Fracturing sand plugging risk early warning method based on ensemble learning and slope inversion method

Abstract

The invention discloses a fracturing sand plugging risk early warning method based on ensemble learning and a slope inversion method. The method comprises the steps of obtaining preprocessed time sequence data; obtaining a plurality of new feature parameters; constructing and training an original deep learning model to obtain a sand blockage prediction single algorithm model, and performing integrated learning on the sand blockage prediction single algorithm model and an XGBoost model by using a Stacking integration method to obtain a sand blockage prediction model, model output oil pressure and prediction time sequence data of casing pressure; and based on the predicted time sequence data of the oil pressure and the casing pressure, judging whether sand blockage occurs or not by using a slope inversion method. According to the method, deep learning and a slope inversion method are combined, so that the early warning effect of the sand blockage accident is remarkably improved. Compared with an existing method for manually observing a curve and manually judging sand blockage, the method solves the problems of lagging, accident, mechanical property and the like, and a certain effect is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of oilfield fracturing sand plugging risk early warning, and in particular relates to a fracturing sand plugging risk early warning method based on ensemble learning and slope inversion method. Background Art

[0002] Fracturing technology is a commonly used method in the development and utilization of oil resources. It cracks underground oil reservoirs and forms fracture channels by injecting high-pressure fracturing fluid, thereby increasing the production capacity of oil wells. During construction, due to the rupture of the formation and the cracks generated, some formation mineral particles may enter the cracks. As the fracturing fluid decreases or stops, these mineral particles may be deposited inside the cracks, forming sand plugs, which will seriously affect the fracturing effect and even cause major accidents. The evolution of sand plugs is complex and nonlinear, so it is necessary to conduct a systematic analysis of sand plugs in fracturing construction, extract accident characteristics, and predict sand plugs to reduce accidents.

[0003] The current fracturing operations have the following problems: (1) In oilfield fracturing operations, there are characteristics of large-scale single construction, long time and many equipment and facilities. At the same time, the large-scale operation of small well sites also makes the on-site operation of the well site more complicated. In this case, factors such as personnel operating errors and unreasonable process design can easily lead to sand plugging. (2) A lot of research work has been done at home and abroad on fracturing construction monitoring, fracturing construction accident analysis and early warning. One method is to understand the downhole fracturing construction conditions by monitoring the changes in the fracturing construction curve. However, this method has lag and is too dependent on the experience and judgment ability of on-site construction technicians. (3) The fracturing effect is affected by dynamic factors such as construction process parameters and crack changes, and the fracturing effect is difficult to estimate. Therefore, it is necessary to develop intelligent fracturing technology based on fracturing transformation technology to achieve high-precision control of fracturing construction.

[0004] At present, the development direction of safety early warning models is mainly divided into three types: monitoring parameter early warning model, machine learning early warning model and risk early warning model. Among them, the monitoring parameter early warning model needs to set a threshold according to the actual situation of the fracturing construction site. When the monitoring parameter exceeds the threshold, the system will issue an alarm. However, this method is greatly affected by experience and subjective factors, and is prone to missed alarms and false alarms. The machine learning early warning model needs to collect a large amount of monitoring data to train a classifier or regression model to evaluate the current working conditions. However, this method requires a large number of data sets, and the model needs to be maintained and updated, and the cost of use is high. The risk early warning model analyzes the accident risk from the perspective of uncertainty, including pre-hazard analysis, accident tree, hierarchical analysis method and fuzzy logic, but these methods usually cannot achieve real-time safety warning and are not effective in practical applications. Therefore, it is extremely important to study and seek a fracturing sand plugging early warning method that can meet the actual needs of on-site fracturing construction, rely on existing fracturing construction site equipment, and be suitable for on-site practical applications, and truly achieve early warning of fracturing sand plugging. Summary of the invention

[0005] In order to solve the above problems, the object of the present invention is to provide a fracturing sand plugging risk early warning method based on ensemble learning and slope reversal method.

[0006] In order to achieve the above object, the method for early warning of fracturing sand plugging risk based on ensemble learning and slope reversal method provided by the present invention comprises the following steps performed in sequence:

[0007] 1) Collect the original data of multiple parameters including oil pressure and casing pressure during construction and construct time series data, taking oil pressure and casing pressure as target parameters and other parameters as feature parameters, and then preprocess the time series data of feature parameters to obtain multiple preprocessed time series data;

[0008] 2) Using wavelet transform and empirical mode decomposition method to process the above preprocessed time series data to obtain multiple new feature parameters;

[0009] 3) Constructing an original deep learning model, inputting the time series data of all the above new feature parameters and target parameters into the above original deep learning model for model training, using the sparrow algorithm SSA to adjust and optimize parameters during training, obtaining a single algorithm model for sand plugging prediction, and then using the Stacking integration method to integrate the above single algorithm model for sand plugging prediction and the XGBoost model to obtain a sand plugging prediction model, and finally the sand plugging prediction model outputs the predicted time series data of oil pressure and casing pressure;

[0010] 4) Based on the predicted time series data of the oil pressure and casing pressure, the slope reversal method is used to determine whether sand plugging occurs.

[0011] In step 1), the original data of multiple parameters including oil pressure and casing pressure during construction are collected and time series data are constructed, with oil pressure and casing pressure as target parameters and other parameters as characteristic parameters, and then the time series data of the characteristic parameters are preprocessed respectively to obtain multiple preprocessed time series data.

[0012] The original data of seven parameters, including oil pressure, casing pressure, discharge flow, sand concentration, stage liquid displacement, total liquid volume and stage cumulative liquid volume, were collected during construction. The oil pressure and casing pressure were used as target parameters, and the other five parameters were used as characteristic parameters. Then, time series data were constructed with time as the X-axis and parameter value as the Y-axis. Then, the difference method was used to fill the missing values, and the data deduplication method was used to remove the duplicate values. The clustering algorithm was used to divide the time series data into multiple clusters, and each cluster was statistically analyzed to detect and process outliers. Finally, normalization was performed to solve the problem of different data value ranges. The normalization formula is:

[0013]

[0014] where Y i represents the normalized data of the ith parameter, X i Represents the original data of the i-th parameter, X imax and X imin Respectively represent the maximum and minimum value of the i-th parameter.

[0015] In step 2), the specific steps of processing the pre-processed time series data using wavelet transform and empirical mode decomposition method to obtain multiple new feature parameters are as follows:

[0016] 2.1) Select db4 wavelet to decompose the preprocessed time series data of the above-mentioned characteristic parameters into 4 layers, complete the data denoising, obtain 4 detail components with the same length as the original data, and then reconstruct and merge these detail components after threshold processing to obtain reconstructed time series data;

[0017] 2.2) The above reconstructed time series data is decomposed using the empirical mode decomposition method to obtain multiple essential mode function IMF components; each essential mode function IMF component represents each frequency component in the original data, and then is arranged in sequence from high frequency to low frequency to obtain a spectrum; then each frequency component in the spectrum is evaluated and screened, and those low-frequency essential mode function IMF components with higher energy, less noise and related to the original fracturing operation curve are retained, and those high-frequency essential mode function IMF components that are greatly affected by noise are removed, because noise will greatly interfere with the system.

[0018] 2.3) Select dozens of low-frequency essential mode function IMF components with strong information as the initial new feature parameters;

[0019] 2.4) Perform sensitivity analysis on the correlation between the above initial new feature parameters and the target parameters to achieve the purpose of dimensionality reduction and feature screening, and finally screen out several new feature parameters.

[0020] In step 3), the original deep learning model is constructed, and the time series data of all the new feature parameters and the target parameters are input into the original deep learning model for model training. During the training, the sparrow algorithm SSA is used to adjust and optimize the parameters to obtain a single algorithm model for sand plugging prediction. Then, the stacking integration method is used to integrate the single algorithm model for sand plugging prediction and the XGBoost model to obtain a sand plugging prediction model. Finally, the specific steps of the sand plugging prediction model outputting the predicted time series data of oil pressure and casing pressure are as follows:

[0021] 3.1) The long short-term memory model LSTM is used as the original deep learning model, and the time series data of all new feature parameters and target parameters within two minutes before the prediction point are used as the input data of the long short-term memory model LSTM. The above long short-term memory model LSTM is trained. During the training process, the sparrow algorithm SSA is used to search for the optimal hyperparameters of the model to improve the accuracy of the parameters, and the three parameters of the number of hidden neurons, learning rate and training times are optimized, so that the original deep learning model can obtain higher-precision oil pressure and casing pressure prediction values, and obtain a single algorithm model for sand plugging prediction;

[0022] 3.2) Use the Stacking integration method to integrate the above-mentioned sand plugging prediction single algorithm model and XGBoost model to obtain a sand plugging prediction model; the output of the sand plugging prediction model is the predicted time series data of oil pressure and casing pressure in the next minute;

[0023] 3.3) The sand plugging prediction model is cross-validated with 5-fold using the above input data to reduce the overfitting of the model to a certain extent, and the prediction results and newly generated data are continuously added to the next stage of sand plugging prediction model training, gradually realizing the learning and prediction of the future prediction stage, and thus obtaining more accurate prediction data.

[0024] In step 4), the method for judging whether sand plugging occurs by using the slope reversal method based on the predicted time series data of the oil pressure and casing pressure is:

[0025] The slopes calculated from the oil pressure and casing pressure data in the historical stable sand adding stage before the start of the sand plugging prediction model are used as the respective alarm thresholds. The slopes of the oil pressure and casing pressure at each point in the predicted time series data of the above oil pressure and casing pressure are calculated. When the calculated absolute value of the slope exceeds the corresponding alarm threshold, it indicates that sand plugging is about to occur, reminding the construction personnel to immediately take measures such as reducing the displacement and reducing the sand ratio to avoid sand plugging and improve construction efficiency and safety.

[0026] The fracturing sand plugging risk warning method based on integrated learning and slope reversal method provided by the present invention has the following beneficial effects: analyzing and calculating the oil pressure, casing pressure, discharge flow, sand concentration data, etc. that are easy to collect during the fracturing construction process, using the Stacking integration method to integrate the LSTM and XGBoost models, respectively predicting the values ​​of oil pressure and casing pressure for time series data, and using the sparrow algorithm SSA to optimize the three parameters of LSTM. The slopes of the oil pressure and casing pressure are calculated respectively and compared with the alarm threshold, thereby realizing intelligent early warning of sand plugging accidents at the fracturing construction site. The present invention adopts a method combining deep learning and slope reversal method to significantly improve the early warning effect of sand plugging accidents. Compared with the existing method of manually observing curves and manually identifying sand plugging, this method solves the problems of hysteresis, accidental and mechanical, and has achieved certain results. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flow chart of the method for early warning of fracturing sand plugging risk based on ensemble learning and slope reversal method provided by the present invention. DETAILED DESCRIPTION

[0028] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] like Figure 1 As shown, the fracturing sand plugging risk early warning method based on ensemble learning and slope reversal method provided by the present invention includes the following steps performed in sequence:

[0030] 1) Collect the original data of multiple parameters including oil pressure and casing pressure during construction and construct time series data, taking oil pressure and casing pressure as target parameters and other parameters as feature parameters, and then preprocess the time series data of feature parameters to obtain multiple preprocessed time series data;

[0031] The original data of seven parameters, including oil pressure, casing pressure, discharge flow, sand concentration, stage liquid displacement, total liquid volume and stage cumulative liquid volume, were collected during construction. The oil pressure and casing pressure were used as target parameters, and the other five parameters were used as characteristic parameters. Then, time series data were constructed with time as the X-axis and parameter value as the Y-axis. Then, the difference method was used to fill the missing values, and the data deduplication method was used to remove the duplicate values. The clustering algorithm was used to divide the time series data into multiple clusters, and each cluster was statistically analyzed to detect and process outliers. Finally, normalization was performed to solve the problem of different data value ranges. The normalization formula is:

[0032]

[0033] where Y i represents the normalized data of the i-th parameter, X i Represents the original data of the i-th parameter, X imax and X imin Respectively represent the maximum and minimum values ​​of the i-th parameter;

[0034] 2) Using wavelet transform and empirical mode decomposition method to process the above preprocessed time series data to obtain multiple new feature parameters;

[0035] The specific steps are as follows:

[0036] 2.1) Select db4 wavelet to decompose the preprocessed time series data of the above-mentioned characteristic parameters into 4 layers, complete the data denoising, obtain 4 detail components with the same length as the original data, and then reconstruct and merge these detail components after threshold processing to obtain reconstructed time series data;

[0037] For example, decomposing the discharge flow into four layers can obtain detail components in level 1, detail components in level 2, detail components in level 3 and detail components in level 4; these detail components are reconstructed and merged after threshold processing, which can better reflect the changing trend of the discharge volume, is closer to the time series data of the actual discharge volume, removes noise interference, and is smoother.

[0038] 2.2) Empirical Mode Decomposition (EMD) is a signal processing and data analysis method used to decompose complex nonlinear and non-stationary signals into a set of Intrinsic Mode Functions (IMFs).

[0039] The above reconstructed time series data are decomposed using the empirical mode decomposition method to obtain multiple essential mode function IMF components; each essential mode function IMF component represents each frequency component in the original data, and then they are arranged in sequence from high frequency to low frequency to obtain a spectrum; then each frequency component in the spectrum is evaluated and screened, and those low-frequency essential mode function IMF components with higher energy, less noise and related to the original fracturing operation curve are retained, and those high-frequency essential mode function IMF components that are greatly affected by noise are removed, because noise will greatly interfere with the system.

[0040] 2.3) Select dozens of low-frequency essential mode function IMF components with strong information as the initial new feature parameters;

[0041] 2.4) Perform sensitivity analysis on the correlation between the above initial new feature parameters and the target parameters to achieve the purpose of dimensionality reduction and feature screening, and finally screen out several new feature parameters.

[0042] 3) Constructing an original deep learning model, inputting the time series data of all the above new feature parameters and target parameters into the above original deep learning model for model training, using the sparrow algorithm SSA to adjust and optimize parameters during training, obtaining a single algorithm model for sand plugging prediction, and then using the Stacking integration method to integrate the above single algorithm model for sand plugging prediction and the XGBoost model to obtain a sand plugging prediction model, and finally the sand plugging prediction model outputs the predicted time series data of oil pressure and casing pressure;

[0043] The specific steps are as follows:

[0044] 3.1) The long short-term memory model LSTM is used as the original deep learning model, and the time series data of all new feature parameters and target parameters within two minutes before the prediction point are used as the input data of the long short-term memory model LSTM. The above long short-term memory model LSTM is trained. During the training process, the sparrow algorithm SSA is used to search for the optimal hyperparameters of the model to improve the accuracy of the parameters, and the three parameters of the number of hidden neurons, learning rate and training times are optimized, so that the original deep learning model can obtain higher-precision oil pressure and casing pressure prediction values, and obtain a single algorithm model for sand plugging prediction;

[0045] 3.2) In order to improve the prediction accuracy of the single algorithm model for sand plugging prediction, the Stacking integration method is used to integrate the single algorithm model for sand plugging prediction and the XGBoost model to obtain a sand plugging prediction model; the output of the sand plugging prediction model is the predicted time series data of oil pressure and casing pressure in the next minute;

[0046] 3.3) The sand plugging prediction model is cross-validated with 5-fold using the above input data to reduce the overfitting of the model to a certain extent, and the prediction results and newly generated data are continuously added to the next stage of sand plugging prediction model training, gradually realizing the learning and prediction of the future prediction stage, and thus obtaining more accurate prediction data.

[0047] 4) Based on the predicted time series data of the oil pressure and casing pressure, the slope reversal method is used to determine whether sand plugging occurs.

[0048] The slope reversal method, also known as the reverse slope method, is an empirical method that can be used to predict sand plugging in real time. The principle is: after the pre-fluid is pumped in, proppant is added, causing the liquid column pressure in the wellbore to gradually increase, and the oil pressure and casing pressure gradually decrease at a relatively constant slope. Once the entire wellbore is filled with sand-carrying fluid, the liquid column pressure reaches the maximum, and the oil pressure and casing pressure stop decreasing. This process is reflected in the construction curve as a negative slope, and the pressure decreases over time. As the construction continues, the pressure remains unchanged for a short time, or begins to increase gradually. When the pressure begins to rise gradually, the slope after the reversal is almost the same as the slope when the wellbore was previously loaded, which is reflected in the construction curve as a positive slope. If the positive slope begins to deviate from the reverse slope, it is an obvious sign of sand plugging in the near-wellbore area. This sign tells the construction personnel that sand plugging is about to occur, and prompts them to replace it immediately to avoid sand plugging.

[0049] The slopes calculated from the oil pressure and casing pressure data in the historical stable sand adding stage before the start of the sand plugging prediction model are used as the respective alarm thresholds. The slopes of the oil pressure and casing pressure at each point in the predicted time series data of the above oil pressure and casing pressure are calculated. When the calculated absolute value of the slope exceeds the corresponding alarm threshold, it indicates that sand plugging is about to occur, reminding the construction personnel to immediately take measures such as reducing the displacement and reducing the sand ratio to avoid sand plugging and improve construction efficiency and safety.

Claims

1. A method for early warning of fracturing sand plugging risk based on ensemble learning and slope reversal method, characterized by: The method comprises the following steps performed in sequence: 1) Collect the original data of multiple parameters including oil pressure and casing pressure during construction and construct time series data, taking oil pressure and casing pressure as target parameters and other parameters as feature parameters, and then preprocess the time series data of feature parameters to obtain multiple preprocessed time series data; 2) Using wavelet transform and empirical mode decomposition method to process the above preprocessed time series data to obtain multiple new feature parameters; 3) Constructing an original deep learning model, inputting the time series data of all the above new feature parameters and target parameters into the above original deep learning model for model training, using the sparrow algorithm SSA to adjust and optimize parameters during training, obtaining a single algorithm model for sand plugging prediction, and then using the Stacking integration method to integrate the above single algorithm model for sand plugging prediction and the XGBoost model to obtain a sand plugging prediction model, and finally the sand plugging prediction model outputs the predicted time series data of oil pressure and casing pressure; 4) Based on the predicted time series data of the oil pressure and casing pressure, the slope reversal method is used to determine whether sand plugging occurs.

2. The method for early warning of fracturing sand plugging risk based on ensemble learning and slope inversion method according to claim 1 is characterized by: In step 1), the original data of multiple parameters including oil pressure and casing pressure during construction are collected and time series data are constructed, with oil pressure and casing pressure as target parameters and other parameters as characteristic parameters, and then the time series data of the characteristic parameters are preprocessed respectively to obtain multiple preprocessed time series data. The original data of seven parameters, including oil pressure, casing pressure, discharge flow, sand concentration, stage liquid discharge, total liquid volume and stage cumulative liquid volume, were collected during construction. The oil pressure and casing pressure were used as target parameters, and the other five parameters were used as characteristic parameters. Then, time series data were constructed with time as the X-axis and parameter value as the Y-axis. Then, the difference method is used to fill the missing values, and the data deduplication method is used to remove duplicate values; The clustering algorithm is used to divide the time series data into multiple clusters, and each cluster is statistically analyzed to detect and process outliers; finally, normalization is performed to solve the problem of different data value ranges. The normalization formula is: where Y i represents the normalized data of the ith parameter, X i Represents the original data of the i-th parameter, X imax and X imin Respectively represent the maximum and minimum value of the i-th parameter.

3. The method for early warning of fracturing sand plugging risk based on ensemble learning and slope inversion method according to claim 1 is characterized by: In step 2), the specific steps of processing the pre-processed time series data using wavelet transform and empirical mode decomposition method to obtain multiple new feature parameters are as follows: 2.1) Select db4 wavelet to decompose the preprocessed time series data of the above-mentioned characteristic parameters into 4 layers, complete the data denoising, obtain 4 detail components with the same length as the original data, and then reconstruct and merge these detail components after threshold processing to obtain reconstructed time series data; 2.2) The above reconstructed time series data is decomposed using the empirical mode decomposition method to obtain multiple essential mode function IMF components; each essential mode function IMF component represents each frequency component in the original data, and then is arranged in sequence from high frequency to low frequency to obtain a spectrum; then each frequency component in the spectrum is evaluated and screened, and those low-frequency essential mode function IMF components with higher energy, less noise and related to the original fracturing operation curve are retained, and those high-frequency essential mode function IMF components that are greatly affected by noise are removed, because noise will greatly interfere with the system. 2.3) Select dozens of low-frequency essential mode function IMF components with strong information as the initial new feature parameters; 2.4) Perform sensitivity analysis on the correlation between the above initial new feature parameters and the target parameters to achieve the purpose of dimensionality reduction and feature screening, and finally screen out several new feature parameters.

4. The method for early warning of fracturing sand plugging risk based on ensemble learning and slope inversion method according to claim 1 is characterized by: In step 3), the original deep learning model is constructed, and the time series data of all the new feature parameters and the target parameters are input into the original deep learning model for model training. During the training, the sparrow algorithm SSA is used to adjust and optimize the parameters to obtain a single algorithm model for sand plugging prediction. Then, the stacking integration method is used to integrate the single algorithm model for sand plugging prediction and the XGBoost model to obtain a sand plugging prediction model. Finally, the specific steps of the sand plugging prediction model outputting the predicted time series data of oil pressure and casing pressure are as follows: 3.1) The long short-term memory model LSTM is used as the original deep learning model, and the time series data of all new feature parameters and target parameters within two minutes before the prediction point are used as the input data of the long short-term memory model LSTM. The above long short-term memory model LSTM is trained. During the training process, the sparrow algorithm SSA is used to search for the optimal hyperparameters of the model to improve the accuracy of the parameters, and the three parameters of the number of hidden neurons, learning rate and training times are optimized, so that the original deep learning model can obtain higher-precision oil pressure and casing pressure prediction values, and obtain a single algorithm model for sand plugging prediction; 3.2) Use the Stacking integration method to integrate the above-mentioned sand plugging prediction single algorithm model and XGBoost model to obtain a sand plugging prediction model; the output of the sand plugging prediction model is the predicted time series data of oil pressure and casing pressure in the next minute; 3.3) The sand plugging prediction model is cross-validated with 5-fold using the above input data to reduce the overfitting of the model to a certain extent, and the prediction results and newly generated data are continuously added to the next stage of sand plugging prediction model training, gradually realizing the learning and prediction of the future prediction stage, and thus obtaining more accurate prediction data.

5. The method for early warning of fracturing sand plugging risk based on ensemble learning and slope inversion method according to claim 1, characterized in that: In step 4), the method for judging whether sand plugging occurs based on the predicted time series data of the oil pressure and casing pressure by using the slope reversal method is: The slopes calculated from the oil pressure and casing pressure data in the historical stable sand adding stage before the start of the sand plugging prediction model are used as the respective alarm thresholds. The slopes of the oil pressure and casing pressure at each point in the predicted time series data of the above oil pressure and casing pressure are calculated. When the calculated absolute value of the slope exceeds the corresponding alarm threshold, it indicates that sand plugging is about to occur, reminding the construction personnel to immediately take measures such as reducing the displacement and reducing the sand ratio to avoid sand plugging and improve construction efficiency and safety.

Images