obtained through the decomposition of the model variance to calculate the parameter sensitivity index, identify the key water wells that have a significant impact on the liquid production volume of the production well, clarify the well-to-well connectivity status, and can provide a reference for the determination and adjustment of the oilfield development plan and improve the recovery factor. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flowchart of the well-to-well connectivity analysis method based on LAE of the present invention.

[0023] Figure 2 It is a schematic structural diagram of the well-to-well connectivity analysis device based on LAE of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0025] As Figure 1 、 2 shown, the well-to-well connectivity analysis method based on LAE of the present invention includes the following steps:

[0026] A. Perform data preprocessing operations on the historical injection-production data of the well group, including data cleaning, dimensionless normalization, time series partitioning, and training set and test set partitioning.

[0027] B. Based on the preprocessed data, use an LSTM encoder to extract features from the input water injection volume sequence, and an LSTM decoder to reconstruct the water injection volume sequence according to the feature mapping; use the time-based backpropagation algorithm to train the LSTM-AE model with minimizing the reconstruction error as the training objective, and continuously optimize the model parameters to achieve the purpose of feature extraction. After the training is completed.

[0028] C. After the training is completed, only retain the LSTM encoder to encode the input water injection volume sequence, extract features, input them into the fully connected layer, and train the weight parameters of the fully connected layer using the error between the true value and the predicted value of the liquid production volume to achieve the prediction of the liquid production volume.

[0029] D. Take the water injection volumes of each water well in the well group as the input variables for sensitivity analysis. Apply the global sensitivity analysis method EFAST to analyze the sensitivity of the input variables of the liquid production prediction model. Through the decomposition of the model variance, obtain the coupling effects between the water injection volumes of each water well and between the water injection volumes of the water wells to calculate the parameter sensitivity index, identify the key water wells that have a significant impact on the liquid production of the production well, and clarify the inter-well connectivity status.

[0030] In step A, the data cleaning operation includes removing null values, duplicate values, and outliers from the dataset. The method used for dimensionless normalization is normalization, and the formula is:

[0031]

[0032] where x represents the parameter to be normalized, x min represents the minimum value of the feature, and x max represents the maximum value of the feature; use the sliding window method to segment the data. Set the time window length to l and the step size to 1, and divide the dataset into N time series; divide 70% of the dataset into the training set, and the remaining 30% as the test set.

[0033] In step B, LSTM-AE refers to the LSTM autoencoder. The working process is as follows. The LSTM encoder is used to extract the feature vector from the input water injection volume sequence X t ={x t , x t+1 , …, xt +l-1}. Concatenate the water injection volume x t at the t-th moment with the hidden state of the LSTM neuron at the (t - 1)-th moment, and input it into the LSTM neuron corresponding to the t-th moment. After encoding, obtain the hidden state of the LSTM. C is the number of neurons in the LSTM hidden layer. The initial state of the first LSTM neuron is c t -1 =h t-1 =0; the hidden state of the (t + l - 1)-th LSTM neuron in the encoder is used as the initial hidden state of the decoder. Through the linear layer operation of the LSTM neuron corresponding to the (t + l - 1)-th moment in the decoder, obtain the reconstructed water injection volume vector x' t+l-1 at this moment; take x' t+l-1 and as the input of the LSTM neuron at the (t + l - 2)-th moment, obtain the hidden state of the LSTM neuron corresponding to the (t + l - 2)-th moment in the decoder and output the reconstructed vector x' t+l-2 , and continue the decoding process until the reconstruction is completed. The water injection volume sequence reconstructed by the decoder is Xt = {x' t+l-1 , x' t+l-2 , …, x' t}。

[0034] In step D, the EFAST method is the Extended Fourier Amplitude Sensitivity Test method, which is one of the methods for global sensitivity analysis. The algorithm believes that the variance of the model is caused by the input parameters and the interactions between the parameters. By performing variance decomposition, the influence degree of each variable and the coupling effect between variables on the model can be obtained to calculate the parameter sensitivity index.

[0035] An inter-well connectivity analysis device based on LAE includes the following modules:

[0036] Data preprocessing module: used to perform preprocessing operations such as data cleaning, dimensionless transformation, time series partitioning, and training set and test set partitioning on the historical injection and production data of the well group;

[0037] Liquid production prediction module based on LSTM-AE: The LSTM encoder is used to extract features from the input water injection volume sequence, and the LSTM decoder reconstructs the water injection volume sequence according to the feature mapping. The backpropagation algorithm based on time is used to train the LSTM-AE model with minimizing the reconstruction error as the training objective, and continuously optimize the model parameters to achieve the purpose of feature extraction. After training, only the LSTM encoder part is retained to encode the input water injection volume sequence, extract features, input into the fully connected layer, and use the error between the true value and the predicted value of the liquid production volume to train the weight parameters of the fully connected layer to realize the prediction of the liquid production volume;

[0038] Inter-well connectivity analysis module: used to take the water injection volumes of each water well in the well group as the input variables of sensitivity analysis, apply the global sensitivity analysis method EFAST to analyze the sensitivity of the input variables of the prediction model, obtain the water injection volumes of each water well and the coupling effect between the water injection volumes of the water wells through the decomposition of the model variance to calculate the parameter sensitivity index, identify the key injection wells that have a significant impact on the liquid production volume of the production well, and clarify the inter-well connectivity status.

[0039] The device first uses a data preprocessing module to preprocess the historical injection and production data of the well group, inputs the preprocessed data into the liquid production prediction module based on LSTM-AE, trains the liquid production prediction model based on LSTM-AE using the backpropagation algorithm, and uses the grid tuning algorithm to select the optimal hyperparameters of the model. The injection volumes of each injection well in the well group are used as input variables for the inter-well connectivity analysis module, and a transformation function is introduced for the input variables, thereby transforming the prediction model into a univariate periodic function. The sampling values of each variable are generated using the transformation function and input into the liquid production prediction module based on LSTM-AE. By decomposing the variance of the prediction model results, the influence of the coupling effect between the injection volumes of each water well and the injection volumes of the water wells on the model results is obtained, the parameter sensitivity index is calculated, the key injection wells that have a significant impact on the liquid production of the production wells are identified, and the inter-well connectivity status is clarified. The working flow chart of the device is shown in the appendix of the specification Figure 2 。

[0040] The following is a detailed description in combination with specific embodiments:

[0041] A well group in the **oilfield** block is used to test this method. This well group includes injection wells I1, I2, I3, I4 and production wells P1, P2, P3, P4. The experimental data are the injection and production data generated by this well group from August 6, 2015 to July 11, 2017, including three fields: production date, daily injection volume of the injection well, and daily liquid production volume of the production well. Part of the historical injection and production data of the well group is shown in Table 1.

[0042] Table 1 Partial historical injection and production data table of the experimental wells

[0043]

[0044]

[0045] The data set is preprocessed by removing missing values and outliers, normalizing the data, and dividing the data using a time window. 70% of the data is divided into the training set, and the remaining 30% of the data is used as the test set.

[0046] Use the LSTM encoder to extract features from the input water injection volume sequence, and the LSTM decoder reconstructs the water injection volume sequence according to the feature mapping. Then, use the time-based backpropagation algorithm to train the model with minimizing the reconstruction error as the training objective, continuously optimize the model parameters, and finally achieve the purpose of feature extraction. After the training of LSTM-AE is completed, only the LSTM encoder part is retained to encode the input water injection volume sequence, extract features, input the fully connected layer, and use the error between the true value and the predicted value of the liquid production volume to train the weight parameters of the fully connected layer to realize the prediction of the liquid production volume. Among them, the optimal values of hyperparameters such as epoch, batch size, units, and dropout in the liquid production volume prediction model based on LSTM-AE are determined through training using the grid search algorithm. Table 2 shows the MAPE error of the liquid production volume prediction model based on LSTM-AE on the test set.

[0047] Table 2 Error Results of the Liquid Production Volume Prediction Model on the Test Set

[0048]

[0049] Apply the global sensitivity analysis method EFAST to analyze the sensitivity of the input variables of the prediction model. Take the water injection volume of each water well in the well group as the input variables of the sensitivity analysis, introduce a transformation function for the input variables, so as to convert the model into a univariate periodic function. Use the transformation function to generate the sampling values of each variable and input them into the liquid production volume prediction model based on LSTM-AE. Through the decomposition of the model variance, obtain the coupling effect between the water injection volume of each water well and the water injection volume of the water wells to calculate the parameter sensitivity index, identify the key water wells that have a significant impact on the liquid production volume of the production well, and clarify the well connection status. The well connection coefficient results obtained by using the present invention for well connection analysis are shown in Table 2.

[0050] Table 3 Table of Well Connection Analysis Results

[0051]

[0052]

[0053] In summary, the content of the present invention is not limited to the above embodiments. Those with knowledge in the same field can easily propose other embodiments within the technical guiding ideology of the present invention, but such embodiments are all included within the scope of the present invention.

Claims

1. A method for analyzing inter-well connectivity based on LAE, characterized in that, It includes the following steps: A. Perform data preprocessing operations on the historical injection-production data of the well group, including data cleaning, dimensionless transformation, time series division, and training set-test set division; the data cleaning operation includes removing null values, duplicate values, and outliers in the dataset, and the method used for dimensionless transformation is normalization, with the formula: Among them, x represents the parameter to be normalized, x min represents the minimum value of the feature, x max represents the maximum value of the feature; the sliding window method is used to segment the data, the time window length is set to l, the step size is 1, and the data set is divided into N time series; 70% of the data set is divided into the training set, and the remaining 30% is used as the test set; B. Based on the preprocessed data, use the LSTM encoder to extract features from the input water injection volume sequence, and the LSTM decoder reconstructs the water injection volume sequence according to the feature mapping; use the time-based backpropagation algorithm to train the LSTM-AE model with minimizing the reconstruction error as the training objective, continuously optimize the model parameters to achieve the purpose of feature extraction, and the training is completed; the LSTM-AE refers to the LSTM autoencoder, and the working process is as follows. The LSTM encoder is used to extract the feature vector from the input water injection volume sequence X t ={x t ,x t +1 ,…,x t+l-1}, and splice the water injection volume x t at the t-th moment with the hidden state of the LSTM neuron at the (t - 1)-th moment, input it into the LSTM neuron corresponding to the t-th moment, and after encoding, obtain the hidden state of the LSTM. c is the number of neurons in the LSTM hidden layer, and the initial state of the first LSTM neuron is c t-1 =h t-1 =0; the hidden state of the (t + l - 1)-th LSTM neuron in the encoder is used as the initial hidden state of the decoder. Through the linear layer operation of the LSTM neuron corresponding to the (t + l - 1)-th moment in the decoder, obtain the reconstructed water injection volume vector x' t+l-1 at this moment; use x' t+l-1 and as the input of the LSTM neuron at the (t + l - 2)-th moment, obtain the hidden state of the LSTM neuron corresponding to the (t + l - 2)-th moment in the decoder and output the reconstructed vector x' t+l-2 , continue the decoding process until the reconstruction is completed, and the water injection volume sequence reconstructed by the decoder is X t '={x' t+l-1 ,x' t+l-2 ,…,x' t}; C. After training, only keep the LSTM encoder to encode the input water injection volume sequence, extract features, input them into the fully connected layer, and use the error between the true value and the predicted value of the liquid production volume to train the weight parameters of the fully connected layer to achieve the prediction of the liquid production volume; D. Take the water injection volume of each water well in the well group as the input variable for sensitivity analysis, apply the global sensitivity analysis method EFAST to analyze the sensitivity of the input variables of the liquid production volume prediction model, obtain the coupling effect between the water injection volumes of each water well and the water injection volumes through the decomposition of the model variance to calculate the parameter sensitivity index, identify the key water injection wells that have a significant impact on the liquid production volume of the production well, and clarify the well-to-well connectivity; the EFAST method is the extended Fourier amplitude sensitivity test method, which is one of the methods for global sensitivity analysis; the algorithm believes that the variance of the model is caused by the input parameters and the interaction between the parameters, and through variance decomposition, the influence degree of each variable and the coupling effect between the variables on the model is obtained to calculate the parameter sensitivity index.

2. An inter-well connectivity analysis device based on LAE, characterized in that, It includes the following modules: Data preprocessing module: used to perform preprocessing operations on the historical injection-production data of the well group, including data cleaning, dimensionless transformation, time series division, and training set-test set division; the data cleaning operation includes removing null values, duplicate values, and outliers in the dataset, and the method used for dimensionless transformation is normalization, with the formula: where x represents the parameter to be normalized, x min represents the minimum value of the feature, and x max represents the maximum value of the feature; the data is segmented using the sliding window method, the time window length is set to l, the step size is 1, and the dataset is divided into N time series; 70% of the dataset is divided into the training set, and the remaining 30% is used as the test set; Liquid production prediction module based on LSTM-AE: Use the LSTM encoder to extract features from the input water injection volume sequence, and the LSTM decoder reconstructs the water injection volume sequence according to the feature mapping; Use the time-based backpropagation algorithm to train the LSTM-AE model with minimizing the reconstruction error as the training objective, and continuously optimize the model parameters to achieve the purpose of feature extraction; After training, only keep the LSTM encoder part to encode the input water injection volume sequence, extract features, input the fully connected layer, and use the error between the true value and the predicted value of the liquid production volume to train the weight parameters of the fully connected layer to achieve the prediction of the liquid production volume; The LSTM-AE refers to the LSTM autoencoder, and the working process is as follows. The LSTM encoder is used to extract the feature vector from the input water injection volume sequence X t ={x t ,x t +1 ,…,x t+l-1}, and splice the water injection volume x t at the t-th moment with the hidden state of the LSTM neuron at the (t - 1)-th moment, input it into the LSTM neuron corresponding to the t-th moment, and obtain the hidden state of the LSTM after encoding. c is the number of neurons in the LSTM hidden layer, and the initial state of the first LSTM neuron is c t-1 =h t-1 =0; The hidden state of the (t + l - 1)-th LSTM neuron in the encoder is used as the initial hidden state of the decoder. Through the linear layer operation of the LSTM neuron corresponding to the (t + l - 1)-th moment in the decoder, the reconstructed water injection volume vector x' t+l-1 at this moment is obtained; Take x' t+l-1 and as the input of the LSTM neuron at the (t + l - 2)-th moment, obtain the hidden state of the LSTM neuron corresponding to the (t + l - 2)-th moment in the decoder and output the reconstructed vector x' t+l-2 , continue the decoding process until the reconstruction is completed. The water injection volume sequence reconstructed by the decoder is X t '={x' t+l-1 ,x' t+l-2 ,…,x' t}; Well-to-well connectivity analysis module: used to take the water injection volume of each water well in the well group as the input variable for sensitivity analysis, apply the global sensitivity analysis method EFAST to analyze the sensitivity of the input variables of the prediction model, obtain the coupling effect between the water injection volumes of each water well and the water injection volumes through the decomposition of the model variance to calculate the parameter sensitivity index, identify the key water injection wells that have a significant impact on the liquid production volume of the production well, and clarify the well-to-well connectivity; the EFAST method is the extended Fourier amplitude sensitivity test method, which is one of the methods for global sensitivity analysis; the algorithm believes that the variance of the model is caused by the input parameters and the interaction between the parameters, and through variance decomposition, the influence degree of each variable and the coupling effect between the variables on the model is obtained to calculate the parameter sensitivity index; The data preprocessed by the data preprocessing module is input into the liquid production volume prediction module based on LSTM-AE, and the water injection volumes of each water injection well obtained by the liquid production volume prediction module based on LSTM-AE are used as the input variables of the well-to-well connectivity analysis module, and the well-to-well connectivity is obtained after processing.

Images