Multi-feature multi-modal fusion tight gas reservoir yield deep learning prediction method

Abstract

The invention discloses a multi-feature multi-modal fusion tight gas reservoir yield deep learning prediction method (MMF), which improves deep learning performance through dynamic production data dimension reduction, data quality enhancement, VMD and other optimization input and LSTM fusion, and belongs to an artificial intelligence method in the technical field of oil-gas exploration and development. The method comprises the following steps: firstly, eliminating production dynamic factors having no important influence on the cumulative yield by using Granger causality test; inputting the screened features into a KF (Kalman filter) to carry out optimization estimation; performing VMD on the optimized cumulative yield to obtain a plurality of groups of sub-modal sequences; finally, multi-modal prediction and fusion are carried out based on LSTM, a prediction result of the cumulative yield is obtained, and compared with RNN, LSTM, GRU, KF-LSTM, CNN-LSTM and CNN-GRU, the R2 of MMF is improved by 0.73%-12.84%; according to the method, through dynamic feature selection, filtering, modal decomposition and deep learning, data representation is optimized, the understanding of the model on dynamic data is enhanced, and the uncertainty, the nonlinear relation and the interpretability among gas well production data are mined, so that the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to an artificial intelligence method in the field of tight gas exploration and development, specifically a multi-feature and multi-modal fusion tight gas reservoir production deep learning prediction method (MMF). Background Art

[0002] As an important component of unconventional natural gas, in the actual exploitation process, tight gas has complex production dynamics that change over time due to the reservoir properties of tight gas reservoirs, such as strong heterogeneity, small pore size, low permeability, strong stress sensitivity, and complex seepage mechanism, showing strong nonlinearity and uncertainty. For example, the output is affected by a variety of factors that may interact with each other (production time, oil pressure, casing pressure, bottom hole flow pressure, formation pressure), resulting in the possibility of multimodal prediction mode for production prediction. In addition, environmental factors, equipment errors, operation and other problems also cause the acquired data to be interfered by noise, resulting in low data quality, which affects the accurate prediction of production.

[0003] At present, in the field of gas reservoir dynamic analysis, traditional production prediction methods are usually based on empirical formulas or statistical regression models. These methods may require strict geological reservoir constraints or fail to consider the mutual influence of production dynamics, making the accuracy of traditional production prediction methods low. With the increasing maturity of machine learning, its excellent performance in dealing with nonlinear and uncertain problems has led to the adoption of machine learning-based production prediction methods in the field of oil and gas exploration and development. As for the analysis of the impact of production dynamic factors on production, the existing technology does not consider the time lag characteristics of data, and in the research on enhancing model performance by optimizing input, a single processing mechanism is often used, which may make the model insufficient in mining information, thus affecting the accuracy of prediction.

[0004] Therefore, the present invention integrates dynamic feature selection, filtering and denoising, variational mode decomposition and deep learning, and establishes a multi-feature and multi-modal fusion deep learning prediction method MMF for tight gas reservoir production to solve the above-mentioned problems. Summary of the invention

[0005] In view of the problems that the existing technology does not consider dynamic data dimensionality reduction, data quality enhancement, and low prediction accuracy, as well as the needs of practical applications, the present invention provides a new tight gas reservoir production prediction model MMF, which aims to solve the problems of insufficient multimodal fusion caused by the mutual influence of production dynamics during the exploitation of tight gas reservoirs, as well as low data quality and low prediction accuracy.

[0006] In order to solve the above problems, the present invention specifically provides the following technical solutions:

[0007] A multi-feature and multi-modal fusion deep learning prediction method for tight gas reservoir production includes the following steps:

[0008] Step S1: Acquire production dynamic data of tight gas wells;

[0009] Step S2: Feature selection of dynamic production data: Granger causality test is used to analyze the important influencing relationship between production dynamic factors and the cumulative production of tight gas, and production dynamic factors without significant influence are eliminated. Production dynamic factors with significant influence are input into the feature subset as the result of dynamic feature selection;

[0010] Step S3: filtering and denoising the dynamic production data, inputting the screened production dynamic factors and the cumulative production of tight gas into a Kalman filter (KF), outputting the optimal estimate of each production dynamic through a prediction and correction module, and obtaining a filtering result;

[0011] Step S4: variational mode decomposition of the cumulative tight gas production, obtaining two or more sub-mode sequences representing different data features through variational mode decomposition (VMD) of the optimal estimated cumulative tight gas production, wherein the optimal number of decomposed sub-modes is determined according to the sample entropy of the sub-mode sequences;

[0012] Step S5: Multi-feature and multi-modal prediction of the cumulative production of tight gas. Combine the further processed data set with the deep neural network to establish a prediction model for each sub-modality. Take the production dynamic factors and sub-modality sequence output by the filtering as input, fuse the relationship between the production dynamic factors and the sub-modality sequence through the LSTM layer, and fuse the information of the multi-dimensional production dynamic factors through the Dense layer to output the prediction results of each sub-modality. Finally, fuse the sub-modality of different data features to realize the multi-feature and multi-modal prediction of the cumulative production of tight gas and obtain the prediction result of the cumulative production of tight gas.

[0013] Preferably, the tight gas well production dynamic data obtained in step S1 include: production time, oil pressure, casing pressure, bottom hole flow pressure, formation pressure and cumulative tight gas production, marked as X=[X1, X2, X3, X4, X5, Y].

[0014] Preferably, the specific steps of step S2 are as follows:

[0015] Step S2.1: Differential stabilization of production dynamic data. Perform a stationary test on the production dynamic data based on the graphical method and the unit root test. If the production dynamic data fails the stationary test, perform a differential operation of formula (1) to obtain a stationary series.

[0016] X t ′=X t -X t-1 (1)

[0017] Among them, X t′ is the value of production dynamics at time t after differentiation, X t , X t-1 are the observed values ​​of production dynamics at time t and t-1;

[0018] Step S2.2: Perform Granger causality test on the stable production dynamics after processing in step S2.1, analyze the important influence relationship between production dynamic factors and the cumulative production of tight gas, and the production dynamic factors that pass the Granger causality test are considered to have a significant impact on the cumulative production of tight gas, and input the feature subset S. The production dynamic factors that do not pass the Granger causality test are considered to have no significant impact on the cumulative production of tight gas, and the production dynamic factors are eliminated. The feature subset S is the result of dynamic production data feature selection.

[0019] Preferably, the specific steps of step S3 are as follows:

[0020] Step S3.1: Prediction: for the feature subset S and the cumulative tight gas production in step S2.2, the predicted value at the current moment is estimated based on the predicted value at the previous moment and combined with external control. The specific formula is as follows:

[0021]

[0022] in, is the a priori estimate of the production dynamics at time t; is the estimated value of production dynamics at time t-1, assuming that u t is the external control input; is the prior estimate of the production dynamic covariance matrix at time t; [P S,t-1 ,P Y,t-1 ] is the estimated value of the production dynamic covariance matrix at time t-1; A is the state transfer matrix; B is the control matrix, because the gas well production process is affected by many factors and is not artificially controllable, it is taken as 0; Q is the mean square error matrix of the process noise, assuming that the process noise is deterministic, it is taken as 0;

[0023] Step S3.2: Correction: Correct the predicted value at the current moment by taking the difference between the measured value and the predicted value at the current moment of the production dynamics, thereby obtaining the optimal estimated value. The specific formula is as follows:

[0024]

[0025] Among them, [K S,t ,K Y,t ] is the Kalman gain; H is the measurement matrix, which indicates the degree of dependence of the actual system state on the calculated data, and is taken as 1 in practice; R is the noise covariance matrix; is the optimal estimate of production dynamics at time t; [St ,Y t ] is the actual observed value of production dynamics at time t; [P S,t ,P Y,t ] is the error covariance matrix of the production dynamics after the production dynamics Kalman filtering at time t.

[0026] Preferably, the specific steps of step S4 are as follows:

[0027] Step S4.1: The cumulative production sequence of tight gas output from step S3 The variational constraint equation based on VMD satisfies the following model:

[0028]

[0029] Where K represents the number of decomposed sub-modes; represents the partial derivative with respect to time; δ(t) represents the Dirac function; represents the submodal sequence; {φ k}={φ1,φ2,…,φ K} represents the center frequency of the sub-mode; * represents the convolution operation;

[0030] Step S4.2: To solve equation (7) in step S4.1, the Lagrangian operator λ(t) and the quadratic multiplication factor α are introduced to construct the augmented Lagrangian function of the structure, and the constrained variation problem is transformed into the unconstrained variation problem of equation (8);

[0031]

[0032] Step S4.3: Use the alternating direction multiplier algorithm to solve equation (8) in step S4.2 and obtain φ k The expression of continuous iterative update of and λ, thus decomposing the cumulative production of tight gas into K sub-modal sequences;

[0033] Step S4.4: Calculate the sample entropy of each submodal sequence and determine the number of submodal components K. The larger the value of the sample entropy, the more complex the submodal sequence, and the more difficult it is for the prediction model to learn the changing trend of the sequence.

[0034] Preferably, the specific steps of step S5 are as follows:

[0035] Step S5.1: normalizing the production dynamic factors optimized in step 3 and the tight gas cumulative production sub-modal sequence decomposed in step 4, and dividing the normalized data set by time window, with the first 80% as a training set and the last 20% as a test set;

[0036] Step S5.2: Combine LSTM and Dense to build a deep neural network prediction model for each sub-modality, where the LSTM layer captures the relationship between production dynamic factors and sub-modalities, and the Dense layer integrates the information of multi-dimensional production dynamic factors. The model is trained with the training set processed in step S5.1, and the weight and bias of the model are adjusted by gradient descent. The hyperparameters are adjusted by the loss function to better fit the influence relationship between the sub-modality and each production dynamic factor.

[0037] Step S5.3: Use the model trained in step S5.2 to predict the test set, input the production dynamic factors processed in step S5.1 into the prediction model, and generate the prediction results of each sub-mode;

[0038] Step S5.4: The prediction results of each sub-mode outputted in step S5.3 are integrated to obtain the prediction result of the cumulative production of tight gas. The integrated reconstruction formula is as follows:

[0039]

[0040] in, is the prediction result of the kth sub-modality, It is the prediction result of the cumulative production of tight gas.

[0041] The present invention discloses the following technical effects:

[0042] The deep learning model MMF proposed in the present invention is based on multi-feature and multi-modal fusion. Firstly, the Granger causality test is used to realize dynamic feature selection, the time lag of production dynamics is considered, and the important influencing relationship between the production dynamics of tight gas wells is analyzed in an interpretable manner, thereby reducing the complexity of the model. Secondly, data filtering is realized by KF, which greatly reduces the noise interference of the data and enhances the data quality. Then, VMD is used to realize the modal decomposition of the cumulative production, deeply analyzes the dynamic change process of the cumulative production, extracts different data features in the cumulative production, and reveals the potential laws of the model. Finally, based on LSTM, the relationship between dynamic data is captured, Dense fuses the information of multi-dimensional dynamic features, and fuses the different data features of each sub-modality, thereby realizing the multi-feature and multi-modal prediction of the cumulative production of tight gas, exploring the uncertainty and nonlinear relationship between the production data of gas wells, and improving the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solution of the embodiment of the present invention, the drawings required for use in the embodiment are briefly introduced below.

[0044] Figure 1 It is a flow chart of the present invention;

[0045] Figure 2 It is a flow chart of a specific embodiment of the present invention;

[0046] Figure 3 Comparison between the actual observation value of production time and the filtered result;

[0047] Figure 4 Comparison between the actual observed value of oil pressure and the filtered result;

[0048] Figure 5 Comparison between the actual observed value of tight gas cumulative production and the filtered result;

[0049] Figure 6 is the result of variational mode decomposition of tight gas cumulative production;

[0050] Figure 7 is the prediction result of sub-mode 1;

[0051] Figure 8 is the prediction result of sub-mode 2;

[0052] Fig. 9 This is the prediction result of the cumulative production of tight gas. DETAILED DESCRIPTION

[0053] In order to more clearly understand the purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods, and the technical solution is further illustrated by a specific embodiment. Obviously, the embodiments described are only partial embodiments of the present invention, and the protection scope of the present invention is not limited to the specific embodiments disclosed below. For those skilled in the art, all other embodiments obtained without creative work are within the protection scope of the present invention.

[0054] Figure 1 The flowchart of the present invention is shown in the following example. Figure 2 , including the following steps:

[0055] Step S1: Acquire production dynamic data of tight gas wells;

[0056] The monthly production data of a tight gas well obtained in the specific embodiment includes: production time, oil pressure, casing pressure, bottom hole flow pressure, formation pressure, and cumulative tight gas production;

[0057] Step S2: Feature selection of dynamic production data;

[0058] The specific implementation steps are as follows:

[0059] Step S2.1: Differential stabilization of production dynamic data. Perform a stationary test on the production dynamic data based on the graphical method and the unit root test. If the production dynamic data fails the stationary test, perform a differential operation of formula (1) to obtain a stationary series.

[0060] Step S2.2: Perform Granger causality test on the production dynamics after differential stabilization in step S2.1, that is, first set the null hypothesis that the production dynamics factor is not the cause of the future changes in the cumulative production of tight gas, then establish the constrained and unconstrained regression models of formula (10) and formula (11) for the processed production dynamics, and then use the least squares method to calculate the coefficient a of the regression model. j With b j Parameter estimation is performed, and then the F statistic F is calculated according to formula (12), which is compared with the critical value F at the significance level α = 0.05. α Compare, if F≤F α , then accept the null hypothesis that the production dynamic factor has no significant effect on the cumulative production of tight gas, and input the feature subset S. On the contrary, if F>F α , then the original hypothesis is rejected, the production dynamic factor has a significant impact on the cumulative production of tight gas, and the production dynamic factor is eliminated;

[0061]

[0062] Among them, x i ′ is the production dynamic factor after differential stabilization of the ith gas, y′ is the cumulative production of tight gas after differential stabilization, a j , b j is the regression coefficient, l and k are the lag orders, T is the sequence length, η 1t , η 2t is a random item, are the predicted values ​​of the constrained and unconstrained regression models, respectively.

[0063] In this example, the P value is used as the evaluation of the test result, P = P (F > F α＝0.05 ), then P>0.05, accept the null hypothesis, production dynamic factors have no significant effect on cumulative gas production, P<0.05, reject the null hypothesis, production dynamic factors have a significant effect on cumulative gas production. Based on the test results shown in Table 1, it can be seen that the test P values ​​based on production time and formation pressure are 0.012 and 0.0026 respectively, which are less than 0.05, reject the null hypothesis, and have a significant impact on the future changes of tight gas cumulative production. Therefore, production time and formation pressure are selected as the result of dynamic feature selection of production data, that is, feature subset S=[X1,X5].

[0064] Table 1 Results of dynamic feature selection

[0065]

[0066] Step S3: Dynamic production data filtering and denoising;

[0067] The specific implementation steps are as follows:

[0068] Step S3.1: prediction: for the production dynamic factors and the cumulative production of tight gas screened out in step S2.2, the prediction values ​​and error covariance matrix at the current moment are estimated based on the prediction values ​​at the previous moment and combined with external control. The calculation formulas are shown in equations (2)-(3);

[0069] Step S3.2: Correction: Correct the predicted value at the current moment by using the difference between the measured value and the predicted value of the production dynamics at the current moment, calculate the Kalman gain at time t, the optimal estimate of the production dynamics, and the error covariance matrix of the production dynamics after Kalman filtering. The calculation formulas are shown in equations (4)-(6).

[0070] The results of filtering the production time, formation pressure and tight gas cumulative production are shown in Table 2 and Figure 3-5 As shown in the figure, after data filtering, the fluctuation of the production time peak becomes smoother, and the sharp inflection points of formation pressure and tight gas cumulative production weaken, which is more consistent with the actual production changes of tight gas wells.

[0071] Table 2 Results of production data filtering

[0072]

[0073]

[0074] Step S4: variational mode decomposition of the cumulative production of tight gas;

[0075] The specific implementation steps are as follows:

[0076] Step S4.1: for the tight gas cumulative production sequence outputted in step S3, a variational constraint equation is constructed based on VMD to satisfy the model shown in formula (7);

[0077] Step S4.2: Introduce the Lagrangian operator λ(t) and the quadratic multiplication factor α to construct the augmented Lagrangian function of the structure, and transform the constrained variational problem of equation (7) into an unconstrained variational problem as shown in equation (8);

[0078] Step S4.3: Use the alternating direction multiplier algorithm to solve formula (8) and obtain φ k The expressions for the continuous iterative update of and λ are shown in formulas (13)-(15), and based on the set number of sub-modes K, φ k and λ are iteratively updated until the iteration stop condition shown in formula (16) is met, and the sub-modal sequence results are output;

[0079]

[0080] in, and yes λ(t) and Fourier transform of ; n is the number of iterations;

[0081] Step S4.4: Calculate the sample entropy of each submodal sequence output in step S4.3 to determine the number of submodal components. The larger the value of the sample entropy, the more complex the submodal sequence, and the more difficult it is for the prediction model to learn the changing trend of the sequence.

[0082] In this example, K=[2,6] is set, and VMD is used to decompose it into K sub-modes. The sample entropy of different sub-modes is calculated as shown in Table 3. The sample entropy of sub-mode 1 and sub-mode 2 is about 0.02, while the sample entropy of sub-mode 3 increases significantly to about 0.2, indicating that the sequence complexity of sub-mode 3 increases, which will lead to an increase in the complexity of the prediction model. Therefore, the optimal decomposition result is to decompose the cumulative production of tight gas into two sub-mode sequences using VMD.

[0083] Table 3 Sample entropy of sub-modes under different K values

[0084]

[0085] The decomposed submodes are as follows: Figure 6 As shown, it reflects the different data characteristics of the cumulative production of tight gas, among which sub-mode 1 represents the data characteristics of the overall increasing trend of the cumulative production of tight gas, and sub-mode 2 represents the data characteristics of the random fluctuation of the cumulative production.

[0086] Step S5: Deep learning model prediction of tight gas cumulative production;

[0087] The specific implementation steps are as follows:

[0088] Step S5.1: normalizing the processed production dynamic factors and the tight gas cumulative production sub-modal sequence, dividing the normalized data set by time window, and taking the first 80% as a training set and the last 20% as a test set;

[0089] Step S5.2: Combine LSTM and Dense to build a deep neural network prediction model for each sub-mode. The model includes an input layer, a hidden layer and an output layer. The hidden units based on the LSTM hidden layer determine the retention and forgetting of data information through the input gate, the forget gate and the output gate to achieve the learning of long-term trend changes in the sequence. The input layer accepts the production dynamic data related to the tight gas well after further processing. Several hidden units of the LSTM hidden layer mine the mutual influence relationship between the production dynamics and the tight gas cumulative production sub-mode. The Dense hidden layer integrates the characteristic information of multi-dimensional production dynamic factors. The output layer outputs the prediction result of the tight gas cumulative production sub-mode. The prediction model is trained using the training set processed in step S5.1. The weight and bias of the model are adjusted by gradient descent. The hyperparameters are adjusted by the loss function to make it better fit the relationship between the sub-mode and each production dynamic.

[0090] Step S5.3: Use the model trained in step S5.2 to predict the test set, input the production dynamic factors processed in step S5.1 into the prediction model, and generate the prediction results of each sub-mode;

[0091] Figure 7-8 The prediction results of sub-modality 1 and sub-modality 2 are shown respectively. The left side of the dotted line is the training set, and the right side is the test set. Figure 7-8 ,The sub-modality shows good fitting effects on both the training set and the test set, indicating that the data features of the sub-modality are fully mined in the model, and can well predict the future change trend of each sub-modality sequence.

[0092] Step S5.4: reconstruct the prediction results of each sub-mode outputted in step S5.3 by formula (9), integrate the different data features of the sub-modes, and obtain the prediction results of the cumulative production of tight gas;

[0093] Fig. 9 The figure shows the prediction results of the cumulative production of tight gas after the fusion of sub-modes. The left side of the dotted line is the training set, and the right side is the test set. Fig. 9 The cumulative production of tight gas shows good prediction results on both the training set and the test set, indicating that the model integrates the feature information of different sub-modalities well, makes accurate predictions on the data, and has good generalization ability.

[0094] The prediction model is evaluated based on the deviation between the prediction result and the actual measured value. The indicators used to evaluate the prediction performance of the model are: determination coefficient (R 2 ), mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE) and mean absolute percentage error (MAPE), where R 2The closer it is to 1, the smaller the difference between the predicted value and the actual observed value, and the stronger the model's prediction ability. The smaller the MAE, RMSE, and MAPE, the smaller the error between the model's predicted value and the actual observed value, and the higher the model's prediction accuracy.

[0095] In this example, RNN, LSTM, GRU, KF-LSTM, CNN-LSTM, CNN-GRU and MMF are used for comparative analysis. Table 4 shows the prediction performance indicators of different models. Among them, the evaluation indicators used are: R 2 , MAE, RMSE and MAPE. Based on Table 1, it can be found that compared with other models, MMF has the best evaluation indicators and significantly better prediction performance.

[0096] Table 4 Prediction performance indicators of different models

[0097]

[0098] The above is only a specific implementation of the invention, but the protection scope of the invention is not limited thereto. Those skilled in the art should understand that the invention includes but is not limited to the contents described in the above specific implementation. Any modification that does not deviate from the functional and structural principles of the invention will be included in the scope of the claims.

Claims

1. A multi-feature and multi-modal fusion deep learning prediction method for tight gas reservoir production, characterized in that: The method specifically comprises: Step S1: Acquire production dynamic data of tight gas wells; Step S2: Feature selection of dynamic production data: Granger causality test is used to analyze the important influencing relationship between production dynamic factors and the cumulative production of tight gas, and production dynamic factors without significant influence are eliminated. Production dynamic factors with significant influence are input into the feature subset as the result of dynamic feature selection; Step S3: Filtering and denoising the dynamic production data, inputting the screened production dynamic factors and the cumulative production of tight gas into the Kalman filter KF, outputting the optimal estimate of each production dynamic through the prediction and correction module, and obtaining the filtering result; Step S4: variational mode decomposition of the cumulative tight gas production, obtaining two or more sub-mode sequences showing different data characteristics through variational mode decomposition (VMD) of the optimal estimated cumulative tight gas production, wherein the optimal number of decomposed sub-modes is determined according to the sample entropy of the sub-mode sequences; Step S5: Multi-feature and multi-modal prediction of the cumulative production of tight gas. Combine the further processed data set with the deep neural network to establish a prediction model for each sub-modality. Take the production dynamic factors and sub-modality sequence output by the filtering as input, fuse the relationship between the production dynamic factors and the sub-modality sequence through the LSTM layer, and fuse the information of the multi-dimensional production dynamic factors through the Dense layer to output the prediction results of each sub-modality. Finally, fuse the sub-modality of different data features to realize the multi-feature and multi-modal prediction of the cumulative production of tight gas and obtain the prediction result of the cumulative production of tight gas.

2. The multi-feature and multi-modal fusion deep learning prediction method for tight gas reservoir production according to claim 1 is characterized in that: The tight gas well production dynamic data obtained in step S1 include: production time, oil pressure, casing pressure, bottom hole flow pressure, formation pressure and tight gas cumulative production, marked as X=[X1, X2, X3, X4, X5, Y].

3. The multi-feature and multi-modal fusion deep learning prediction method for tight gas reservoir production according to claim 1 is characterized in that: The specific steps of step S2 are as follows: Step S2.1: Differential stabilization of production dynamic data. Perform a stationary test on the production dynamic data based on the graphical method and the unit root test. If the production dynamic data fails the stationary test, perform a differential operation of formula (1) to obtain a stationary series. X t ′=X t -X t-1 (1) Among them, X′ t is the value of the production dynamics at time t after differentiation, X t , X t-1 are the observed values ​​of production dynamics at time t and t-1; Step S2.2: Perform Granger causality test on the stable production dynamics after processing in step S2.1, analyze the important influence relationship between production dynamic factors and the cumulative production of tight gas, and the production dynamic factors that pass the Granger causality test are considered to have a significant impact on the cumulative production of tight gas, and input the feature subset S. The production dynamic factors that do not pass the Granger causality test are considered to have no significant impact on the cumulative production of tight gas, and the production dynamic factors are eliminated. The feature subset S is the result of dynamic production data feature selection.

4. The multi-feature and multi-modal fusion deep learning prediction method for tight gas reservoir production according to claim 1 is characterized in that: The specific steps of step S3 are as follows: Step S3.1: Prediction: for the feature subset S and the cumulative tight gas production in step S2.2, the predicted value at the current moment is estimated based on the predicted value at the previous moment and combined with external control. The specific formula is as follows: in, is the a priori estimate of the production dynamics at time t; is the estimated value of production dynamics at time t-1, assuming that u t is the external control input; is the prior estimate of the production dynamic covariance matrix at time t; [P S,t-1 ,P Y,t-1 ] is the estimated value of the production dynamic covariance matrix at time t-1; A is the state transfer matrix; B is the control matrix, because the gas well production process is affected by many factors and is not artificially controllable, it is taken as 0; Q is the mean square error matrix of the process noise, assuming that the process noise is deterministic, it is taken as 0; Step S3.2: Correction: Correct the predicted value at the current moment by taking the difference between the measured value and the predicted value at the current moment of the production dynamics, thereby obtaining the optimal estimated value. The specific formula is as follows: Among them, [K S,t ,K Y,t ] is the Kalman gain; H is the measurement matrix, which indicates the degree of dependence of the actual system state on the calculated data, and is taken as 1 in practice; R is the noise covariance matrix; is the optimal estimate of production dynamics at time t; [S t ,Y t ] is the actual observed value of production dynamics at time t; [P S,t ,P Y,t ] is the error covariance matrix of the production dynamics after the production dynamics Kalman filtering at time t.

5. The multi-feature and multi-modal fusion deep learning prediction method for tight gas reservoir production according to claim 1 is characterized in that: The specific steps of step S4 are as follows: Step S4.1: The cumulative production sequence of tight gas output from step S3 The variational constraint equation based on VMD satisfies the following model: Where K represents the number of decomposed sub-modes; represents the partial derivative with respect to time; δ(t) represents the Dirac function; represents the submodal sequence; {φ k }={φ1,φ2,…,φ K } represents the center frequency of the sub-mode; * represents the convolution operation; Step S4.2: To solve equation (7) in step S4.1, the Lagrangian operator λ(t) and the quadratic multiplication factor α are introduced to construct the augmented Lagrangian function of the structure, and the constrained variation problem is transformed into the unconstrained variation problem of equation (8); Step S4.3: Use the alternating direction multiplier algorithm to solve equation (8) in step S4.2 and obtain φ k The expression of continuous iterative update of and λ, thus decomposing the cumulative production of tight gas into K sub-modal sequences; Step S4.4: Calculate the sample entropy of each submodal sequence and determine the number of submodal components K. The larger the value of the sample entropy, the more complex the submodal sequence, and the more difficult it is for the prediction model to learn the changing trend of the sequence.

6. The multi-feature and multi-modal fusion deep learning prediction method for tight gas reservoir production according to claim 1 is characterized in that: The specific steps of step S5 are as follows: Step S5.1: normalizing the production dynamic factors optimized in step S3 and the tight gas cumulative production sub-modal sequence decomposed in step 4, and dividing the normalized data set by time window, with the first 80% as a training set and the last 20% as a test set; Step S5.2: Combine LSTM and Dense to build a deep neural network prediction model for each sub-modality, where the LSTM layer captures the relationship between production dynamic factors and sub-modalities, and the Dense layer integrates the information of multi-dimensional production dynamic factors. The model is trained with the training set processed in step S5.1, and the weight and bias of the model are adjusted by gradient descent. The hyperparameters are adjusted by the loss function to better fit the influence relationship between the sub-modality and each production dynamic factor. Step S5.3: Use the model trained in step S5.2 to predict the test set, input the production dynamic factors processed in step S5.1 into the prediction model, and generate the prediction results of each sub-mode; Step S5.4: The prediction results of each sub-mode outputted in step S5.3 are integrated to obtain the prediction result of the cumulative production of tight gas. The integrated reconstruction formula is as follows: in, is the prediction result of the kth sub-modality, It is the prediction result of the cumulative production of tight gas.

Images