A method for predicting rate of penetration based on physical consistency constraints

Abstract

This invention proposes a mechanical drilling rate prediction method based on physical consistency constraints. By collecting and preprocessing drilling time-series data from multiple wells, a sliding window is used to construct historical operating conditions and future prediction sequences. A prediction model integrating a gated cyclic unit and a multilayer perceptron is constructed to achieve joint modeling of historical features and future controllable parameters. Simultaneously, a physical consistency loss function based on a mechanical specific energy model is introduced and jointly optimized with data error loss. A dynamic weighting mechanism enables collaborative learning of data-driven and physical constraints. This method can achieve multi-step prediction of mechanical drilling rate, improve the accuracy and physical rationality of the prediction results, and enhance the model's generalization ability under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of petroleum engineering technology, and more specifically to a method for predicting mechanical drilling rate based on physical consistency constraints. Background Technology

[0002] The rate of penetration (ROP) is a core indicator for measuring drilling efficiency, directly affecting drilling cycle and overall cost. In actual drilling operations, the ROP is influenced by a combination of factors such as formation lithology, drilling pressure, rotational speed, and displacement, exhibiting complex characteristics of strong nonlinearity, high dimensionality, and multi-scale temporal coupling. Accurate prediction of ROP is of significant engineering value for optimizing drilling parameters and preventing complex downhole situations.

[0003] In recent years, with the widespread application of artificial intelligence technology in the petroleum engineering field, pure data-driven models based on machine learning and deep learning have gradually become the mainstream method for predicting mechanical drilling speed. For example, patent CN202311024201.0 proposes a prediction scheme combining federated learning and bidirectional LSTM. This scheme mainly focuses on using deep learning networks to mine the temporal features of drilling data, and uses a distributed federated framework to perform multi-node joint training while protecting data privacy, thereby improving the generalization performance of the pure data-driven model; patent CN202010231812.2 discloses a method based on spatial and multi-dimensional features. The constructed machine learning-based rate of drilling (RDR) prediction method primarily extracts composite features such as wellbore structure, drill string, and formation in the target well section and adjacent areas, directly inputting these features into a pre-trained machine learning model for "black box" mapping, aiming to predict the mechanical rate of drilling at a specific well depth. However, the aforementioned existing technologies still suffer from problems such as a lack of physical mechanism constraints and insufficient accuracy in future multi-step predictions, making it difficult to meet the reliability and forward-looking optimization requirements of drilling operations under complex conditions. Therefore, this paper proposes a mechanical rate of drilling prediction method that can combine historical data, future controllable parameters, and introduce physical consistency constraints to improve prediction accuracy and physical rationality. Summary of the Invention

[0004] To address the problems of neglecting physical laws, poor physical plausibility of prediction results, and insufficient generalization ability in existing pure data-driven mechanical drilling rate prediction methods, this invention proposes a mechanical drilling rate prediction method based on physical consistency constraints. This method significantly improves the prediction accuracy and physical plausibility of mechanical drilling rate by introducing a physical consistency loss function based on the Teale mechanical specific energy formula and jointly optimizing it with the error loss function.

[0005] To solve the above-mentioned technical problems, the main steps of the technical solution adopted by the present invention include:

[0006] 1) Collect drilling data from multiple wells during the drilling process and preprocess the collected drilling data;

[0007] 2) Arrange the preprocessed drilling data in chronological order to construct a time series with time step markers, and generate historical operating condition sequences and future prediction sequences through a sliding window, and then divide them into training sets and test sets;

[0008] 3) Construct a mechanical drilling rate prediction model with built-in physical consistency constraints, and introduce physical guidance features based on mechanical specific energy into the model structure;

[0009] 4) Train the model by using historical operating condition sequences and future controllable parameter sequences as joint inputs;

[0010] 5) Construct a multi-constraint joint loss function during model training and optimize it through a dynamic weighting mechanism;

[0011] 6) Use the trained model to make predictions on the test set, and use model evaluation metrics to evaluate the prediction results.

[0012] In step 1), drilling data from multiple wells is collected during the drilling process, and the collected drilling data is preprocessed, including the following steps:

[0013] 1.1) Collect drilling data from multiple wells during the drilling process. The drilling data includes drilling condition parameters, drill bit parameters, and predicted target parameters. The drilling condition parameters include: measured depth (MD), vertical depth (TVD), rotary table speed (RPM), weight on bit (WOB), inlet displacement (MFIA), outlet displacement (MDIA), outlet density (MDOA), hook load (HKLA), torque (TQA), standpipe pressure (SPPA), and drill bit position (BDTI). The drill bit parameters include the drill bit diameter (D). The predicted target parameter is the rate of penetration (ROP).

[0014] 1.2) Select parameters from drilling parameters that have a regulatory effect on mechanical drilling rate to form a controllable parameter set: {RPM, WOB, MFIA, MDIA}

[0015] Among them, the selected parameters can achieve real-time control of drilling speed by adjusting the drill bit rotation speed, drilling pressure and mud circulation flow rate;

[0016] 1.3) Perform data cleaning on the collected drilling data to remove missing data and outlier data exceeding three times the standard deviation;

[0017] 1.4) Normalize the cleaned drilling data to eliminate differences in the dimensions and numerical ranges of different parameters.

[0018] In step 2), constructing time-series data with time-step markers includes:

[0019] 2.1) Arrange the drilling data into a sequence in chronological order. And a time step marker t is introduced to form a timing representation:

[0020] X t =[x t ,t]

[0021] Where, x t It includes historical operating parameters, drill bit parameters, and controllable parameters; t is the time step index; T is the total time length.

[0022] 2.2) Extract historical operating condition sequences and corresponding future prediction sequences based on the sliding window method. Historical sequence X t-L：t Used as model input, predicting sequence Y t：t+H Used for supervised training:

[0023] X t-L：t ={X t-L X t-L+1 ,...,X t}

[0024] Y t:t+H ={ROP t ROP t+1 , ..., ROP t+H}

[0025] Where L is the historical window length; H is the prediction step size; X t-L:t For historical operating condition sequences; Y t：t+H This is the predicted sequence.

[0026] In step 3), the mechanical drilling rate prediction model includes:

[0027] 3.1) Feature Extraction Module: Utilizes a gated recurrent unit (GRU) to extract features from historical operating condition sequences.

[0028] h t =GRU(x t ,h t-1 )

[0029] in, The hidden state vector;

[0030] 3.2) Feature Fusion Module: Fuses historical features with future controllable parameters.

[0031] z t =[h t ;u t ]

[0032] Among them, u t z is a sequence vector of controllable parameters; tFor fusion feature representation;

[0033] 3.3) Prediction Output Module: The Multilayer Perceptron (MLP) maps the fused features to predicted values.

[0034]

[0035] in, To predict the mechanical drilling rate at corresponding time steps, it can be extended to multi-step prediction.

[0036] In step 5), a multi-constraint joint loss function is constructed during model training and optimized through a dynamic weighting mechanism, specifically including the following steps:

[0037] 5.1) Define the data fitting loss as:

[0038]

[0039] Where N is the sample size. ROP represents the mechanical drilling rate predicted by the model. i This represents the actual mechanical drilling speed.

[0040] 5.2) The mechanical specific energy is calculated based on the Teale mechanical specific energy model, and its expression is:

[0041]

[0042] Where TQA is torque; RPM is rotational speed; D is drill bit diameter; WOB is drill pressure; For predicted mechanical specific energy; MSE i This represents the true mechanical specific energy.

[0043] 5.3) Define the joint loss function and dynamic weight optimization:

[0044] L total =L data +λL physics

[0045]

[0046] Where λ is the weight coefficient; t is the training round; T is the total rounds; and α is the adaptive scaling factor.

[0047] In step 6), the trained mechanical drilling rate prediction model is used to predict the test data, and the prediction results are evaluated using model evaluation metrics, including:

[0048] 6.1) Use the trained mechanical drilling rate prediction model to predict the mechanical drilling rate from the test data;

[0049] 6.2) Obtain the error between the predicted value and the actual value;

[0050] 6.3) Using the coefficient of determination R 2 The mean relative error (MRE) and mean absolute error (MAE) are used to evaluate the prediction results.

[0051] The advantages of this invention compared to the prior art are mainly reflected in:

[0052] (1) This invention addresses the problem of insufficient physical rationality in purely data-driven models: It innovatively introduces a physical consistency loss function based on Teale's mechanical specific energy formula to impose drilling energy conservation constraints during model training. Through this constraint, the deep learning model is forced to follow the "physical matching relationship between drilling pressure, rotational speed, drilling rate, and mechanical specific energy" while fitting historical drilling data. This mechanism effectively avoids anomalies in prediction results that violate physical laws, and significantly improves the prediction accuracy and generalization ability of the model in complex and unknown formation environments.

[0053] (2) Achieved efficient fusion of historical time series features and future control parameters: This invention uses gated cyclic units (GRUs) to efficiently extract the time series evolution features of historical drilling conditions and concatenates them with the future controllable operation parameter sequence as conditional features. This fusion strategy can accurately characterize the synergistic effect of historical formation trends and future artificial adjustment parameters on mechanical drilling rate, thereby providing a more reliable decision-making basis for forward-looking parameter optimization in drilling operations, while enhancing the model's responsiveness to future operation adjustments. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the embodiments will be briefly described below.

[0055] Figure 1 This is an overall flowchart of the method of the present invention;

[0056] Figure 2 This is a structural diagram of the mechanical drilling speed prediction model of the present invention;

[0057] Figure 3 This diagram illustrates a comparison between the method of this invention and the mechanical drilling rate prediction results of a method without physical consistency constraints. Detailed Implementation

[0058] Example 1

[0059] See Figure 1 The present invention proposes a mechanical drilling rate prediction method based on physical consistency constraints, which specifically includes the following steps:

[0060] 1) Collect drilling data from multiple wells during the drilling process and preprocess the collected drilling data;

[0061] 2) Arrange the preprocessed drilling data in chronological order to construct a time series with time step markers, and generate historical operating condition sequences and future prediction sequences through a sliding window, and then divide them into training sets and test sets;

[0062] 3) Construct a mechanical drilling rate prediction model with built-in physical consistency constraints, and introduce physical guidance features based on mechanical specific energy into the model structure;

[0063] 4) Train the model by using historical operating condition sequences and future controllable parameter sequences as joint inputs;

[0064] 5) Construct a multi-constraint joint loss function during model training and optimize it through a dynamic weighting mechanism;

[0065] 6) Use the trained model to make predictions on the test set, and use model evaluation metrics to evaluate the prediction results.

[0066] In this embodiment, step 1) involves collecting drilling data from multiple wells using a sensor system and drilling monitoring system at an oilfield drilling site. The collected drilling data is then cleaned and normalized, and training and testing datasets are constructed. Specifically, this includes:

[0067] 1.1) Drilling Data Acquisition: Drilling data from multiple wells was acquired using a sensor system and a drilling monitoring system at an oilfield drilling site. The drilling data includes drilling condition parameters, drill bit parameters, and predicted target parameters. Specifically: Drilling condition parameters include measured well depth (MD), vertical well depth (TVD), rotary table speed (RPM), weight on bit (WOB), inlet displacement (MFIA), outlet displacement (MDIA), outlet density (MDOA), hook load (HKLA), torque (TQA), riser pressure (SPPA), and drill bit position (BDTI); drill bit parameters include drill bit diameter (D); the predicted target parameter is the rate of penetration (ROP); historical drilling rate sample data is shown in Table 1.

[0068] Table 1 Example of historical drilling rate sample data

[0069]

[0070] 1.2) Screening of controllable parameters: Screen the parameters that have the effect of regulating mechanical drilling rate from the drilling condition parameters to form a set of controllable parameters {RPM, WOB, MFIA, MDIA}. The selected parameters can achieve real-time control of drilling rate by adjusting the drill bit rotation speed, drilling pressure and mud circulation flow rate.

[0071] 1.3) Data cleaning: Remove missing records from the dataset, use the 3σ criterion to identify and remove outlier data exceeding three times the standard deviation in each feature column; at the same time, delete invalid records with zero or negative ROP to ensure that the model training only uses valid data under normal drilling conditions.

[0072] 1.4) Data normalization: In order to eliminate the differences in numerical range between different physical dimensions, all feature data are processed using the min-max normalization method.

[0073] In this embodiment, step 2) involves arranging the preprocessed drilling data in chronological order to construct a time-series sequence with time step markers. A sliding window is then used to generate historical operating condition sequences and future prediction sequences, which are subsequently divided into training and testing sets. Specifically, this includes:

[0074] 2.1) Time Series Construction: Arrange drilling data into a time sequence. And a time step marker t is introduced to form a timing representation:

[0075] X t =[x t ,t]

[0076] Where, x t It includes historical operating parameters, drill bit parameters, and controllable parameters; t is the time step index; T is the total time length.

[0077] 2.2) Sliding window sampling: Based on the sliding window method, historical operating condition sequences and corresponding future prediction sequences are extracted. The historical sequence X t-L：t Used as model input, predicting sequence Y t：t+H Used for supervised training:

[0078] X t-L：t ={X t-L X t-L+1 , ..., X t}

[0079] Y t：t+H ={ROP t ROP t+1 , ..., ROP t+H}

[0080] Where L is the historical window length; H is the prediction step size; X t-L：t For historical operating condition sequences; Y t：t+H To predict the sequence, a sliding window technique is used to construct training samples, with 100 consecutive steps of historical drilling condition parameter sequences as the historical condition sequence X. t-L:t Using the mechanical drilling rate sequence corresponding to the next 50 steps as the prediction target sequence Y t：t+HMeanwhile, the corresponding sequence of controllable parameters for the next 50 steps (RPM, WOB, MFIA, MDIA) is introduced as the input of the future controllable parameters, where L equals 100 as the historical window length and H equals 50 as the prediction step size.

[0081] 2.3) Dataset splitting: Data from 3 wells was used as the training set, and data from 1 well was used as the test set to evaluate the model's generalization ability.

[0082] In this embodiment, step 3) involves constructing a mechanical drilling rate prediction model with built-in physical consistency constraints. Physical guidance features based on mechanical specific energy are introduced into the model structure, such as... Figure 2 As shown, it specifically includes:

[0083] 3.1) Feature Extraction Module: A Gated Recurrent Unit (GRU) network is used, containing two GRU hidden layers, each with 128 neurons, to extract temporal features from the input historical working condition sequence. The formula is as follows:

[0084] h t =GRU(x t h t-1 )

[0085] in, To hide the state vector, extract the temporal dependency features of historical drilling conditions;

[0086] 3.2) Feature Fusion Module: This module fuses the historical time-series features h extracted by the GRU network. t With the future controllable parameter sequence vector u t The features are concatenated along the feature dimension to form a fused feature representation z. t :

[0087] z t =[h t ;u t ]

[0088] Among them, u t This is a sequence vector of controllable parameters for the next 50 steps, containing four parameters: RPM, WOB, MFIA, and MDIA.

[0089] 3.3) Prediction Output Module: Employs a Multilayer Perceptron (MLP) structure, including two fully connected layers, to map the fused feature vectors into a predicted sequence of future mechanical drilling rates.

[0090]

[0091] in, To predict the mechanical drilling rate at corresponding time steps, it can be extended to multi-step prediction. It enables continuous prediction of the mechanical drilling rate for the next 50 steps.

[0092] In this embodiment, step 4) involves using the historical operating condition sequence and the future controllable parameter sequence as joint inputs to train the model, specifically including:

[0093] 4.1) Data input: The mechanical drilling rate prediction model is constructed by inputting historical data and future controllable parameters to achieve the synergistic use of historical working condition information and future controllable operation information;

[0094] 4.2) Model parameter update: The backpropagation algorithm is used to calculate the gradient of the loss function with respect to the model parameters, and the Adam optimizer is used to iteratively update the model parameters. The training epochs are set to 300 epochs and the training set batch size is set to 32.

[0095] 4.3) Training convergence determination: When the change of the loss function is lower than the preset threshold in a number of consecutive training rounds, or when the maximum number of training rounds of 300 is reached, the model training is determined to be complete, and the final mechanical drilling speed prediction model is obtained.

[0096] In this embodiment, step 5) involves constructing a multi-constraint joint loss function during model training and optimizing it using a dynamic weighting mechanism. Specifically, this includes:

[0097] 5.1) Data Fitting Loss: The mean squared error is used as the data fitting loss, and its expression is as follows:

[0098]

[0099] Where N is the sample size. ROP represents the mechanical drilling rate predicted by the model. i This represents the actual mechanical drilling speed.

[0100] 5.2) Physical consistency loss: The mechanical specific energy is calculated based on the Teale mechanical specific energy model, and its expression is:

[0101]

[0102] Where TQA is torque; RPM is rotational speed; D is drill bit diameter; WOB is drilling pressure; and the model-predicted mechanical drilling speed is... and ROP (Real Mechanical Drilling Rate) i Substituting into the above formula, the predicted mechanical specific energy is calculated. MSE (Mean Energy Specific Energy) of Real Machines i Construct the physical consistency loss function:

[0103]

[0104] By minimizing the physical consistency loss function, the model prediction results are made to satisfy the energy constraint relationship in the drilling process, thereby improving the physical rationality of the prediction results;

[0105] 5.3) Joint Loss Function and Dynamic Weight Optimization: The data fitting loss and physical consistency loss are weighted and combined to construct a multi-constraint joint optimization objective function:

[0106] L total =L data +λL physics

[0107]

[0108] Where λ is the physical constraint weight coefficient; t is the current training round; T is the total training rounds; α is the adaptive amplification coefficient, which ranges from 0.5 to 2.0. In this embodiment, α is 1.0. In the early stage of model training, λ0 is set to 0.1 so that the model learns the data distribution characteristics first. As the training rounds increase, λ gradually increases according to the above formula, so that the model strengthens the constraint on physical consistency in the later stage, thereby achieving a dynamic balance between data-driven and physical constraints.

[0109] In this embodiment, step 6) involves using the trained mechanical drilling rate prediction model to predict the test data and evaluate the model performance, specifically including:

[0110] 6.1) Model prediction: The trained mechanical drilling rate prediction model is tested on the test set. Using historical data of the current 100 drilling steps and controllable parameter data of the next 50 steps, the mechanical drilling rate of the next 50 steps is predicted.

[0111] 6.2) Error Acquisition: Compare the model's predicted values ​​with the actual values ​​on the test set to obtain the prediction error, which will be used for subsequent evaluation calculations;

[0112] 6.3) Calculation of evaluation indicators: The coefficient of determination R is used. 2 The mean relative error (MRE) and mean absolute error (MAE) are used to evaluate the predicted and actual values. The formulas for the model evaluation metrics are as follows:

[0113]

[0114] Among them, y i For the true value, For predicted values, is the mean of the actual values, and n is the total number of samples.

[0115] In this embodiment, to verify the effectiveness of the method of the present invention, a comparative model without physical consistency constraints is introduced as a benchmark. Both models are trained and evaluated using the same training and test datasets to ensure experimental fairness.

[0116] In this embodiment, step 6) involves a systematic analysis of the prediction results of the trained model of the present invention and the comparison model. The evaluation is conducted from multiple perspectives, including error distribution, trend fitting accuracy, and prediction stability, to comprehensively measure the model performance. Figure 3 This section compares the prediction results of the two models on the test set; the main evaluation metric includes the coefficient of determination R. 2 The mean relative error (MRE) and mean absolute error (MAE) of the model are shown in Table 2.

[0117] Table 2 Model Evaluation Results

[0118]

[0119] As can be seen from Table 2, the method of the present invention in R 2 The method significantly outperforms the comparative model without physical constraints in terms of MRE and MAE, indicating that combining physical consistency constraints effectively improves the accuracy and stability of mechanical drilling rate prediction. The method of this invention can not only fit the trend of historical drilling data more accurately, but also respond more reasonably to future operating parameters, providing reliable technical support for real-time decision-making, parameter optimization and risk control in the drilling process.

Claims

1. A method for predicting mechanical drilling rate based on physical consistency constraints, characterized in that, Includes the following steps: 1) Collect drilling data from multiple wells during the drilling process and preprocess the collected drilling data; 2) Arrange the preprocessed drilling data in chronological order to construct a time series with time step markers, and generate historical operating condition sequences and future prediction sequences through a sliding window, and then divide them into training sets and test sets; 3) Construct a mechanical drilling rate prediction model with built-in physical consistency constraints, and introduce physical guidance features based on mechanical specific energy into the model structure; 4) Train the model by using historical operating condition sequences and future controllable parameter sequences as joint inputs; 5) Construct a multi-constraint joint loss function during model training and optimize it through a dynamic weighting mechanism; 6) Use the trained model to make predictions on the test set, and use model evaluation metrics to evaluate the prediction results.

2. The method according to claim 1, characterized in that, In step 1), drilling data from multiple wells is collected during the drilling process, and the collected drilling data is preprocessed, including the following steps: 1.1) Collect drilling data from multiple wells during the drilling process. The drilling data includes drilling condition parameters, drill bit parameters, and predicted target parameters. The drilling condition parameters include: measured depth MD, vertical depth TVD, rotary table speed RPM, weight on bit WOB, inlet displacement MFIA, outlet displacement MDLA, outlet density MDOA, hook load HKLA, torque TQA, standpipe pressure SPPA, and drill bit position BDTI. The drill bit parameters include the drill bit diameter D. The predicted target parameter is the rate of penetration (ROP). 1.2) Select parameters from drilling operating parameters that have a regulatory effect on mechanical drilling rate to form a controllable parameter set: {RPM,WOB,MFIA,MDIA} The selected parameters enable real-time control of drilling speed by adjusting the drill bit rotation speed, drilling pressure, and mud circulation flow rate. 1.3) Perform data cleaning on the collected drilling data to remove missing data and outlier data exceeding three times the standard deviation; 1.4) Normalize the cleaned drilling data to eliminate differences in the dimensions and numerical ranges of different parameters.

3. The method according to claim 1, characterized in that, In step 2), constructing time-series data with time-step markers includes: 2.1) Arrange the drilling data into a sequence in chronological order. And a time step marker t is introduced to form a timing representation: X t =[x t ,t] Where, x t It includes historical operating parameters, drill bit parameters, and controllable parameters; t is the time step index; T is the total time length. 2.2) Extract historical operating condition sequences and corresponding future prediction sequences based on the sliding window method. Historical sequence X t-L：t Used as model input, predicting sequence Y t：t+H Used for supervised training: X t-L：t ={X t-L ,X t-L+1 ,...,X t } Y t：t+H ={ROP t ROP t+1 ..., ROP t+H } Where L is the historical window length; H is the prediction step size; X t-L:t For historical operating condition sequences; Y t：t+H This is the predicted sequence.

4. The method according to claim 1, characterized in that, In step 3), the mechanical drilling rate prediction model includes: 3.1) Feature Extraction Module: Utilizes a gated recurrent unit (GRU) to extract features from historical operating condition sequences. h t =GRU(x t ,h t-1 ) in, The hidden state vector; 3.2) Feature Fusion Module: Fuses historical features with future controllable parameters. z t =[h t ;u t ] Among them, u t z is a sequence vector of controllable parameters. t For fusion feature representation; 3.3) Prediction Output Module: The Multilayer Perceptron (MLP) maps the fused features to predicted values. in, To predict the mechanical drilling rate at corresponding time steps, it can be extended to multi-step prediction. .

5. The method according to claim 1, characterized in that, In step 5), a multi-constraint joint loss function is constructed during model training and optimized through a dynamic weighting mechanism, specifically including the following steps: 5.1) Define the data fitting loss as: Where N is the number of samples; The mechanical drilling rate predicted by the model; ROP i This represents the actual mechanical drilling speed. 5.2) The mechanical specific energy is calculated based on the Teale mechanical specific energy model, and its expression is: Where TQA is torque; RPM is rotational speed; D is drill bit diameter; WOB is drill pressure; For predicted mechanical specific energy; MSE i This represents the true mechanical specific energy. 5.3) Define the joint loss function and dynamic weight optimization: THE total =L data +λL physics Where λ is the weight coefficient; t is the training round; T is the total round; and α is the adaptive amplification coefficient.

6. The method according to claim 1, characterized in that, In step 6), the trained mechanical drilling rate prediction model is used to predict the test data, and the prediction results are evaluated using model evaluation metrics, including: 6.1) Use the trained mechanical drilling rate prediction model to predict the mechanical drilling rate from the test data; 6.2) Obtain the error between the predicted value and the actual value; 6.3) Using the coefficient of determination R 2 The prediction results are evaluated using mean relative error (MRE) and mean absolute error (MAE).

Images