A formation pore pressure logging prediction method based on LSTM-PINN method

Abstract

The present invention relates to a formation pore pressure logging prediction method based on LSTM-PI NN method, comprising the following steps: collecting geological, logging, and formation test related data in the study area; screening logging data points in the study area and establishing a database; constructing a neural network prediction model using LSTM-PI NN method; training the neural network prediction model based on the database to obtain a trained neural network prediction model; and predicting formation pore pressure using the trained neural network prediction model. This application is based on the collaborative drive of physics and data, and the predicted formation pore pressure can meet the basic theoretical laws, and the prediction effect is more accurate, which is of great significance to oil exploration and development, underground engineering construction, etc.

Description

Technical Field

[0001] The present invention relates to the technical field of formation pore pressure prediction, and in particular to a formation pore pressure logging prediction method based on LSTM-PINN method. Background Art

[0002] The prediction of formation pore pressure has important guiding significance for petroleum exploration and development, underground engineering construction, etc., especially its importance in the oil and gas drilling process is gradually emerging. The theory and method of accurately predicting formation pore pressure has become an important research topic for geological exploration and development personnel at home and abroad.

[0003] Although the measured formation pore pressure can be obtained relatively accurately in actual formation testing, it is difficult to obtain and has high economic cost, so it is difficult to be widely used.

[0004] At present, the methods for predicting formation pressure mainly include geology, geophysics, engineering geology and mathematical modeling. Geophysical logging is one of the important means of obtaining underground information. The data is easy to obtain and has high vertical resolution, which can effectively reflect the characteristics of each well. The more mature and commonly used traditional methods for predicting formation pore pressure using logging data are the equilibrium depth method, Bowers method and Eaton index method. Recently, the prediction of formation pore pressure based on machine learning algorithm is also a more effective prediction method.

[0005] Among them, for unloading formations, the Bowers method establishes the loading curve equation and the unloading curve equation between the vertical effective stress and the acoustic time difference based on the vertical stress being equal to the sum of the vertical effective stress and the formation pore pressure, and then determines the formation pore pressure. The Bowers method based on the effective stress theory can relatively accurately predict the formation pore pressure of unloading formations, but only uses the acoustic time difference logging curve, which is greatly affected by the quality of the logging curve. Although the traditional machine learning prediction method has a certain degree of accuracy, it has problems such as poor interpretability of the results and the need for a large amount of data. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a formation pore pressure logging prediction method based on the LSTM-PINN method.

[0007] The present invention is achieved through the following technical solutions:

[0008] A formation pore pressure logging prediction method based on LSTM-PINN method includes the following steps:

[0009] Collect relevant data on geology, well logging, and formation testing in the study area;

[0010] Screen the well logging data points in the study area and establish a database;

[0011] Based on Python language, build a neural network prediction model based on LSTM-PINN method;

[0012] Training the neural network prediction model based on the database to obtain a trained neural network prediction model;

[0013] The trained neural network prediction model is used to predict the formation pore pressure.

[0014] Optionally, the method for screening the well logging data points in the study area includes: selecting data points of mudstone layers with an effective thickness greater than 2 m.

[0015] Furthermore, the method for screening the well logging data points in the study area includes: selecting data points of the mudstone layer section with a diameter expansion rate less than 15%.

[0016] Preferably, six logging curves, namely, natural gamma, natural potential, acoustic time difference, compensated density, neutron, and deep resistivity, are selected to establish a database.

[0017] Preferably, the input parameters of the neural network prediction model include depth DEPTH, longitudinal wave velocity V p , compensation density DEN, vertical stress S v , neutron CNL, natural gamma GR.

[0018] A formation pore pressure logging prediction method based on the LSTM-PINN method is described, characterized in that the first physical loss function indicates that the prediction result needs to meet the judgment condition that the vertical stress is greater than the effective stress;

[0019] The second physical function indicates that the longitudinal wave velocity V in the input parameter must be satisfied. p When it is equal to the set value, the prediction result of the model is equal to 0;

[0020] The third physical loss requires judging the value of the partial derivative. If the partial derivative value calculated by the model is closer to the partial derivative value of the physical formula in the Bowers model, it means that the network model has learned this physical knowledge.

[0021] Optionally, the second physical function indicates that when the longitudinal wave velocity Vp in the input parameter is equal to 1524 m / s, the prediction result of the model is equal to 0.

[0022] Preferably, the learning rate of the neural network prediction model is 0.001, the Epoch is 15000, the hidden_size is 24, and the Num_layers is 2.

[0023] Compared with the prior art, this application has at least the following beneficial effects:

[0024] This application is based on the collaborative drive of physics and data. The predicted formation pore pressure can meet the basic theoretical laws, and the prediction effect is more accurate, which is of great significance to oil exploration and development, underground engineering construction, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0026] Figure 1 It is the loading curve fitting diagram of the normal compaction section in the study area in the embodiment;

[0027] Figure 2 It is a Bowers method unloading curve fitting diagram based on the unloading point in the embodiment;

[0028] Figure 3 A comparison diagram of formation pore pressure prediction results based on the Bowers method in the embodiment;

[0029] Figure 4 It is a correlation analysis diagram between effective stress and various parameters in the embodiment;

[0030] Figure 5 It is a PINN framework diagram in the embodiment;

[0031] Figure 6 This is a diagram of the model training process and results in the embodiment;

[0032] Figure 7 A comparison diagram of the model formation pore pressure prediction results in the embodiment;

[0033] Figure 8 This is a diagram showing the prediction results of the formation pore pressure in the study area in the example. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0035] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the various embodiments can be referred to each other.

[0036] The present embodiment discloses a formation pore pressure logging prediction method based on the LSTM-PINN method, comprising the following steps:

[0037] S1. Collect relevant data on geology, well logging, and formation testing in the study area.

[0038] The geological data include logging lithology profiles, stratigraphic divisions and other data; the logging data include conventional logging data such as natural gamma, compensated sonic waves, compensated density, neutrons, deep and shallow resistivity, etc.; the formation testing data are the actual formation pore pressure obtained by static pressure testing.

[0039] S2. Filter the data points in the study area and establish a database.

[0040] Mudstone formations are more stable than other lithological formations. Therefore, the prediction of formation pore pressure usually requires the screening of logging data points. This embodiment selects mudstone sections with an effective thickness greater than 2m based on the collected logging lithology profiles and logging data. Since the diameter expansion will affect the quality of the logging curve, the diameter expansion rate k is calculated based on the drill bit diameter data and the well diameter logging curve, and finally the data points of the layer with an expansion rate less than 15% are screened out to ensure the good quality of the logging curve. At the same time, the formation pore pressure obtained by the formation test is matched with the logging curve to obtain the logging value of the corresponding depth, including natural gamma GR, natural potential SP, acoustic time difference AC, neutron CNL, and deep resistivity RT.

[0041] Among them, the expansion rate k is calculated using the following formula:

[0042]

[0043] In the above formula, CAL is the well diameter in the logging curve, in; BS is the drill bit diameter in the drilling data, in.

[0044] S3. Calculate vertical stress S v and the longitudinal wave velocity V p .

[0045] Among them, based on the compensated density logging curve, the vertical stress can be obtained, as shown in formula (2):

[0046]

[0047] In the above formula, S vis the vertical stress, MPa; H0 is the starting depth of logging, m; ρ0 is the formation density without logging information, g / cm 3 ; DEN is the logging compensation density, g / cm 3 ; g is the acceleration due to gravity, which is 9.8m / s 2 .

[0048] Among them, the longitudinal wave velocity V p It is calculated from the acoustic time difference in the logging curve, as shown in formula (3):

[0049]

[0050] In the above formula, V p is the longitudinal wave velocity, m / s; AC is the logging acoustic wave time difference, μs / ft;

[0051] S4. The traditional Bowers method was used to study the formation pore pressure in the study area.

[0052] Since the loading curve is consistent with the normal compaction trend, the normal pressure section is selected for fitting, and the loading curve fitting formula is obtained as shown in formula (4). The loading curve fitting result is shown in Figure 1 As shown;

[0053] V p =1524+274.52σ m 0.6478 (4)

[0054] In the above formula, σ m is the effective stress, MPa.

[0055] Based on the acoustic velocity and overlying formation pressure corresponding to the measured formation pressure point, the plasticity coefficient in the unloading curve can be inverted, and finally the unloading curve fitting formula is obtained, as shown in formula (5); the unloading curve fitting result is shown in Figure 2 As shown;

[0056]

[0057] In the above formula, σ m is the effective stress, MPa; σ max is the effective stress corresponding to the beginning of formation unloading.

[0058] Based on the loading curve and unloading curve, the formation pore pressure P in the study area is calculated using formula (6): p , the results are as follows Figure 3 As shown:

[0059] P p =S v -σ m (6)

[0060] In the above formula, P p is the formation pore pressure, MPa; S v is the vertical stress, MPa, σ m is the effective stress, MPa;

[0061] Based on the measured value of formation pore pressure, the mean relative error MRE is calculated using formula (7):

[0062]

[0063] In the above formula, y i is the measured value of formation pore pressure; is the predicted value of formation pore pressure; n is the number of tests.

[0064] In this embodiment, the calculated result MRE is 12.42%. It can be seen that the prediction effect is not ideal.

[0065] S5. Screen appropriate input parameters for neural network prediction.

[0066] The statistically measured effective stress and related parameters were analyzed for correlation. The results are as follows: Figure 4 As shown. It can be seen that the effective stress has a good correlation with depth, acoustic wave time difference (velocity), neutron, natural gamma, and vertical stress. According to Bowers theory, the change of effective stress in unloading formations will have a special response in acoustic wave time difference (velocity) and density. Therefore, after comprehensive analysis, this embodiment selects depth DEPTH, longitudinal wave velocity V p , compensation density DEN, vertical stress S v , neutron CNL, natural gamma GR and other six parameters are used as input parameters of the neural network.

[0067] S6. In order to improve the prediction accuracy, the LSTM-PINN method was used to construct a neural network prediction model and predict the formation pore pressure in the study area.

[0068] Specifically, we use Python language to build an LSTM neural network model. Combined with the PINN theory, we improve the LOSS loss function based on the LSTM model and introduce three physical loss functions to achieve the effect of embedding physical knowledge.

[0069] The physics-embedded neural network (PINN) can embed physical prior knowledge into machine learning models and algorithms, which can improve learning efficiency and ability, and also improve model interpretability. After adding physical knowledge, the amount of data required will be reduced compared to pure data-driven, effectively solving the problem of insufficient measured data. In order to realize the function of this model, the loss function module of the model is modified and three physical loss functions are introduced. The framework diagram of the physics-embedded neural network (PINN) is shown below: Figure 5 As shown, among the three physical loss functions from top to bottom, the first physical loss function indicates that the prediction result needs to meet the judgment condition that the vertical stress is greater than the effective stress; the second physical function indicates that when the longitudinal wave velocity Vp in the input parameter is equal to 1524m / s, the prediction result of the model is equal to 0; the third physical loss needs to judge the value of the partial derivative. If the partial derivative value calculated by the model is closer to the partial derivative value of the physical formula in the Bowers model, it means that the network model has learned this physical knowledge.

[0070] Since the logging curve has time series characteristics, the long short-term memory network (LSTM) can be well applied in this type of data. Therefore, combining LSTM with PI NN can improve the prediction effect. The hyperparameter values ​​are shown in Table 1:

[0071] Table 1: Model hyperparameter values

[0072]

[0073] Finally, the prediction effect is Figure 7 As shown, the average relative error can be calculated by formula (7). The calculation results show that the average relative error MRE of the training set is 5.4769%, and the average relative error MRE of the test set is 7.8687%. Compared with the traditional Bowers method, the neural network prediction model constructed in this embodiment predicts the formation pore pressure more accurately.

[0074] S7. After training the LSTM-PI NN model based on the previously established database, save the model and apply it to the well logging profile in the study area, such as Figure 8 It can be seen that the formation pore pressure predicted by the LSTM-PI NN method can meet the basic theoretical laws and has better prediction effect than the traditional Bowers prediction method.

[0075] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A formation pore pressure logging prediction method based on LSTM-PINN method, characterized in that: The following steps are involved: Collect relevant data on geology, well logging, and formation testing in the study area; The well logging data points in the study area were screened, and six logging curves, including natural gamma, natural potential, acoustic time difference, compensated density, neutron, and deep resistivity, were selected to establish a database; Based on Python language, build LSTM neural network model, introduce three physical loss functions based on PINN theory, and construct LSTM-PINN neural network prediction model; Training the LSTM-PINN neural network prediction model based on the database to obtain a trained neural network prediction model; Use the trained neural network prediction model to predict formation pore pressure; The input parameters of the neural network prediction model include depth DEPTH, longitudinal wave velocity V p , compensation density DEN, vertical stress S v , neutron CNL, natural gamma GR; The first physical loss function indicates that the prediction result needs to meet the judgment condition that the vertical stress is greater than the effective stress; The second physical loss function indicates that the longitudinal wave velocity V in the input parameter needs to be satisfied. p When it is equal to the set value, the prediction result of the model is equal to 0; The third physical loss function needs to judge the value of the partial derivative. If the partial derivative value calculated by the model is closer to the partial derivative value of the physical formula in the Bowers model, it means that the network model has learned this physical knowledge.

2. The formation pore pressure logging prediction method based on the LSTM-PINN method according to claim 1 is characterized in that: The method of screening the well logging data points in the study area includes: selecting the data points of the mudstone layer with an effective thickness greater than 2m.

3. A formation pore pressure logging prediction method based on LSTM-PINN method according to claim 1 or 2, characterized in that: The method of screening the well logging data points in the study area includes: selecting the data points of the mudstone layer section with a diameter expansion rate less than 15%.

4. The formation pore pressure logging prediction method based on LSTM-PINN method according to claim 1 is characterized in that: The calculation formula for vertical stress is: In the above formula, S v is the vertical stress, MPa; H0 is the starting depth of logging, m; ρ0 is the formation density without logging information, g / cm 3 ; DEN is the logging compensation density, g / cm 3 ; g is the acceleration due to gravity, which is 9.8m / s 2 .

5. The formation pore pressure logging prediction method based on LSTM-PINN method according to claim 1 is characterized in that: The second physical loss function indicates that when the longitudinal wave velocity Vp in the input parameter is equal to 1524m / s, the model's prediction result is equal to 0.

6. A formation pore pressure logging prediction method based on LSTM-PINN method according to claim 1, 4 or 5, characterized in that: The learning rate of the neural network prediction model is 0.001, the Epoch is 15000, the hidden_size is 24, and the Num_layers is 2.

7. The formation pore pressure logging prediction method based on LSTM-PINN method according to claim 3 is characterized in that: The expansion rate k is calculated using the following formula: In the above formula, CAL is the well diameter in the logging curve, in; BS is the drill bit diameter in the drilling data, in.

Images