Machine learning-based credibility-containing formation pore pressure while-drilling updating method

Abstract

The invention belongs to the field of oil and gas drilling, and particularly relates to a credibility-containing formation pore pressure while-drilling updating method based on machine learning. According to the method, the credibility-containing formation pore pressure of the target well is updated on the basis of the MLP-BiGRU-MC machine learning model, and the drilling well seismic characteristics are enhanced on the basis of the well seismic fusion data so as to predict the formation pore pressure more accurately, so that the characteristics of multi-source data can be extracted; the relevance between the formation information can be processed, and the formation pore pressure containing credibility can be predicted; in addition, while-drilling updating can be carried out on the stratum pore pressure which is predicted before drilling and contains credibility through while-drilling data, the prediction accuracy and precision of the stratum pressure are expected to be further improved, then the reasonable drilling fluid safety density is determined, and therefore it is guaranteed that the drilling process is safe and efficient.

Description

Technical Field

[0001] The present invention belongs to the field of oil and gas drilling, and specifically relates to a method for updating formation pore pressure while drilling with credibility based on machine learning. Background Art

[0002] Due to the complex and changeable geological conditions of deep formations and the lack of understanding of the formation to be drilled, it is difficult to accurately predict and monitor the formation pressure for internal complex geology. For the existing single-value formation pressure prediction method, it is difficult to determine the empirical coefficient, it is difficult to accurately obtain the model parameters, and there are large errors in the prediction results.

[0003] In recent years, with the development of data-driven methods such as machine learning and neural networks, domestic and foreign scholars have introduced these methods into the field of formation pressure prediction. Traditional methods such as the Dc index method and the Sigma method have problems of regional limitations and cannot accurately predict the formation pore pressure. The method for updating formation pore pressure while drilling has also changed from the initial physical model based on empirical equations to an intelligent model based on data-driven. Chinese Patent CN118134022A provides a method for predicting formation pressure based on an improved machine learning algorithm, which improves the accuracy of formation pressure prediction based on the PSO-LSTM model. However, when using the LSTM model for pressure prediction, not only does it take a long time, is prone to overfitting, but also requires a lot of high-precision data. Summary of the Invention

[0004] In order to solve the above deficiencies of the prior art, the present invention provides a method for updating formation pore pressure while drilling with credibility based on machine learning. Creatively, based on the machine learning model of MLP-BiGRU-MC, the formation pore pressure with credibility of the target well is updated, and the present invention also innovatively proposes to enhance the seismic characteristics of the well being drilled based on well-seismic fusion data to more accurately predict the formation pore pressure. The MLP-BiGRU-MC machine learning model constructed by the present invention can not only extract the features of multi-source data, but also process the correlation between formation information to predict the formation pore pressure with credibility. In addition, the present invention can also use the data while drilling to update the formation pore pressure with credibility predicted before drilling, which is expected to further improve the prediction accuracy and precision of the formation pressure, and then determine a reasonable drilling fluid safety density, so as to ensure the safety and efficiency of the drilling process.

[0005] The present invention provides a method for updating formation pore pressure while drilling with credibility based on machine learning. The technical implementation route is as follows: S1. Collect characteristic data and perform preprocessing to obtain drilling element data, and perform normalization processing; S2. Perform a correlation analysis on the normalized drilling element data, and screen to obtain the input parameters of the MLP-BiGRU-MC model; S3. Build and analyze the MLP-BiGRU-MC model to obtain the single-valued formation pore pressure updated while drilling; S4. Perform an uncertainty analysis on the single-valued formation pore pressure updated while drilling.

[0006] Among them, in step S1, the characteristic data includes well logging data, mud logging data, seismic data, and well depth data of adjacent wells.

[0007] The preprocessing steps include: S101. Perform smoothing processing on the well logging data, mud logging data, and seismic data, and perform depth adjustment processing on the well depth data to obtain drilling element data; The smoothing processing methods include the unascertained filtering method and the five-point bell-shaped function smoothing method.

[0008] The well logging data, mud logging data, and seismic data are all affected by many factors, resulting in null values and outliers, and serious burr phenomena in the curves. In order to eliminate this interference and retain the useful data representing the formation properties normally, the unascertained filtering method and the five-point bell-shaped function smoothing method are used to process the above data.

[0009] Among them, the five-point bell-shaped function smoothing method is: (1) In the formula, is the value after smoothing at the current sampling point i, is the value before smoothing at the current sampling point i, , , , are the values before smoothing at the (i - 2)-th, (i - 1)-th, (i + 1)-th, and (i + 2)-th sampling points respectively; The unascertained filtering method is: (2) (3) In the formula, A is the characteristic data of the current sampling point after processing; is the characteristic data at the current sampling point, is the credibility distribution function of the characteristic data, n is the number of samples, represents the number of contained in the neighborhood { }, where , and are parameters related to the neighborhood range.

[0010] Use the mathematical expectation of the processed characteristic data A of the current sampling point to replace A to achieve outlier rejection.

[0011] Considering the prediction of the formation pore pressure of the well to be drilled using adjacent wells, the thickness and burial depth of the same formation in two adjacent wells are not necessarily the same. Therefore, depth adjustment technology is used for correction. The depth adjustment formula is as follows: (4) where is the depth of the position to be determined for the well to be drilled, in m; is the corresponding adjacent well depth, in m; is the depth of the well to be drilled, in m; is the corresponding adjacent well depth, in m; is the corresponding adjacent well depth, in m.

[0012] S102. Normalize the drilling element data processed in S101.

[0013] Considering that the dimensions of the drilling element data processed in S101 are different, normalize it. The formula used is: (5) where is the drilling element data after normalization, is the original data of the drilling element, are the minimum and maximum values of the original drilling element data set respectively.

[0014] In step S2, use the Pearson correlation coefficient to perform a correlation analysis on the normalized drilling element data. The calculation formula of the Pearson correlation coefficient is: (6) In the formula, is the Pearson correlation coefficient, is the value of the characteristic data at the current sampling point , represents the mean value of the characteristic data, is the value of the formation pore pressure at the current sampling point , represents the mean value of the formation pore pressure, is the number of samples.

[0015] Among them, the Pearson correlation coefficient The value range of is from -1 to 1. When this value is closer to 1, it indicates a stronger positive correlation between the two parameters; when it is closer to -1, it indicates a stronger negative correlation between the two parameters; when it is closer to 0, it indicates a weaker correlation between the two parameters.

[0016] Furthermore, determine according to actual requirements the value range, and screen the corresponding normalized drilling element data as the input parameters of the MLP-BiGRU-MC model.

[0017] In step S3, the process of building and analyzing the MLP-BiGRU-MC model includes: S301. Input the input parameters screened in step S2 into the MLP model to construct pre-drilling well-seismic fusion data; S302. Input the pre-drilling well-seismic fusion data into the BiGRU model for training, and output the predicted value of the single-value formation pore pressure; S303. Use the MC model to update the predicted value of the single-value formation pore pressure to obtain the updated single-value formation pore pressure while drilling.

[0018] After the target well obtains real-time data during the drilling process, it is imported into the MLP-BiGRU-MC model again to correct the well-seismic fusion data in real time, and update the predicted value of the single-value formation pore pressure derived from the MLP-BiGRU-MC model while drilling.

[0019] The MLP consists of multiple layers of neurons, and each neuron is connected to the neurons in the previous layer and the next layer through weights. The MLP usually consists of three parts: an input layer for receiving external data; a hidden layer that performs a non-linear transformation on the input through an activation function to extract the features of the data; and an output layer that outputs the result according to requirements.

[0020] The pre-drilling well-seismic fusion data F(x) is: (7) where x represents the input layer, and the output of the hidden layer can be expressed as , , represents the weight of the MLP model; , represents the bias of the MLP model, and the functions f and are functions.

[0021] The BiGRU model has 2 GRU layers and 4 control units; among them, the control units include an update gate and a reset gate.

[0022] The update gate is responsible for adjusting the fusion ratio of the information of the previous state and the information of the current moment, and its calculation formula is as follows: (8) In the formula, is the output of the update gate, is the activation function, is the input at the current moment (i.e., the pre-drilling well-seismic fusion data obtained by MLP processing), is the hidden state of the previous state, is the weight of the update gate, is the bias of the update gate.

[0023] The reset gate determines the correlation between the hidden state at the current moment and the hidden state at the previous moment, and selects how much to retain, discarding the rest. If the output value of the reset gate is close to 0, it means that the hidden state at the current moment will be more dependent on the current input information, so the input of the previous unit is more likely to be completely discarded; if it is close to 1, it means that the hidden state at the current moment is more dependent on the hidden state at the previous moment, so the input of the previous unit is more likely to be completely retained. Its calculation formula is as follows: (9) In the formula, is the output of the reset gate, is the weight of the reset gate, is the bias of the reset gate.

[0024] The essence of the Markov chain (MC) is that when a random process is given the current state and all past states, the conditional probability distribution of its future state depends only on the current state. Given the current state, the random process is conditionally independent of the past states. The MC model satisfies the following conditions: (10) In the formula, is the state at time t, is the transition probability between two states.

[0025] In the Markov chain, the probability of state transition at a certain moment depends only on its previous state. By finding the transition probability between any two states in the system, the Markov chain model can be obtained, which makes the Markov chain applicable to the case of drilling in the same formation, and the optimal exploration range and the transition probability of each formation can be judged through multiple trainings according to the lithology and physical properties of each formation. During the drilling process, the MLP-BiGRU-MC model can be corrected by inputting real-time drilling data to obtain appropriate probabilities for updating the formation pore pressure while drilling.

[0026] In step S4, the uncertainty analysis refers to expanding the single-valued formation pore pressure curve updated while drilling into a probability region according to the principles of sequence stratigraphy.

[0027] Specifically, according to the principles of sequence stratigraphy, similar stratigraphic sedimentation ages have the same physical responses. The smaller the spacing between measurement points, the greater the similarity of formation pressures. Based on the above idea, the formation pressure data between adjacent measurement points within the same group of strata can be used as a set of measurement samples. The cumulative probability distribution of formation pressure at each measurement point can be determined, and then the formation pressures with the same cumulative probability distribution are connected into a line. Thus, the formation pressure curve of the entire well section under a specific cumulative distribution probability can be obtained, and the formation pore pressure of the entire well section with credibility in the vertical direction can be obtained. In this way, a single-valued pressure prediction curve is transformed into a region composed of different cumulative probability distribution curves, and the region formed by the selected probability interval can be used to increase the credibility of the pressure prediction result.

[0028] The method for updating formation pore pressure with credibility during drilling based on machine learning is evaluated by the coefficient of determination ( ).

[0029] Further, the is: (11) where n is the number of samples, is the true value, is the predicted value, is the average value.

[0030] is used to reflect the fitting degree of the model, ranging from 0 to 1. The closer its value is to 1, the better the fitting effect of the two curves, that is, the closer the prediction result is to the true value.

[0031] Further, the RMSE is: (12) where n is the number of samples, is the true value, is the predicted value.

[0032] The root mean square error is very sensitive to extremely large or extremely small errors in a set of measurements. Therefore, the root mean square error can well reflect the precision of the measurement.

[0033] Traditional recurrent neural network models such as GRU are unidirectional neural networks that can only predict the results of the next moment based on the forward time series relationship. However, due to the influence of sedimentation on the formation all year round, the composition of the upper and lower surrounding rocks has a certain correlation. Combining forward and backward information is beneficial for the model to learn more time series features. Therefore, the improved structure of BiGRU is adopted in the present invention. The BiGRU model is a recurrent neural network composed of two independent GRU units. One processes data forward according to the time series, and the other processes data backward according to the time series. Through this bidirectional structure, the BiGRU model can capture both the forward and backward information of the sequence data at the same time. The final result is the splicing of the results in two directions, which can reduce the computational overhead and training time while ensuring the prediction accuracy through efficient calculation and the ability to model time series. Especially in the tasks of real-time data prediction and large-scale data processing during the drilling process, BiGRU has more advantages than LSTM, thus obtaining better prediction performance and improving the prediction accuracy of the model.

[0034] Advantageous effects: 1. The present invention updates the formation pore pressure with credibility for the target well for the first time based on the machine learning model of MLP-BiGRU-MC, and proposes to enhance the seismic characteristics of the well being drilled based on well-seismic fusion data to more accurately predict the formation pore pressure.

[0035] 2. The MLP-BiGRU-MC machine learning model constructed by the present invention can not only extract the features of multi-source data, but also process the correlation between formation information, and then predict the formation pore pressure with credibility.

[0036] 3. The present invention can also use the data while drilling to update the formation pore pressure with credibility predicted before drilling, further improving the prediction accuracy and precision of the formation pressure, and then determining the reasonable safety density of the drilling fluid, thus well guaranteeing the safety and efficiency of the drilling process. Description of the drawings

[0037] Figure 1 It is a flow chart of the method for updating the formation pore pressure with credibility based on machine learning; Figure 2 It is a comparison chart before and after smoothing the density, acoustic travel time, and natural gamma data; Figure 3 It is a heat map of Pearson correlation coefficients; Figure 4 It is a comparison chart of the formation pore pressure predicted by different models; Figure 5 It is a comparison chart of the predicted values of the formation pore pressure before and after the update of the MC model; Figure 6It is a formation pore pressure map with credibility. Detailed implementation manners

[0038] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying 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 based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0039] Taking a well in the Ledong block as an example, this area is characterized by high temperature and high pressure in deep formations and complex formation conditions. The temperature can reach 174.3 °C at a well depth of 3,982 meters, and the pressure coefficient is 2.17. Traditional methods cannot normally estimate the formation pressure, and there is a problem of incomplete interpretation data. Therefore, it is necessary to correct the pre-drilling prediction results of formation pore pressure based on the data while drilling to improve its credibility.

[0040] A method for correcting formation pressure with credibility based on the MLP-BiGRU-MC method disclosed in this embodiment, as Figure 1 shown, includes the following steps: S1. Collect characteristic data, preprocess it to obtain drilling element data, and perform normalization processing: S101. Use the five-point bell-shaped function smoothing method and the unascertained filtering method to smooth the logging, mud logging, and seismic data, and perform depth adjustment on the well depth data to obtain drilling element data.

[0041] As Figure 2 shown is the comparison chart of density, acoustic travel time, and natural gamma data before and after smoothing. The abscissa is the well depth, and the ordinates are density (den), acoustic travel time (dtco), and natural gamma (GR) respectively. Among them, the blue is the original data, and the orange is the processed curve.

[0042] S102. Perform normalization processing on the drilling element data obtained after the processing of S101.

[0043] S2. Perform correlation analysis on the normalized drilling element data, and select and screen them as the input parameters of the MLP-BiGRU-MC model. The final result is as Figure 3 shown in the Pearson correlation coefficient heat map.

[0044] The selected characteristic data include: depth, drilling rate in mud logging data, hook load, rotation speed, torque, natural gamma in logging data, acoustic travel time, and formation velocity, low-frequency component, phase cosine value, root mean square amplitude, and sweet spot attribute in seismic data.

[0045] S3. Build the MLP-BiGRU-MC model for analysis to obtain the updated in-situ formation pore pressure value while drilling: First, use MLP to construct the pre-drilling well-seismic fusion data for the input parameters obtained in step S2, then import it into the BiGRU model for training to output the predicted value of the in-situ formation pore pressure, and finally use MC to update the predicted value of the in-situ formation pore pressure. The specific steps for using MC to update the predicted value of the in-situ formation pore pressure are as follows: When real-time data is obtained during the drilling of the target well, import it into the MLP model again to correct the well-seismic fusion data in real time, and update the predicted value of the in-situ formation pore pressure exported by the BiGRU model while drilling.

[0046] As Figure 5 shown in the comparison chart of the predicted values of the formation pore pressure before and after the update of the MC model, where the pre-drilling prediction pp is the predicted value of the in-situ formation pore pressure without being updated by the MC model, and the corrected pp is the predicted value of the in-situ formation pore pressure updated by the MC model. Figure 5 Use the in-drilling data of the first 3540 meters of well depth to update the predicted formation pore pressure value of the last 60 meters of well depth (i.e., the well depth from 3540 to 3600 meters) while drilling.

[0047] S4. Conduct uncertainty analysis on the updated in-situ formation pore pressure value while drilling Take the formation pressure data between adjacent measuring points within the same formation as a set of measurement samples, determine the cumulative probability distribution of the formation pressure at each measuring point, and then connect the formation pressures with the same cumulative probability distribution into a line to obtain the formation pressure curve of the entire well section under a specific cumulative distribution probability, so as to obtain the formation pore pressure with credibility in the vertical direction of the entire well section. The obtained results are as Figure 6 shown.

[0048] S5. Evaluate the prediction results of the model Select the coefficient of determination ( ) and the root mean square error (RMSE) to evaluate the MLP-BiGRU-MC, CNN-LSTM, and GRU models respectively. The results are as Figure 4 and Table 1 shown (where the input parameters of the CNN-LSTM and GRU models are the same as those of the MLP-BiGRU-MC): Table 1 Evaluation results of each model

[0049] It can be seen from Table 1 that compared with the traditional CNN-LSTM model, the two GRU models using the gate control unit have higher fitting degrees and more accurate prediction results, and the training time of the model is shorter. The is 0.82, and the RMSE is 0.83%, indicating that it is more accurate and applicable in predicting and correcting formation pore pressure.

[0050] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, rather than limitations on the specific implementation manners of the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the claims of the present invention shall be included within the protection scope of the claims of the present invention.

Claims

1. A machine learning-based method for updating pore pressure of formations with credibility while drilling, characterized in that The specific steps include: S1. Collect characteristic data, preprocess it to obtain drilling element data, and perform normalization processing; S2. Perform correlation analysis on the normalized drilling element data, and screen to obtain the input parameters of the MLP-BiGRU-MC model; S3. Build and analyze the MLP-BiGRU-MC model to obtain the in-situ single-value formation pore pressure updated while drilling; S4. Perform uncertainty analysis on the in-situ single-value formation pore pressure updated while drilling; Among them, step S3 specifically includes: S301. Input the input parameters screened in step S2 into the MLP model to construct pre-drilling well-seismic fusion data; S302. Input the pre-drilling well-seismic fusion data into the BiGRU model for training, and output the predicted value of the single-value formation pore pressure; S303. Use the MC model to update the predicted value of the single-value formation pore pressure to obtain the in-situ single-value formation pore pressure updated while drilling.

2. The method for updating formation pore pressure while drilling with credibility based on machine learning according to claim 1, characterized in that In step S1, the characteristic data includes logging data, mud logging data, seismic data, and well depth data of adjacent wells; the preprocessing steps include: S101. Smooth the logging data, mud logging data, and seismic data, and perform depth adjustment on the well depth data to obtain drilling element data; S102. Perform normalization processing on the drilling element data processed in S101.

3. The method for in-situ updating of formation pore pressure with credibility based on machine learning according to claim 1, characterized in that in step S2, the Pearson correlation coefficient is used to perform correlation analysis on the normalized drilling element data, and the calculation formula of the Pearson correlation coefficient is: （6） In the formula, is the Pearson correlation coefficient, is the drilling element data after normalization processing, represents the mean value of the drilling element data after normalization processing, is the current sampling point where the value of the formation pore pressure is located, represents the mean value of the formation pore pressure, is the number of samples; According to Based on the value range, the corresponding normalized drilling element data is screened as the input parameter of the MLP-BiGRU-MC model.

4. The method for in-situ updating of formation pore pressure with credibility based on machine learning according to claim 1, characterized in that the pre-drilling well-seismic fusion data is: （7） Where F(x) is the pre-drilling well-seismic fusion data, represents the input layer, and the output of the hidden layer is expressed as , , represents the weights of the MLP model; , represents the bias of the MLP model, and the functions and the function are functions.

5. The method for in-situ updating of formation pore pressure with credibility based on machine learning according to claim 1, characterized in that the BiGRU model has 2 GRU layers and 4 control units; among them, the control unit includes an update gate and a reset gate; The calculation formula of the update gate is as follows: （8） wherein, is the output of the update gate, is the activation function, is the input at the current moment, is the hidden state of the previous state, is the weight of the update gate, is the bias of the update gate; The calculation formula of the reset gate is as follows: （9） Wherein, is the output of the reset gate, is the weight of the reset gate, is the bias of the reset gate.

6. The method for in-situ updating of formation pore pressure with credibility based on machine learning according to claim 1, characterized in that the MC model satisfies the following conditions: （10） Wherein, is the state at time t, is the transition probability between two states.

7. The method for in-situ updating of formation pore pressure with credibility based on machine learning according to claim 2, characterized in that the smoothing processing methods include the unascertained filtering method and the five-point bell-shaped function smoothing method; Among them, the five-point bell-shaped function smoothing method （1） Wherein, is the value after smoothing of the current sampling point i, is the value before smoothing of the current sampling point i, , , , are respectively the values before smoothing of the (i - 2)th, (i - 1)th, (i + 1)th, and (i + 2)th sampling points; The unascertained filtering method uses the mathematical expectation of the feature data A of the processed current sampling point to replace A and achieve the elimination of outliers: （2） （3） Where A is the feature data of the current sampling point after processing, is the feature data of the current sampling point, is the credibility distribution function of the feature data, is the sampling point, n is the number of samples, denotes the neighborhood { } contains the number of, where, , and are parameters related to the neighborhood range; The depth adjustment formula is: （4） wherein, is the depth of the to-be-drilled well at the to-be-sought position, m; is the corresponding offset well depth, m; is the depth of the to-be-drilled well, m; is the corresponding offset well depth, m; is the corresponding offset well depth, m.

8. The method for in-situ updating of formation pore pressure with credibility based on machine learning according to claim 2, characterized in that the normalization processing formula is: （5） Among them, is the drilling element data after normalization processing, is the original data of the drilling element, are the minimum and maximum values of the original data set of the drilling element, respectively.

9. The method for in-situ updating of formation pore pressure with credibility based on machine learning according to claim 1, characterized in that The machine learning-based formation pore pressure update method while drilling with credibility is evaluated by and RMSE; The said is as follows: （11） where n is the number of samples, is the true value, is the predicted value, is the average value; the RMSE is: （12） where n is the number of samples, is the true value, is the predicted value.

Images