Convolutional bidirectional time sequence indicator diagram prediction method based on specific data driving

Abstract

The invention discloses a convolution bidirectional time sequence indicator diagram prediction method based on specific data driving, and belongs to the technical field of oil extraction indicator diagram prediction, and the method comprises the following steps: 1, obtaining indicator diagram data, and recording a dynamic corresponding relation between a polished rod load and a suspension center displacement through an indicator diagram; 2, performing data cleaning, data alignment and data normalization on the indicator diagram data, wherein the data obtained after normalization processing is indicator diagram load specific data; step 3, constructing a time sequence sample set based on indicator diagram load specific data; 4, constructing and training a convolution bidirectional time sequence model; and step 5, applying the trained convolution bidirectional time sequence model to prediction of an oil well field oil pumping unit indicator diagram. The method has the capability of accurately predicting various working conditions and even complex combined working conditions, strives for more time for pump repair after pump inspection, and has the characteristics of high calculation speed, low economic cost and high characterization precision.

Description

Technical Field

[0001] The invention belongs to the technical field of oil production dynamometer diagram prediction, and in particular relates to a convolution bidirectional time series dynamometer diagram prediction method based on specific data driving. Background Art

[0002] In the field of oil extraction, ensuring the efficient and stable operation of the pumping unit and accurately grasping the working conditions play a key role in improving extraction efficiency, reducing costs and ensuring safe production. In actual production, the working conditions of the pumping unit are complex and diverse, including not only a single working condition, but also a combination of multiple working conditions intertwined with each other.

[0003] The existing dynamometer prediction model cannot fully cover these working conditions. Once a complex combination of working conditions is encountered, the prediction accuracy drops significantly and cannot provide a reliable decision-making basis for oilfield production. Therefore, how to use oilfield big data and use deep learning methods to accurately predict a variety of working conditions is of great significance to oil well production. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a convolutional bidirectional timing indicator diagram prediction method based on specific data driving, which can accurately predict multiple working conditions, even complex combined working conditions.

[0005] The technical solution of the present invention is as follows: A convolution bidirectional timing indicator diagram prediction method based on specific data driving comprises the following steps: Step 1, obtaining the dynamometer diagram data, which records the dynamic correspondence between the polished rod load and the suspension point displacement; Step 2: Clean, align and normalize the dynamometer data. The data obtained after normalization is the load-specific data of the dynamometer; Step 3: construct a time series sample set based on the load-specific data of the dynamometer diagram; Step 4: Build and train a convolutional bidirectional time series model; Step 5: Apply the trained convolutional bidirectional time series model to the prediction of the pumping unit indicator diagram at the oil well site.

[0006] Furthermore, the specific process of step 2 is as follows: Step 2.1, clean the dynamometer data; the specific process is: first perform non-null value integrity check, remove the well condition records with null values ​​in the load displacement data dimension, and then obtain the dynamometer time series data set with block ownership, well group ownership, acquisition time and time series characteristics through directional extraction; Step 2.2: Align the dynamometer data. The specific process is as follows: Step 2.2.1, divide the displacement load data points of the dynamometer diagram into upstroke data and downstroke data, and splice the downstroke data in reverse order after the upstroke data. The calculation formula is as follows: (1); in, It is the displacement point of the downstroke after splicing; is the maximum displacement value of the dynamometer diagram; It is the displacement point of the downstroke before splicing; Step 2.2.2, determine the fixed difference displacement point, keep the number of data points before and after interpolation unchanged, the formula is: (2); in, is the fixed difference displacement point; is the number of data points after data alignment; A function that generates a specific sequence of values; Step 2.2.3, use cubic spline interpolation to interpolate the original data, by fitting a cubic polynomial between every two adjacent displacement load points, ensuring that the first-order derivative and the second-order derivative of each segment of the polynomial are continuous at the connection point, and satisfying the natural boundary condition that the second-order derivative at the endpoint is zero, thereby generating a smooth curve, and then obtaining the load data at the fixed differential displacement point, and replacing the original displacement load data with the displacement load data obtained by interpolation, to obtain new dynamometer diagram data of the same differential displacement point; Step 2.3: normalize the new dynamometer data. The data obtained after normalization is the load-specific data of the dynamometer. The specific formula is as follows: (3); (4); in, is the normalized displacement; is the actual displacement; is the minimum displacement value of the dynamometer diagram; is the normalized load value; is the actual load value; , are the maximum load value and the minimum load value respectively.

[0007] Furthermore, the specific process of step 3 is as follows: Step 3.1, group the load-specific data of the dynamometer diagram according to blocks and well groups, ensure that the data of the well group in each block is processed independently, and take a section of time series data of a well in a block as a sample; Step 3.2, sort the data in each block and well group according to the acquisition time; Step 3.3, using the sliding window sampling method, extract the data pairs of the historical 28 days and the future 28 days from the data of each block and well group; the specific process of the sliding window sampling method is: starting from the starting position of the time series data, select windows of length 56 in sequence with a step size of 1 until a complete window can no longer be formed; for each window, the first 28 data points are used as historical data, and the last 28 data points are used as future data, so as to extract the data pairs of the historical 28 days and the future 28 days; Step 3.4: Collect the data pairs of all blocks to obtain the final time series sample set; Step 3.5: Divide the final time series sample set into training set, validation set and test set in proportion.

[0008] Furthermore, the specific process of step 4 is as follows: Step 4.1: Construct a convolutional bidirectional time series model for dynamometer data prediction; the convolutional bidirectional time series model consists of a one-dimensional convolutional layer, two bidirectional long short-term memory network layers, and a time-distributed fully connected layer; Step 4.2: Input the training set divided in step 3.5 into the model for model training, and then evaluate the model performance based on the test set and validation set to check the generalization ability of the model.

[0009] Furthermore, in step 4.1, the specific working process of the convolutional bidirectional time series model is: Step 4.1.1, the one-dimensional convolutional layer processes the input data and determines the initial target features; Step 4.1.2, input the initial target feature into the first bidirectional long short-term memory network layer to confirm the first temporal target feature; Step 4.1.3, input the first time series target feature into the second bidirectional long short-term memory network layer to confirm the second time series target feature; Step 4.1.4: Input the second time series target feature into the time distribution fully connected layer to obtain the prediction result of the future dynamometer data.

[0010] Furthermore, in step 4.2, the prediction accuracy of the convolutional bidirectional time series model is evaluated using the mean absolute error and the relative mean error, and the formulas are: (5); (6); in, is the mean absolute error; is the relative average error; For the The true value of samples; For the The predicted value of samples; is the sample size; Calculate the mean absolute error and relative mean error of the training set, validation set, and test set respectively, and then compare them. Output the model that meets the specified conditions; otherwise, re-collect data in step 1 for model training; The specified conditions are: when the relative average errors of the test set and the validation set are both smaller than the relative average error of the training set, or the difference between the relative average error of the test set and the relative average error of the training set is less than or equal to a preset threshold.

[0011] Furthermore, in step 5, the indicator diagram data of the oil pumping unit in the oil field is obtained in real time, and is processed according to the process of steps 1 and 2 to obtain the indicator diagram load specific data, and the indicator diagram load specific data is input into the trained model to calculate the future indicator diagram of the oil pumping unit in real time, thereby realizing continuous prediction of the indicator diagram of the oil pumping unit.

[0012] The beneficial technical effects brought about by the present invention: The convolution bidirectional timing dynamometer diagram prediction method based on specific data driving is concise and practical. Compared with the existing dynamometer diagram prediction method, it has the ability to accurately predict a variety of working conditions, even complex combination working conditions, so as to buy more time for pump inspection and repair. At the same time, it has the characteristics of fast calculation speed, low economic cost and high characterization accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The present invention is a flow chart of a convolution bidirectional timing dynamometer diagram prediction method based on specific data driven.

[0014] Figure 2 This is a prediction result diagram of the future dynamometer diagram under complex working conditions dominated by gas influence using the convolution bidirectional time series model in the experiment of the present invention.

[0015] Figure 3 This is a graph showing the prediction results of the future dynamometer diagram under complex working conditions dominated by wax deposition using the convolutional bidirectional time series model in the experiment of the present invention.

[0016] Figure 4 It is a line graph of the training rounds and the mean absolute error of the training set and the validation set in the experiment of the present invention. DETAILED DESCRIPTION

[0017] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: The present invention proposes a convolution bidirectional time series dynamometer diagram prediction method based on specific data drive. Before constructing the convolution bidirectional time series model, the difference displacement point is first fixed and the load data of the dynamometer diagram is processed by cubic spline interpolation to simplify the prediction target to load, thereby reducing the model complexity and improving the prediction accuracy. First, the dynamometer diagram data including blocks, well groups, acquisition time and time series characteristics are obtained; the acquired data is cleaned, aligned and normalized, and the data obtained after normalization is the load-specific data of the dynamometer diagram; based on the load-specific data of the dynamometer diagram, a time series sample set is constructed through features such as blocks and well groups; based on the processed time series sample set, a convolution bidirectional time series model of the dynamometer diagram is constructed and trained; based on the trained convolution bidirectional time series model, the dynamometer diagram is predicted. The present invention predicts the dynamometer diagram by a deep learning method; the model is evaluated using the mean absolute error, and the generalization ability of the model is checked using the test set data; finally, the future dynamometer diagram of the oil well pumping unit is calculated using the trained model.

[0018] A convolutional bidirectional timing indicator diagram prediction method based on specific data drive, the brief process is as follows Figure 1 As shown, the specific steps include: Step 1: Obtain the dynamometer diagram data. The dynamometer diagram fully presents the mechanical characteristics of the equipment in the complete stroke cycle by recording the dynamic correspondence between the polished rod load and the suspension point displacement. Obtaining high-quality dynamometer diagram data is the key to ensuring the accuracy of the dynamometer diagram prediction.

[0019] Step 2: Clean, align and normalize the dynamometer data. The data obtained after normalization is the load-specific data of the dynamometer. The specific process is as follows: Step 2.1, dynamometer data cleaning, first perform non-null value integrity check, remove well condition records with null values ​​in the load displacement data dimension, and then obtain dynamometer time series data set with block affiliation, well group affiliation, acquisition time and time series characteristics through directional extraction, in preparation for the subsequent construction of dynamometer sample set with time-space correlation characteristics.

[0020] Step 2.2, align the dynamometer data with sufficient features. Since the dynamometer collects displacement and load data at time intervals within a stroke, rather than load data at displacement intervals, the displacement and load of each point in an dynamometer data at adjacent time intervals are changing. For the prediction problem of the dynamometer diagram, obtaining the displacement and its corresponding load at the same time will increase the complexity of modeling. Therefore, by performing load interpolation processing on each set of data in the dynamometer diagram, fixing the differential displacement points, and only predicting the load, the complexity of the model is simplified. The details are as follows: Step 2.2.1: First, divide the displacement load data points of the dynamometer diagram into upstroke data and downstroke data. To keep the time sequence of each dynamometer diagram data coherent, the downstroke data is spliced ​​in reverse order after the upstroke data. The calculation formula is as follows: (1); in, It is the displacement point of the downstroke after splicing; is the maximum displacement value of the dynamometer diagram; It is the displacement point of the downstroke before splicing; Step 2.2.2, determine the fixed difference displacement point, keep the number of data points before and after interpolation unchanged, the formula is: (2); in, is the fixed difference displacement point; is the number of data points after data alignment; is a function that generates a specific sequence of values. The starting value is The end value is , Represents the interval of each increment or decrement.

[0021] Step 2.2.3: Use cubic spline interpolation to interpolate the original data. Fit a cubic polynomial between every two adjacent displacement load points to ensure that the first-order derivative and second-order derivative of each segment of the polynomial are continuous at the connection and meet the natural boundary condition that the second-order derivative at the endpoint is zero, thereby generating a smooth curve and obtaining the load data at the fixed differential displacement point. Replace the original displacement load data with the interpolated displacement load data to obtain the dynamometer data of the same differential displacement point.

[0022] Step 2.3: normalize the new dynamometer data. The data obtained after normalization is the load-specific data of the dynamometer. The specific formula is as follows: (3); (4); For each point in the dynamometer data series, is the normalized displacement; is the actual displacement; is the minimum displacement value of the dynamometer diagram; is the normalized load value; is the actual load value; , are the maximum load value and the minimum load value respectively; Since each set of displacement data is the same after normalization, only the normalized load data needs to be predicted.

[0023] Step 3: Based on the load-specific data of the dynamometer diagram, a time series sample set is constructed through features such as blocks and well groups. The specific process is as follows: Step 3.1, group the load-specific data of the dynamometer diagram according to blocks and well groups, ensure that the data of the well group in each block is processed independently, and take a section of time series data of a well in a block as a sample; Step 3.2: Sort the data in each block and well group according to the acquisition time to ensure the time sequence of the data; Step 3.3: Use the sliding window sampling method to extract the data pairs of the historical 28 days and the future 28 days from the data of each block and well group; the specific process of the sliding window sampling method is: starting from the starting position of the time series data, select windows of length 56 in sequence with a step size of 1 until a complete window can no longer be formed. For each window, the first 28 data points are used as historical data and the last 28 data points are used as future data, thereby extracting the data pairs of the historical 28 days and the future 28 days. In this way, multiple input-output sample pairs can be extracted from the time series data for constructing a time series sample set.

[0024] Step 3.4: Collect the data pairs of all blocks to obtain the final time series sample set; the time series sample set includes the input sample set and the output sample set. The data of all blocks in the past 28 days is the input sample set, and the data of all blocks in the next 28 days is the output sample set. Step 3.5: Divide the final time series sample set into training set, validation set and test set in proportion; specifically: divide the input sample set and output sample set in the time series sample set into training set, validation set and test set in the proportion of 70%:15%:15%.

[0025] Step 4: Build and train a convolutional bidirectional time series model; Step 4.1: Construct a convolutional bidirectional time series model for dynamometer data prediction. The convolutional bidirectional time series model mainly consists of a one-dimensional convolutional layer, two bidirectional LSTM (long short-term memory network) layers, and a time-distributed fully connected layer. The specific working process of the convolutional bidirectional time series model is as follows: Step 4.1.1, the one-dimensional convolutional layer processes the input data and determines the initial target features; Step 4.1.2, input the initial target features into the first bidirectional LSTM layer to confirm the first time series target features; Step 4.1.3, input the first time series target feature into the second bidirectional LSTM layer to confirm the second time series target feature; Step 4.1.4: Input the second time series target feature into the time distribution fully connected layer to obtain the prediction result of the future dynamometer data.

[0026] The parameters and specific settings of the convolutional bidirectional timing model are as follows: The number of convolution kernels is 128; the convolution kernel size is 3; the stride size is 1; the padding method is causal padding; the number of units in the first bidirectional LSTM layer is 128; the number of units in the second bidirectional LSTM layer is 128; the normalization method is MaxMin normalization (deviation normalization); the optimizer is Adam; the loss function is mean square error MSE; the number of iterations is 50; and the training batch is 16.

[0027] The input dimension of the model is ( ,28,200), where is the number of samples, 28 is the time step, 200 is the feature dimension, and the data dimension of the input layer after a one-dimensional convolution becomes ( ,28,256), and then pass through two bidirectional LSTM layers in sequence. Since the load of the next 200 points needs to be output in the end, a time-distributed fully connected layer with 200 neurons is added for each time step to obtain the final output.

[0028] Step 4.2: Input the training set divided in step 3.5 into the model for model training, and then evaluate the model performance based on the test set and the validation set to check the generalization ability of the model; the present invention uses the mean absolute error (MAE) and the relative mean error (MRE) to evaluate the prediction accuracy of the convolutional bidirectional time series model, and the formulas are: (5); (6); in, is the mean absolute error; is the relative average error; For the The true value of samples; For the The predicted value of samples; Calculate the average absolute error and relative average error of the training set, validation set, and test set respectively, and then compare them; the smaller the average absolute error value, the better the prediction effect of the model; when the relative average errors of the test set and validation set are both smaller than or slightly larger than the relative average error of the training set, it indicates that the network model can be output; when any relative average error in the test set and validation set is much larger than the relative average error of the training set, it indicates that the model may be overfitted and cannot be generalized well to new data, and it is necessary to return to step 1 to collect data again for model training. The judgment standard for slightly larger than is that the difference between the relative average error of the test set and the relative average error of the training set is less than or equal to the preset threshold; the judgment standard for much larger than is that the difference between the relative average error of the test set and the relative average error of the training set is greater than the preset threshold.

[0029] Step 5: Apply the convolutional bidirectional time series model trained in step 4 to the prediction of the pumping unit indicator diagram at the oil well site. The indicator diagram data of the pumping unit in the oil field is obtained in real time, and the load-specific data of the indicator diagram is obtained after processing according to the process of steps 1 and 2. The load-specific data of the indicator diagram is input into the trained model to calculate the future indicator diagram of the pumping unit in real time, thereby realizing the continuous prediction of the indicator diagram of the pumping unit.

[0030] In order to prove the feasibility of the method of the present invention, the following experiment was carried out.

[0031] The experimental data comes from a block in an oil field. The block to be studied has several oil wells and contains data under several different production systems.

[0032] Through steps 2 and 3, the historical production data of the oil wells in the block were sorted out to obtain a total of 6982 valid data sets. The time series sample sets obtained after processing were randomly divided into training sets, verification sets and test sets according to a certain ratio (training set: verification set: test set = 70%: 15%: 15%), including 4886 training set data, 1048 verification set data and 1048 test set data. Some historical displacement data of oil wells in a certain block are shown in Table 1.

[0033] Table 1 Some historical displacement data of oil wells in a certain block .

[0034] As shown in Table 1, the length of displacement data of each well is consistent, but the length of displacement data between each well is inconsistent. First, use formula (1) to splice the downstroke data in reverse order after the upstroke data.

[0035] Some historical load data of oil wells in a certain block are shown in Table 2.

[0036] Table 2 Some historical load data of oil wells in a certain block .

[0037] As shown in Table 2, it can be seen that the length of the load data of each well is consistent with the length of the displacement data. After the fixed difference displacement point is determined by formula (2), the load data at the fixed difference displacement point is interpolated, and the new displacement load data is used to replace the original displacement load data. Then, the displacement data and load data are normalized by formula (3) and formula (4) respectively. The normalized displacement data are the same, so only the normalized load data needs to be predicted. The load data of the historical 28-day dynamometer diagram is used as the input of the model, and the model outputs the dynamometer diagram for the next 28 days. The training set is input into the model, and the model is trained to obtain the trained convolutional bidirectional time series model. The training set, validation set, and test set are respectively input into the trained model, and their MAE and MRE are calculated using formula (5) and formula (6), respectively. The results are shown in Table 3.

[0038] Table 3. Mean absolute error and relative mean error results of training set, validation set and test set .

[0039] It can be seen from Table 3 that the relative mean error (MRE) between the validation set and the test set is smaller than the relative mean error of the training set, indicating that the network has good generalization ability. The trained convolutional bidirectional time series model can be used to predict the future dynamometer diagram of the oil well pumping unit.

[0040] Some results calculated based on the convolutional bidirectional timing model are as follows: Figure 2 , Figure 3 As shown, Figure 2 It is the prediction result of the future indicator diagram under complex working conditions dominated by gas influence using the convolution bidirectional time series model. Figure 3 It is the prediction result of the future dynamometer diagram under complex working conditions dominated by wax deposition using the convolution bidirectional time series model. It can be seen from both figures that the prediction of the dynamometer diagram data for the next 28 days based on the historical dynamometer diagram data for the past 28 days maintains the trend of working condition changes. This trend can provide reference information for the oil field in judging the severity of subsequent working conditions.

[0041] Table 4. Partial training rounds of training set and validation set and the corresponding mean absolute error results .

[0042] The training rounds and the corresponding mean absolute error (MAE) obtained by calculating the training set and the validation set are shown in Table 4 (due to the large amount of data, Table 4 only shows part of the data); draw their line graphs, as shown in Figure 4 shown.

[0043] Combining Table 4 and Figure 4 The mean absolute error of the training set continued to decrease, indicating that the model was learning the training data, and the gap between the mean absolute error of the training and validation sets gradually narrowed. The mean absolute error of the validation set finally stabilized at a low level (0.0175), indicating that the model eventually stabilized and had practical application potential. The load data calculated by the model was very close to the actual load data, which could meet the needs of oilfield working condition diagnosis.

[0044] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A convolution bidirectional timing indicator diagram prediction method based on specific data drive, characterized in that: The steps include: Step 1, obtaining the dynamometer diagram data, which records the dynamic correspondence between the polished rod load and the suspension point displacement; Step 2: Clean, align and normalize the dynamometer data. The data obtained after normalization is the load-specific data of the dynamometer; Step 3: construct a time series sample set based on the load-specific data of the dynamometer diagram; Step 4: Build and train a convolutional bidirectional time series model; Step 5: Apply the trained convolutional bidirectional time series model to the prediction of the pumping unit indicator diagram at the oil well site.

2. The convolution bidirectional timing indicator diagram prediction method based on specific data drive according to claim 1 is characterized in that: The specific process of step 2 is: Step 2.1, clean the dynamometer data; the specific process is: first perform non-null value integrity check, remove the well condition records with null values ​​in the load displacement data dimension, and then obtain the dynamometer time series data set with block ownership, well group ownership, acquisition time and time series characteristics through directional extraction; Step 2.2, align the dynamometer data; The specific process is: Step 2.2.1, divide the displacement load data points of the dynamometer diagram into upstroke data and downstroke data, and splice the downstroke data in reverse order after the upstroke data. The calculation formula is as follows: (1)； in, It is the displacement point of the downstroke after splicing; is the maximum displacement value of the dynamometer diagram; It is the displacement point of the downstroke before splicing; Step 2.2.2, determine the fixed difference displacement point, keep the number of data points before and after interpolation unchanged, the formula is: (2)； in, is the fixed difference displacement point; is the number of data points after data alignment; A function that generates a specific sequence of values; Step 2.2.3, use cubic spline interpolation to interpolate the original data, by fitting a cubic polynomial between every two adjacent displacement load points, ensuring that the first-order derivative and the second-order derivative of each segment of the polynomial are continuous at the connection point, and satisfying the natural boundary condition that the second-order derivative at the endpoint is zero, thereby generating a smooth curve, and then obtaining the load data at the fixed differential displacement point, and replacing the original displacement load data with the displacement load data obtained by interpolation, to obtain new dynamometer diagram data of the same differential displacement point; Step 2.3: normalize the new dynamometer data. The data obtained after normalization is the load-specific data of the dynamometer. The specific formula is as follows: (3)； (4)； in, is the normalized displacement; is the actual displacement; is the minimum displacement value of the dynamometer diagram; is the normalized load value; is the actual load value; , are the maximum load value and the minimum load value respectively.

3. The convolution bidirectional timing indicator diagram prediction method based on specific data drive according to claim 2 is characterized in that: The specific process of step 3 is as follows: Step 3.1, group the load-specific data of the dynamometer diagram according to blocks and well groups, ensure that the data of the well group in each block is processed independently, and take a section of time series data of a well in a block as a sample; Step 3.2, sort the data in each block and well group according to the acquisition time; Step 3.3, using the sliding window sampling method, extract the data pairs of the historical 28 days and the future 28 days from the data of each block and well group; the specific process of the sliding window sampling method is: starting from the starting position of the time series data, select windows of length 56 in sequence with a step size of 1 until a complete window can no longer be formed; for each window, the first 28 data points are used as historical data, and the last 28 data points are used as future data, so as to extract the data pairs of the historical 28 days and the future 28 days; Step 3.4: Collect the data pairs of all blocks to obtain the final time series sample set; Step 3.5: Divide the final time series sample set into training set, validation set and test set in proportion.

4. The convolution bidirectional timing indicator diagram prediction method based on specific data drive according to claim 3 is characterized in that: The specific process of step 4 is as follows: Step 4.1: Construct a convolutional bidirectional time series model for dynamometer data prediction; the convolutional bidirectional time series model consists of a one-dimensional convolutional layer, two bidirectional long short-term memory network layers, and a time-distributed fully connected layer; Step 4.2: Input the training set divided in step 3.5 into the model for model training, and then evaluate the model performance based on the test set and validation set to check the generalization ability of the model.

5. The convolution bidirectional timing indicator diagram prediction method based on specific data drive according to claim 4 is characterized in that: In step 4.1, the specific working process of the convolutional bidirectional time series model is: Step 4.1.1, the one-dimensional convolutional layer processes the input data and determines the initial target features; Step 4.1.2, input the initial target feature into the first bidirectional long short-term memory network layer to confirm the first temporal target feature; Step 4.1.3, input the first time series target feature into the second bidirectional long short-term memory network layer to confirm the second time series target feature; Step 4.1.4: Input the second time series target feature into the time distribution fully connected layer to obtain the prediction result of the future dynamometer data.

6. The method for predicting a convolutional bidirectional timing dynamometer diagram based on specific data drive according to claim 5, characterized in that: In step 4.2, the prediction accuracy of the convolutional bidirectional time series model is evaluated using the mean absolute error and the relative mean error, and the formulas are: (5)； (6)； in, is the mean absolute error; is the relative average error; For the The true value of samples; For the The predicted value of samples; is the sample size; Calculate the mean absolute error and relative mean error of the training set, validation set, and test set respectively, and then compare them. Output the model that meets the specified conditions; otherwise, re-collect data in step 1 for model training; The specified conditions are: when the relative average errors of the test set and the validation set are both smaller than the relative average error of the training set, or the difference between the relative average error of the test set and the relative average error of the training set is less than or equal to a preset threshold.

7. The convolution bidirectional timing indicator diagram prediction method based on specific data drive according to claim 6 is characterized in that: In step 5, the dynamometer diagram data of the oil pumping unit in the oil field is obtained in real time, and is processed according to the process of steps 1 and 2 to obtain the dynamometer diagram load specific data, and the dynamometer diagram load specific data is input into the trained model to calculate the future dynamometer diagram of the oil pumping unit in real time, thereby realizing the continuous prediction of the dynamometer diagram of the oil pumping unit.

Images