A method for obtaining gas-water relative permeability curve of tight sandstone based on DAE-DNN model

Abstract

The present invention discloses a method for obtaining a gas-water relative permeability curve of tight sandstone based on a DAE-DNN model. The relative permeability curve is obtained by a Brooks-Corey model, and combined with fracture parameters, construction parameters, and seepage characteristic parameters, production data is obtained through numerical simulation of gas reservoirs, including water production, gas production, and pressure. The production data, fracture parameters, construction parameters, and seepage characteristic parameters are subjected to dimensionality reduction processing using DAE, and the dimensionality reduction data is used as input, and the parameters related to the relative permeability curve are used as output. The parameters related to the relative permeability curve are imported into DNN to predict, and finally the relative permeability curve is determined by normalizing 10 groups of relative permeability curves predicted by the prediction simulation set. The present invention can quickly and accurately calculate the gas-water relative permeability curve, better reflect the actual gas-water flow capacity, and can provide strong support and guidance for gas field development and production.

Description

Technical Field

[0001] The invention relates to the technical field of oil and gas field development, and in particular to a method for obtaining a tight sandstone gas-water relative permeability curve based on a DAE-DNN model. Background Art

[0002] Tight sandstone reservoirs generally have a high initial water saturation, and are prone to gas-water two-phase flow characteristics during the development process, resulting in prominent water production problems in some gas wells, affecting the gas well productivity. The gas-water relative permeability curve is a key factor affecting the accurate simulation and prediction of tight gas well productivity. The common gas-water relative permeability curve determination methods can be divided into direct determination and indirect determination. The direct determination method includes a stable method and an unstable method. The stable method takes a long time and cannot represent the actual flow conditions of the reservoir fluid. The unstable method has a simple calculation process and a short time, but in terms of calculation, due to the lax simplified assumptions, it brings a lot of unreliability in interpretation. For gas-driven reservoirs, it is difficult to obtain gas-liquid relative permeability curves using both steady-state and non-steady-state methods. Therefore, it can be obtained by indirect determination. Common indirect determination methods include capillary force curve calculation, field data calculation, and empirical formula calculation. Among them, the use of empirical formulas to calculate relative permeability can avoid the trouble of laboratory determination, which is very beneficial for predicting relative permeability and numerical simulation.

[0003] Based on the existing patents and papers, it is found that when using the indirect method to obtain the relative permeability curve of tight sandstone, the following contents are not taken into account:

[0004] (1) Production data is not taken into account in the method of obtaining the relative permeability curve of tight sandstone. The dynamic changes of gas reservoir physical properties, development methods, etc. will be comprehensively reflected in the production data. Therefore, the relative permeability curve obtained using the dynamic data of the gas reservoir can more truly reflect the actual gas-water flow capacity. In the existing technology, when obtaining the relative permeability curve through production data, the material balance method is used, which is difficult to calculate and time-consuming.

[0005] (2) When existing methods use machine learning to obtain the phase permeability curve, they lack dimensionality reduction processing of the input data. Since production data has the characteristics of high dimensionality, in order to reduce the difficulty of solution, reduce calculation time, and improve learning ability, it is necessary to reduce the dimensionality of the input data.

[0006] (3) When existing methods use machine learning to obtain phase permeability curves, they do not use deep neural networks. Deep neural networks refer to neural networks with more than three layers. Their powerful nonlinear fitting ability makes them the first choice for solving various complex nonlinear mapping problems. In addition, they do not take into account the multi-solution nature of neural network models. Summary of the invention

[0007] In view of the characteristics of the conventional relative permeability curve acquisition calculation process being complex, time-consuming and costly, with the development of artificial intelligence, machine learning methods in artificial intelligence have been applied to research in multiple fields of petroleum engineering. This paper proposes a method for obtaining tight sandstone gas-water relative permeability curve based on the DAE-DNN model to simplify the relative permeability curve acquisition process, reduce time consumption and reduce costs.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] A method for obtaining a tight sandstone gas-water relative permeability curve based on a DAE-DNN model comprises the following steps:

[0010] Step 1: Generate multiple sets of relative permeability curves according to the Brooks-Corey model;

[0011] Step 2: Simulate the gas-water two-phase capacity according to the relative permeability curve;

[0012] Step 3: Sample set construction;

[0013] Step 4: Use the DAE-DNN model to construct a tight sandstone gas-water relative permeability curve prediction model set;

[0014] Step 5: Normalize the phase permeability curve obtained by the prediction model set and verify its accuracy.

[0015] The specific process of generating multiple sets of relative permeability curves by the Brooks-Corey model in step 1 is as follows:

[0016] The Brooks-Corey model is expressed as:

[0017]

[0018]

[0019]

[0020] Where: k ro , k rw , k rg —relative permeabilities of oil phase, water phase and gas phase, respectively;

[0021] k ro ,max,k rw ,max,k rg ,max—the maximum relative permeability of oil phase, water phase and gas phase respectively;

[0022] s o 、s w 、sg — respectively, the saturation of oil phase, water phase and gas phase;

[0023] s or 、s wc 、s gc —respectively residual saturation of oil phase, water phase and gas phase;

[0024] n o 、n w 、n g —Shape factor, which determines the shape of the relative permeability curve and has a value range of 1-6;

[0025] According to the above Brooks-Corey model, the irreducible water saturation (s wc ), residual gas saturation (s gc ), shape factor (n w 、n g ) are variable parameters, and the other parameters are fixed parameters, and multiple sets of relative permeability curves are generated.

[0026] In step 2, the gas-water two-phase capacity simulation is performed according to the relative permeability curve, which specifically includes:

[0027] Step 21: Simulation of fracture network formed by hydraulic fracturing;

[0028] Step 22: gas-water two-phase flow simulation;

[0029] Step 23: The gas reservoir numerical model obtains three sets of production data: water production (FWPR), gas production (FGPR), and pressure (FPR).

[0030] In the simulation of the fracture network formed by hydraulic fracturing in step 21, the fracturing scheme design is simulated.

[0031] In step 22, the gas-water two-phase flow simulation is performed by setting the critical flow saturation and the starting pressure gradient.

[0032] The sample set construction in step 3 specifically includes: using production data as input layer samples, where the production data includes water production, gas production, and pressure; using bound water saturation (s wc ), residual gas saturation (s gc ), shape factor (n w 、n g ) constitute the output layer samples.

[0033] In step 4, the DAE-DNN model is used to construct a tight sandstone gas-water relative permeability curve prediction model set, which specifically includes:

[0034] Step 41: Data cleaning and processing;

[0035] Step 42: The encoder part of the DAE (deep autoencoder) reduces the dimension of the input sample; the decoder part of the DAE reconstructs the reduced dimension data, and outputs the reduced dimension data when the root mean square error (RMSE) meets the requirements;

[0036] Step 43: The reduced-dimensional input samples are passed through DNN models with different structures to obtain relative permeability curves.

[0037] In step 41, the data is cleaned and processed, specifically: invalid values ​​and abnormal values ​​in the data are removed, missing data are supplemented, and the integrity and validity of the data during the entire production period are ensured. The data is processed using the Min-max standardization method, so that the values ​​corresponding to all factors are uniformly mapped to the [0, 1] interval.

[0038] In step 42, the DAE (deep autoencoder) consists of a 5-layer structure, including 1 input layer, 1 output layer, and 3 hidden layers. When the root mean square error (RMSE) between the reconstructed data and the original input data is less than 0.01, the error requirement is met.

[0039] The construction of the DNN model in step 43 is specifically as follows:

[0040] Step 431: training algorithm determination;

[0041] Step 432: forming a prediction model set by setting different hidden layers;

[0042] The training algorithm is determined in step 431. There are three deep neural network training algorithms: Levenberg-Maquardt (LM), Scaled Conjugate Gradient (SCG), and Bayesian Regularization (BR). When the hidden layer settings are consistent, the above three training algorithms are used to train the model respectively, where the maximum number of training rounds is 1000, and the training algorithm with a smaller mean square error is selected as the training algorithm.

[0043] In step 432, a prediction model set is formed by setting different hidden layers. The purpose is to set multiple groups of different hidden layer numbers (ranging from 3 to 5 layers) and node numbers (ranging from 1 to 100) for model training through the selected training algorithm. When the root mean square error (RMSE) is less than 0.01, it can become a group of prediction models. Finally, 10 groups of models form a prediction model set.

[0044] In step 5, the phase permeability curve obtained by the prediction model set is normalized and the accuracy is verified. Specifically:

[0045] The prediction model set normalizes 10 phase permeability curves, and obtains the numerical model water production rate (FWPR), gas production rate (FGPR), and pressure (FPR) curves through the numerical model. The model is verified by the mean square error between the numerical model production curve and the actual production curve.

[0046] Compared with the prior art, the advantages of the present invention are:

[0047] Compared with the direct method, the present invention reduces costs and avoids the trouble of laboratory measurement; compared with the indirect method, the present invention fully utilizes production data based on machine learning, more truly reflects the actual gas-water flow capacity, and avoids complex calculations, reduces time consumption, and reduces costs.

[0048] Compared with the existing neural network for phase permeation, the present invention uses a deep neural network, adds a deep autoencoder to reduce the dimension of the data, and takes into account the multi-solution nature of the deep neural network, thereby reducing the difficulty of solution, reducing the calculation time, and improving the learning ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the model formation process of the present invention;

[0050] Figure 2 It is the DAE-DNN model network structure of the present invention;

[0051] Figure 3a is a porosity distribution map of the three-dimensional geological model of the present invention;

[0052] Figure 3b is a permeability field distribution map of the three-dimensional geological model of the present invention;

[0053] Figure 4 It is a graph of relative permeability of some gases and water produced randomly by the present invention;

[0054] Figure 5 It is the actual production dynamic curve of the present invention;

[0055] Figure 6 It is a comparison diagram of the sample core phase permeability curve and the predicted phase permeability curve of the present invention;

[0056] Figure 7a It is a comparison diagram of the water production curve of the present invention and the predicted water production curve;

[0057] Figure 7b is an error diagram between the water production curve of the present invention and the predicted water production curve;

[0058] Figure 8a is a comparison diagram of the pressure curve of the present invention and the predicted pressure curve;

[0059] Figure 8bIt is an error diagram between the pressure curve of the present invention and the predicted pressure curve. DETAILED DESCRIPTION

[0060] In order to make the purpose, research method and advantages of the present invention more clearly understood, the following examples are listed to further illustrate the present invention in detail.

[0061] like Figure 1 The process shown in the figure, a method for obtaining a tight sandstone gas-water relative permeability curve based on a DAE-DNN model of the present invention comprises the following steps:

[0062] Step 1: According to Figure 2 The Brooks-Corey model shown generates multiple sets of relative permeability curves.

[0063] 500 sets of relative permeability curves were randomly generated using the Brooks-Corey model, and the parameters of the generated relative permeability curves are shown in Table 1.

[0064] Table 1 Parameters of randomly generated relative permeability curves

[0065]

[0066]

[0067] Step 2: Perform gas-water two-phase capacity simulation based on the relative permeability curve.

[0068] Taking the X well group in a gas reservoir in Sichuan Basin as the research object, a three-dimensional geological model was established. A numerical model was established on the basis of the geological model. On the basis of the phase permeability curve produced in Table 1, the fracture parameters, construction parameters, and seepage characteristic parameters (as shown in Table 2) were set, and the 7275-day water production (FWPR), gas production (FGPR), and pressure (FPR) data were obtained through numerical simulation of the gas reservoir.

[0069] Taking the X well group in a gas reservoir in Sichuan Basin as the research object, a three-dimensional geological model was established. Based on the geological model, a numerical model was established. Through numerical simulation of the reservoir, the 7275-day water production (FWPR), gas production (FGPR) and pressure (FPR) data were obtained.

[0070] Figure 3a and Figure 3b They are the porosity and permeability field distribution maps of the three-dimensional geological model.

[0071] Figure 4 This is a graph of relative permeability of gas and water for some random production.

[0072] Step 3: Sample set construction.

[0073] The production data is used as the input layer sample, where the production data includes water production, gas production, and pressure; the bound water saturation (s wc ), residual gas saturation (s gc ), shape factor (n w 、n g ) constitute the output layer samples.

[0074] Randomly select 70% of the data as the training set, 15% of the data as the validation set, and 15% of the data as the test set.

[0075] Step 4: Use the DAE-DNN model to construct a tight sandstone gas-water relative permeability curve prediction model set.

[0076] Step 41: Data cleaning and processing;

[0077] Invalid values ​​and abnormal values ​​in the data are removed, missing data are filled, and the integrity and validity of the data during the entire production period are ensured. The data is processed using the Min-max standardization method, so that the values ​​corresponding to all factors are uniformly mapped to the [0, 1] interval.

[0078] Step 42: The encoder part of the DAE (deep autoencoder) reduces the dimension of the input sample; the decoder part of the DAE reconstructs the reduced dimension data and outputs the reduced dimension data when the root mean square error (RMSE) meets the requirements;

[0079] DAE (Deep Autoencoder) consists of a 5-layer structure, with 1 input layer, 1 output layer, and 3 hidden layers. When the root mean square error (RMSE) between the reconstructed data and the original input data is less than 0.01, the error requirement is met. The standardized production data, crack parameters, construction parameters, and seepage characteristic parameters 21825 error requirement data were input into DAE. After analyzing different dimensionality reduction numbers (as shown in Table 2), it was determined that the dimensionality reduction was reduced to 1000 data.

[0080] Table 2 RMSE corresponding to different dimensionality reduction numbers

[0081]

[0082] Step 43: The reduced-dimensional input samples are passed through DNN models with different structures to obtain relative permeability curves.

[0083] Step 431: Training algorithm determination

[0084] There are three neural network training algorithms: Levenberg-Maquardt (LM), Scaled Conjugate Gradient (SCG), and Bayesian Regularization (BR). When the number of samples and the hidden layer settings are consistent, the above three training algorithms are used to train the model. The maximum number of training rounds is 1000. The training results are shown in Table 3. It can be seen from Table 3 that the LM algorithm has a faster training time and a smaller error; the SCG algorithm has a short training time and high memory efficiency, but a larger error; the BR algorithm has the smallest error and strong generalization ability, but a slow training speed. In summary, the LM algorithm is selected as the training algorithm for the model.

[0085] Table 3 Training status of three training algorithms

[0086]

[0087] Step 432: forming a prediction model set by setting different hidden layers;

[0088] The models were trained under ten different sets of hidden layers, and the following training results were obtained, as shown in Table 4. The ten sets of prediction models were formed into a prediction model set.

[0089] Table 4 Prediction model set

[0090]

[0091]

[0092] Step 5: Normalize the phase permeability curve obtained by the prediction model set and verify its accuracy.

[0093] Taking the actual production dynamic curve as input, the relative permeability curve is predicted by the prediction model set in Table 4. The 10 predicted relative permeability curves are normalized, and the numerical model water production (FWPR), gas production (FGPR), and pressure (FPR) curves are obtained through the numerical model. The mean square error between the numerical model production curve and the actual production curve is small.

[0094] Figure 5 It is the actual production dynamic curve;

[0095] Figure 6 It is a comparison diagram between the sample core phase permeability curve and the predicted phase permeability curve;

[0096] Figure 7a and Figure 7b They are the comparison of water production curve and predicted water production curve and the error diagram;

[0097] Figure 8a and Figure 8bThey are the comparison between the pressure curve and the predicted pressure curve and the error diagram.

[0098] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the implementation methods of the present invention, and should be understood that the protection scope of the present invention is not limited to such special statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical enlightenment disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for obtaining the gas-water relative permeability curve of tight sandstone based on DAE-DNN model, characterized in that: The following steps are involved: Step 1: Generate multiple sets of relative permeability curves according to the Brooks-Corey model; Step 2: Simulate the gas-water two-phase capacity according to the relative permeability curve; Step 3: Sample set construction; Step 4: Use the DAE-DNN model to construct a tight sandstone gas-water relative permeability curve prediction model set; Step 5: Normalize the phase permeability curve obtained by the prediction model set and verify its accuracy.

2. The method for obtaining the gas-water relative permeability curve of tight sandstone based on the DAE-DNN model according to claim 1 is characterized in that: The specific process of generating multiple sets of relative permeability curves by the Brooks-Corey model in step 1 is as follows: The Brooks-Corey model is expressed as: Where: k ro , k rw , k rg —relative permeabilities of oil phase, water phase and gas phase, respectively; k ro ,max,k rw ,max,k rg ,max—the maximum relative permeability of oil phase, water phase and gas phase respectively; s o 、s w 、s g — respectively, the saturation of oil phase, water phase and gas phase; s or 、s wc 、s gc —respectively residual oil saturation, irreducible water saturation, and residual gas saturation; n o 、n w 、n g —Shape factor, which determines the shape of the relative permeability curve and has a value range of 1-6; According to the above Brooks-Corey model, the irreducible water saturation s is selected. wc , residual gas saturation s gc , shape factor n w 、n g is a variable parameter, and the other parameters are fixed parameters, and multiple sets of relative permeability curves are generated.

3. The method for obtaining the gas-water relative permeability curve of tight sandstone based on the DAE-DNN model according to claim 1 is characterized in that: In step 2, gas-water two-phase capacity simulation is performed according to the relative permeability curve, specifically including: Step 21: Simulation of fracture network formed by hydraulic fracturing; Step 22: gas-water two-phase flow simulation; Step 23: The gas reservoir numerical model obtains three sets of production data: water production FWPR, gas production FGPR, and pressure FPR.

4. The method for obtaining the gas-water relative permeability curve of tight sandstone based on the DAE-DNN model according to claim 3 is characterized in that: In the simulation of the fracture network formed by hydraulic fracturing in step 21, the fracturing scheme design is simulated; in the simulation of the gas-water two-phase flow in step 22, the critical flow saturation and the starting pressure gradient are set for simulation.

5. The method for obtaining the gas-water relative permeability curve of tight sandstone based on the DAE-DNN model according to claim 1, characterized in that: The sample set construction in step 3 specifically includes: using production data as input layer samples, where the production data includes water production, gas production, and pressure; using bound water saturation s wc , residual gas saturation s gc , shape factor n w 、n g Constitute the output layer samples.

6. The method for obtaining the gas-water relative permeability curve of tight sandstone based on the DAE-DNN model according to claim 1, characterized in that: The step 4 specifically includes: Step 41: Data cleaning and processing; Step 42: The encoder part of the deep autoencoder DAE reduces the dimension of the input sample; the decoder part of the DAE reconstructs the reduced dimension data, and outputs the reduced dimension data when the root mean square error RMSE meets the error requirement; Step 43: The reduced-dimensional input samples are passed through DNN models with different structures to obtain relative permeability curves.

7. The method for obtaining the gas-water relative permeability curve of tight sandstone based on the DAE-DNN model according to claim 6, characterized in that: In step 41, the data is cleaned and processed, specifically: invalid values ​​and abnormal values ​​in the data are removed, missing data are supplemented, and the integrity and validity of the data during the entire production period are ensured; and the Min-max standardization method is used to process the data, so that the values ​​corresponding to all factors are uniformly mapped to the [0, 1] interval.

8. The method for obtaining the gas-water relative permeability curve of tight sandstone based on the DAE-DNN model according to claim 7, characterized in that: In step 42, the deep autoencoder DAE consists of a 5-layer structure, including 1 input layer, 1 output layer, and 3 hidden layers; when the root mean square error RMSE between the reconstructed data and the original input data is less than 0.01, the error requirement is met.

9. The method for obtaining the gas-water relative permeability curve of tight sandstone based on the DAE-DNN model according to claim 6, characterized in that: In step 43, the construction of the DNN model is specifically as follows: Step 431: training algorithm determination; There are three deep neural network training algorithms: Levenberg-Maquardt, Scaled Conjugate Gradient, and Bayesian Regularization. When the hidden layer settings are consistent, the above three training algorithms are used to train the model. The maximum number of training rounds is 1000, and the training algorithm with the smallest mean square error is selected as the training algorithm. Step 432: forming a prediction model set by setting different hidden layers; Specifically, through the selected training algorithm, multiple groups of different numbers of hidden layers and nodes are set for model training. When the root mean square error RMSE is less than 0.01, it becomes a group of prediction models. Finally, 10 groups of models form a prediction model set.

10. The method for obtaining tight sandstone gas-water relative permeability curve based on DAE-DNN model according to claim 1, characterized in that: The step 5 is specifically as follows: The prediction model set normalizes 10 phase permeability curves, and obtains the numerical model water production FWPR, gas production FGPR, and pressure FPR curves. The model is verified by the mean square error between the numerical model production curve and the actual production curve.

Images