Drilling rate prediction method based on mechanism-deep learning and hybrid model updating strategy

Abstract

The invention belongs to the technical field of oil and gas drilling, and particularly relates to a mechanical drilling speed prediction method based on mechanism-deep learning and a hybrid model updating strategy. The method comprises the steps that 1, a mechanical drilling speed prediction mechanism model considering PDC drill bit abrasion and the bottom hole temperature is established; 2, establishing a deep learning model architecture; 3, creating a hybrid mechanical drilling speed prediction model through the WT model and the CAG model; 4, establishing a model hybrid updating strategy combining regular full-amount updating and real-time incremental updating; 5, the drilling speed prediction model is updated according to the mixed model updating strategy for the drilling parameters obtained during drilling; and step 6. And according to the drilling parameters obtained at the current moment, inputting the updated mechanical drilling speed prediction model so as to successfully predict the drilling speed at the next moment. According to the method, transient changes of complex working conditions in the drilling process can be captured easily, the drilling period can be shortened easily, the drilling cost can be reduced, and intelligent regulation and control of the drilling process can be assisted.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of oil and gas drilling, and particularly relates to a mechanical drilling speed prediction method based on a mechanism-depth learning and a hybrid model updating strategy. BACKGROUND

[0002] With the continuous development of high-quality oil and gas resources, the exploitation of remaining resources is becoming more and more difficult and the cost is becoming higher and higher. Dry hot rock type geothermal energy has more development potential and prospect due to its universality of distribution, higher heat storage temperature and clean and low-carbon renewable nature. Dry hot rock is located at a depth of 3 to 10 km, with a temperature higher than 180°C and containing no or only a small amount of fluid, and has complex geological conditions: high temperature, high hardness, high stress, high density and high risk. The complex formation conditions result in high dry hot rock geothermal energy development cost, of which 50%-60% comes from drilling cost, the drilling speed prediction accuracy is limited, and a drilling speed prediction method for dry hot rock formation drilling is urgently needed.

[0003] There are many existing drilling speed prediction methods. From the structure of the model, it is generally divided into three types of prediction models based on mechanism (physics), data (deep learning) and hybrid (coupling of mechanism model and data model). The mechanism model follows the physical objective law and has rigorous theory, but the prediction accuracy is limited. The model based on deep learning can skip the understanding of the physical process, and when the data quality and quantity are sufficient, it can achieve relatively high prediction accuracy, but it has the problems of insufficient generalization and migration. The hybrid model can couple the mechanism model and the data model to achieve better prediction effect, but it is mostly a static prediction model. The complex drilling environment of dry hot rock formation has higher demand for dynamic updating of the model, and it is difficult to accurately predict the drilling speed during drilling. Some methods use a sliding window updating strategy for dynamic updating, although the real-time changes of the drilling conditions are considered, but as the drilling section deepens, accidental factors will accumulate, and then affect the drilling speed prediction accuracy of the next tripping or other adjacent blocks. A new drilling speed prediction model dynamic updating strategy needs to be researched to improve the robustness and accuracy of the drilling speed prediction model. SUMMARY

[0004] In view of the shortcomings of the above-mentioned existing drilling speed prediction technology, the application proposes a mechanical drilling speed prediction method based on mechanism-depth learning and hybrid model updating strategy suitable for dry hot rock formation. The method helps to capture the transient changes of complex conditions during drilling, avoids the deviation of the overall law caused by the accumulation of transient errors, helps to shorten the drilling cycle, reduce the drilling cost, and assist the intelligent control of the drilling process. A mechanical drilling speed prediction method with engineering practical value is formed.

[0005] The technical scheme adopted by the present application is: a mechanical drilling speed prediction method based on mechanism-deep learning and a hybrid model updating strategy, the mechanical drilling speed prediction method comprising the following steps: Step one, establishing a mechanical drilling speed prediction mechanism model (WT model) covering PDC bit wear and bottom hole temperature Aiming at the drilling characteristics of hot dry rock formation, a mechanical drilling speed prediction mechanism model WT covering bottom hole temperature and bit wear is proposed, as shown in formula 1.

[0006] (1) In the formula, W f is a function of PDC bit wear, as shown in formula 2; WOB is the drilling pressure, unit kN; RPM is the rotational speed, unit r / min; is the displacement of drilling fluid, unit L / s; T is the temperature of the drilled formation, unit °C; P is the bottom hole differential pressure, unit MPa; S is the compressive strength of the drilled rock, unit MPa; D d is the bit diameter, unit mm; K is the comprehensive coefficient of the model, a, b, c, d Both are the coefficients of the model parameters, dimensionless.

[0007] (2) In the formula, is the IADC bit dullness grade, dimensionless.

[0008] Collect historical data of completed wells, calculate the comprehensive coefficients in formula (1) from them, and thus obtain the mechanism model of drilling speed prediction.

[0009] Step two, establishing a deep learning model architecture CNN-GRU-Attention (CGA model) (1) Establish a one-dimensional convolutional neural network composed of an input layer, a convolutional layer, a pooling layer, a Flatten layer, a fully connected layer and an output layer.

[0010] (2) Establish a gated recurrent unit to automatically extract the corresponding drilling features.

[0011] (3) Build an attention mechanism to dynamically adjust the attention degree to the input data or time steps.

[0012] (4) Combine one-dimensional convolutional neural network, gated recurrent unit and attention mechanism to build a CNN-GRU-Attention deep learning model with decoder and encoder architecture for mechanical drilling speed prediction.

[0013] Step three, create a hybrid mechanical drilling speed prediction model (HRM) through the WT model and the CAG model (1) Calculate the preliminary drilling speed prediction value using the mechanism model fitted from the drilling data.

[0014] (2) Calculate the difference between the predicted value and the actual drilling speed value, i.e. the residual.

[0015] (3) Use drilling data and prediction residuals as input for deep learning model training to obtain the final drilling speed prediction value.

[0016] Step four, establish a model hybrid update strategy combining periodic full update and real-time incremental update (1) Use the initially established hybrid drilling speed prediction model as the steady-state mechanical drilling speed prediction model.

[0017] (2) When a single rod or stand is drilled, or the mechanical specific energy or mechanical drilling speed fluctuates continuously, (within 2m or 20 minutes, the fluctuation exceeds 40%, the specific adjustment needs to be made according to the block situation), collect drilling data and start real-time incremental update.

[0018] (3) Lock the fully connected layer of the hybrid model, only update the hidden layer with real-time drilling sequence data, and train only one round.

[0019] (4) When tripping in and out, (round trip: lifting the drill string from the wellbore is called tripping out, and lowering the drill into the wellbore is called tripping in; short trip: according to the needs of drilling technology, some drill tools are lifted from the well and then lowered into the well.) Collect all drilling data for the current tripping and randomize, and start periodic full update.

[0020] (5) Discard the last real-time incremental update drilling speed prediction model, update the hidden layer based on the last steady-state drilling speed prediction model, train one round, and use the updated drilling speed prediction model as the basis for the real-time incremental update model.

[0021] Step five, update the drilling speed prediction model according to the drilling parameters obtained during drilling according to the hybrid model update strategy.

[0022] Step six, input the updated mechanical drilling speed prediction model according to the drilling parameters obtained at the current time, thereby successfully predicting the drilling speed at the next time.

[0023] Further, the collected data of the completed well in the target block includes the rate of penetration, the drilling pressure, the rotation speed, the displacement, the bottom hole temperature, the bottom hole differential pressure, the compressive strength of the drilled rock, the drill bit diameter and the wear data of the drill bit.

[0024] Further, the making a connection time, the tripping and the abnormal fluctuation of the drilling speed are taken as the model updating conditions, and the drilling speed prediction model is updated on the basis of the stratum drilling through the hybrid model updating strategy.

[0025] Further, when the prediction accuracy or the MSE fluctuation is too large, the drilling speed prediction model is updated.

[0026] Further, when the tripping or the short tripping is performed, the transient drilling speed prediction model of each single or stand is abandoned, the data in the same is randomly shuffled on the basis of the previous steady drilling speed prediction model, a round of update is performed, and the steady drilling speed prediction model is taken as the basis model for the next single or stand, so that the online prediction of the drilling speed is realized.

[0027] Further, the initial steady drilling speed prediction model is trained by using the historical data of the completed well.

[0028] Further, the steady drilling speed prediction model is taken as the basis at each time of drilling.

[0029] Further, the drilling data is preprocessed, one round of training is performed by using the data of each single or stand, and the drilling speed prediction model is updated in real time.

[0030] The present application has the following advantages: 1. The characteristics of the hot dry rock stratum and the characteristics of the PDC drill bit are considered, the bottom hole temperature, the drill bit wear and the rock compressive strength are introduced into the drilling speed prediction mechanism model, and the prediction accuracy of the model is improved.

[0031] 2. The deep learning model of the CGA architecture is constructed, the key drilling features are automatically extracted, the long-term dependent time sequence mode is calculated, the model parameters are dynamically adjusted when the drilling condition changes, the space features, the time sequence dependence and the dynamic attention are fused, the balance between the high accuracy and the strong generalization of the drilling speed prediction is realized.

[0032] 3. The coupling of the mechanism model (WT) and the deep learning model (CGA) improves the accuracy of mechanical drilling speed prediction and increases the applicability of the model.

[0033] 4. The mixed model update strategy of real-time incremental update and periodic full update is better than separate full update or real-time incremental update, which takes into account real-time features, global and overall trends, and realizes real-time online prediction of mechanical drilling speed in complex hot dry rock strata drilling. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is the structural block diagram of the CGA model.

[0035] Figure 2 is a schematic diagram of the coupling mode of the mechanism model and the deep learning model.

[0036] Figure 3 is a process diagram of the mixed update strategy.

[0037] Figure 4 is an update flowchart of the mixed model update strategy. DETAILED DESCRIPTION

[0038] EMBODIMENT

[0039] Referring to the drawings, a mechanical drilling speed prediction method based on mechanism-deep learning and model mixed update strategy is provided. The deep learning model architecture, coupling mode and model mixed update strategy process are shown in the drawings. The modeling work is divided into four stages: establishing a mechanism drilling speed prediction model, establishing a drilling speed prediction model based on deep learning, establishing a coupled drilling speed prediction model, and establishing a mixed model update strategy. In the first stage, based on the proposed WT model, drilling data is collected and the relevant model coefficients in the mechanism model are fitted. In the second stage, a CNN-GRU-Attention model is constructed to establish a drilling speed prediction model based on deep learning. In the third stage, a WT-CAG model is established to establish a mixed drilling speed prediction model in a residual manner. In the fourth stage, a model update strategy combining real-time incremental update and periodic full update is established to update the mechanical drilling speed model. The specific implementation scheme is as follows: A mechanical drilling speed prediction method based on mechanism-deep learning and mixed model update strategy includes the following steps: Step 1: Establish a mechanical drilling speed prediction mechanism model (WT model) covering PDC bit wear and bottom hole temperature According to the drilling characteristics of hot dry rock strata, a mechanical drilling speed prediction mechanism model WT covering bottom hole temperature and bit wear is proposed, as shown in Equation 1.

[0040] (1) In the formula, W f It is a function of PDC drill bit wear, as shown in Formula 2; WOB It is drilling pressure, measured in kN. RPM It refers to rotational speed, measured in r / min. This is the drilling fluid displacement, measured in L / s; T It is the temperature of the strata encountered during drilling, in °C; P It is the bottom hole pressure difference, in MPa; S It is the compressive strength of the rock encountered during drilling, measured in MPa. D d This is the drill bit diameter, in mm; K These are the model's overall coefficients. a, b, c, d These are all coefficients of the model parameters, and are dimensionless.

[0041] (2) In the formula, It is the IADC bit passivation grade, dimensionless.

[0042] Collect historical data of completed wells and calculate the various comprehensive coefficients in formula (1) to obtain the mechanism model for drilling speed prediction.

[0043] Step 2: Establish a deep learning model architecture CNN-GRU-Attention (CGA model) (1) Establish a one-dimensional convolutional neural network, consisting of an input layer, a convolutional layer, a pooling layer, a flattened layer, a fully connected layer, and an output layer.

[0044] (2) Establish a gated loop unit to automatically extract the corresponding drilling features.

[0045] (3) Construct an attention mechanism to dynamically adjust the degree of attention to input data or time steps.

[0046] (4) Combining one-dimensional convolutional neural networks, gated recurrent units and attention mechanisms, a CNN-GRU-Attention deep learning model using a decoder and encoder architecture is constructed to predict the mechanical drilling speed.

[0047] Step 3: Create a hybrid mechanical drilling rate prediction model (HRM) using the WT model and the CAG model. (1) Calculate the preliminary drilling speed prediction value using the mechanism model fitted by drilling data.

[0048] (2) Calculate the difference between the predicted value and the actual drilling speed, i.e., the residual.

[0049] (3) Use drilling data and prediction residuals as input to train the deep learning model to obtain the final drilling speed prediction value.

[0050] Step 4: Establish a hybrid update strategy that combines periodic full updates and real-time incremental updates. (1) The initially established hybrid drilling rate prediction model is used as the steady-state mechanical drilling rate prediction model.

[0051] (2) When a single or vertical shaft is drilled, and the mechanical energy or mechanical drilling speed fluctuates continuously (within 2m or 20min, the continuous fluctuation exceeds 40%, and the specific adjustment needs to be made according to the block situation), collect drilling data and start real-time incremental updates.

[0052] (3) Lock the fully connected layers of the hybrid model, update the hidden layers only using real-time drilling sequence data, and train only one round.

[0053] (4) When performing tripping or short tripping, collect all drilling data for the current tripping, randomly shuffle it, and start a periodic full update.

[0054] (5) Discard the last real-time incremental update drilling rate prediction model, and based on the previous steady-state drilling rate prediction model, lock the fully connected layer, update only the hidden layer, train for one round, and use the updated drilling rate prediction model as the base model of the real-time incremental update model.

[0055] Step 5: Update the drilling speed prediction model using the drilling parameters obtained during drilling according to the hybrid model update strategy.

[0056] Step 6: Based on the drilling parameters obtained at the current moment, input the updated mechanical drilling rate prediction model to successfully predict the drilling rate at the next moment.

[0057] Example 2 Taking industrial data from seven wells at a drilling site in a certain region's hot dry rock area as an example Completed drilling data for the target block were collected, including mechanical drilling rate, drilling pressure, rotational speed, displacement, bottom hole temperature, bottom hole pressure differential, compressive strength of the rock encountered, drill bit diameter, and drill bit wear. The model coefficients of the corresponding mechanistic models were fitted and obtained as follows: 456.26, -0.81, 0.16, and 0.05.

[0058] By combining one-dimensional convolutional neural networks, gated recurrent units, and attention mechanisms, a CNN-GRU-Attention (CGA) deep learning model is constructed for predicting mechanical drilling speed using a decoder and encoder architecture, such as... Figure 1 As shown in Table 1, the optimal parameters of the model for this method are: Table 1. Optimal hyperparameters of the model proposed in this invention Hyperparameters Values Batch size 8 Learning rate 0.0001 Training epochs 1000 Sequence length 5 Dropout 0.1 Weight decay 0.001 Number of layers 2 Hidden layer size 64 Attention heads 1 Optimization algorithm Adam Loss function MSELoss Using the time of single-joint connection, tripping in and out of the well, and abnormal fluctuations in drilling speed as model update conditions, the drilling speed prediction model is updated through a hybrid model update strategy based on formation drilling.

[0059] This method demonstrates good performance in integrating transient and long-term steady-state changes during drilling. It also measures the root mean square error (RMSE), average absolute percentage error (MAPE), and coefficient of determination (R²). 2 The values ​​of 0.23, 26.76, and 0.77 respectively showed good results.

[0060] The method proposed in this invention considers the effects of bottom hole temperature and drill bit wear during drilling. By integrating the characteristics of physical mechanism models and machine learning models, it improves both the accuracy and applicability of the drilling speed prediction model. By comprehensively considering both transient and steady-state changes in the drilling speed process, a hybrid update strategy for model updating is established. This strategy not only captures accidental and sudden factors during drilling in real time but also avoids the accumulation of errors from long-term accidental factors through periodic full updates, thus incorporating the overall drilling pattern and better reflecting the actual field conditions of drilling projects.

[0061] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A drilling rate prediction method based on mechanism-deep learning and a hybrid model update strategy, characterized in that: The mechanical drilling rate prediction method includes the following steps: Step 1: Establish a mechanical drilling rate prediction mechanism model that includes PDC drill bit wear and bottom hole temperature. To address the drilling characteristics of hot and dry rock formations, a mechanical drilling rate prediction mechanism model WT, which includes bottom hole temperature and drill bit wear, is proposed, as shown in Formula 1. (1) In the formula, W f It is a function of PDC drill bit wear, as shown in Formula 2; WOB It is drilling pressure, measured in kN. RPM It refers to rotational speed, measured in r / min. This is the drilling fluid displacement, measured in L / s; T It is the temperature of the strata encountered during drilling, in °C; P It is the bottom hole pressure difference, in MPa; S It is the compressive strength of the rock encountered during drilling, measured in MPa. D d This is the drill bit diameter, in mm; K These are the model's overall coefficients. a, b, c, d These are all coefficients of the model parameters, and are dimensionless; (2) In the formula, It is the IADC bit passivation grade, dimensionless; Collect historical data of completed wells and calculate the various comprehensive coefficients in formula (1) to obtain the mechanism model for drilling speed prediction; Step 2: Establish a deep learning model architecture CNN-GRU-Attention (1) Establish a one-dimensional convolutional neural network, consisting of an input layer, a convolutional layer, a pooling layer, a flattened layer, a fully connected layer, and an output layer; (2) Establish a gated loop unit to automatically extract the corresponding drilling features; (3) Construct an attention mechanism to dynamically adjust the degree of attention to input data or time steps; (4) Combining one-dimensional convolutional neural networks, gated recurrent units and attention mechanisms, a CNN-GRU-Attention deep learning model using a decoder and encoder architecture is constructed to predict the mechanical drilling speed. Step 3: Create a hybrid mechanical drilling rate prediction model using the WT model and the CAG model. (1) Calculate the preliminary drilling speed prediction value using the mechanism model fitted from the drilling data; (2) Calculate the difference between the predicted value and the actual drilling speed; (3) Use drilling data and prediction residuals as input to train a deep learning model to obtain the final drilling speed prediction value; Step 4: Establish a hybrid update strategy that combines periodic full updates and real-time incremental updates. (1) The initially established hybrid drilling rate prediction model is used as the steady-state mechanical drilling rate prediction model; (2) When a single root or vertical root is drilled, and the mechanical energy or mechanical drilling rate fluctuates continuously, collect drilling data and start real-time incremental updates; (3) Lock the fully connected layers of the hybrid model, update the hidden layers only using real-time drilling sequence data, and train only one round; (4) When performing tripping or short tripping, collect all drilling data for the current tripping, randomly shuffle it, and start a periodic full update; (5) Discard the last real-time incremental update drilling rate prediction model, and based on the previous steady-state drilling rate prediction model, lock the fully connected layer, update only the hidden layer, train for one round, and use the updated drilling rate prediction model as the base model of the real-time incremental update model. Step 5: Update the drilling rate prediction model using the drilling parameters obtained during drilling according to the hybrid model update strategy; Step 6: Based on the drilling parameters obtained at the current moment, input the updated mechanical drilling rate prediction model to successfully predict the drilling rate at the next moment.

2. The drilling rate prediction method based on mechanism-deep learning and hybrid model update strategy according to claim 1, characterized in that: Collect complete drilling data for the target block, including mechanical drilling rate, drilling pressure, rotational speed, displacement, bottom hole temperature, bottom hole pressure differential, compressive strength of the rock encountered, drill bit diameter, and drill bit wear data.

3. The drilling rate prediction method based on mechanism-deep learning and hybrid model update strategy according to claim 1, characterized in that: Using the time of single-joint connection, tripping in and out of the well, and abnormal fluctuations in drilling speed as model update conditions, the drilling speed prediction model is updated through a hybrid model update strategy based on formation drilling.

4. The drilling rate prediction method based on mechanism-deep learning and hybrid model update strategy according to claim 1, characterized in that: If the prediction accuracy or MSE fluctuates too much, update the drilling speed prediction model.

5. The drilling rate prediction method based on mechanism-deep learning and hybrid model update strategy according to claim 1, characterized in that: When performing tripping or short tripping, the transient rate of drilling prediction model for each single or vertical shaft is discarded. Based on the previous steady-state rate of drilling prediction model, the data in the model is randomly shuffled and updated for one cycle. This updated model serves as the basis for the mechanical rate of drilling prediction model for the next single or vertical shaft, thus enabling online prediction of mechanical rate of drilling.

Images