Gas well liquid drop entrainment rate prediction method based on big data
By using a big data-based method for predicting droplet entrainment rate, and employing data cleaning and multi-model fusion techniques, the applicability and accuracy of droplet entrainment rate prediction were addressed, achieving efficient and accurate prediction under different operating conditions.
Patent Information
- Application Number
- CN202511848235.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-09
- Publication Date
- 2026-03-31
AI Technical Summary
Existing droplet entrainment rate prediction methods have limited applicability and low prediction accuracy, making it difficult to accurately predict gas well liquid accumulation under different operating conditions.
By collecting experimental data on droplet entrainment rate in gas wells, we conducted collaborative verification and residual feedback preprocessing, grouped the data based on fluid physics mechanism characteristics, and used physical information neural networks, genetic programming, and random forest algorithms to perform multi-model fusion and dynamically weighted fusion of the final results.
It achieves accurate and efficient prediction of droplet entrainment rate under a wide range of operating conditions, overcomes the limitations of traditional methods in terms of applicability and high computational cost, shortens prediction time, and improves prediction accuracy.
Smart Images
Figure CN121765240A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of gas well production technology, and in particular relates to a method for predicting the entrainment rate of gas well droplets based on big data. Background Technology
[0002] Liquid accumulation in gas wells occurs because, as gas well production progresses, formation pressure decreases, gas production gradually declines, and the reduced gas flow velocity is insufficient to carry the liquid to the surface, leading to liquid accumulation in the wellbore. Currently, most gas fields in my country have entered the mid-to-late stages of development, and water production significantly impacts well productivity. Low pressure and low production in gas wells result in extremely severe wellbore liquid accumulation problems. The droplet entrainment rate is a key parameter for predicting gas well liquid accumulation. Studies show that during normal liquid production, the flow pattern within the wellbore is mostly annular with high gas flow velocity. At this time, a liquid film adheres to the wellbore wall, and some droplets are carried into the gas core by the gas. These droplets affect the gas core density and, together with the gas-liquid interface friction coefficient, influence the calculation of the gas-liquid interface shear force, thus affecting the stress on the liquid film on the wellbore wall. Therefore, accurately predicting the gas well droplet entrainment rate is of great significance for predicting wellbore liquid accumulation.
[0003] Several calculation methods have been proposed in the industry for predicting droplet entrainment rates. These include the earliest Paleev & Filippovic formula, and improved versions such as the Wallis, Oliemans, and Ishii & Mishima formulas, most of which use the Weber number as a characteristic parameter. These methods are largely empirical formulas based on fitting experimental data, often applicable only to specific conditions and with limited applicability. For example, droplet entrainment rate calculations using data fitted under low-pressure conditions are difficult to accurately predict droplet entrainment rates under high-pressure conditions, and vice versa. Summary of the Invention
[0004] To address the aforementioned shortcomings in existing technologies, this invention provides a big data-based method for predicting the entrainment rate of gas well droplets, which solves the problems of limited applicability and low prediction accuracy of existing droplet entrainment rate prediction methods.
[0005] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a method for predicting the entrainment rate of gas well droplets based on big data, comprising: Collect experimental data on droplet entrainment rate in gas wells, summarize the influencing factors considered under different experimental conditions, and form an initial dataset. The initial dataset is preprocessed with co-validation and residual feedback to obtain a clean preprocessed dataset. Grouping of fluid physics mechanism features based on clean datasets; Droplet entrainment rate is predicted based on different feature groups, and prediction results for multiple operating conditions are obtained. The prediction results of multiple operating conditions are dynamically weighted and fused based on the current operating conditions of the gas well to obtain the final droplet entrainment rate prediction result.
[0006] The beneficial effects of this invention are as follows: 1. By constructing a collaborative data cleaning architecture, an optimization effect of "maximizing strengths and minimizing weaknesses" is achieved. Through a two-stage process: box plotting is used to initially identify statistically significant outliers; then, the Isolation Forest algorithm and residual threshold supervision are used to remove noise and outliers from the dataset, improving data quality. This mechanism enables the invention to effectively clean the dataset while ensuring that the final model's prediction accuracy does not decrease but rather increases, overcoming the drawbacks of traditional methods where cleaning and modeling are disconnected, making it difficult to balance efficiency and accuracy.
[0007] 2. By integrating gas-liquid phase operating parameters, dimensionless parameters, and system geometric characteristics, the relevant data are divided into entrainment dynamic characteristics, entrainment impedance characteristics, and system geometric and state characteristics. Different models are used for training to establish a nonlinear mapping relationship between input feature parameters and droplet entrainment rate. The complex nonlinear relationship in high-dimensional data is explored, and accurate and efficient prediction of droplet entrainment rate is achieved under a wide range of operating conditions.
[0008] 3. The prediction method provided by this invention overcomes the strong dependence of traditional empirical formulas on specific flow patterns and operating conditions, thus exhibiting high universality. Furthermore, it avoids the high computational cost and long processing time caused by traditional CFD simulation methods that require calculating large numbers of grids. It achieves a reduction in prediction time from hours or even days to seconds while maintaining accuracy, significantly reducing manpower, material resources, and computing power. Attached Figure Description
[0009] Figure 1 A flowchart of a gas well droplet entrainment rate prediction method based on big data is provided for an embodiment. Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 This is a box plot corresponding to the 18 sets of experimental data in the example; Figure 7 A comparison chart of droplet entrainment rate and dataset was generated for the random forest model. Figure 8 This is a comparison chart of the calculated droplet entrainment rate and the Wolf experimental results. Detailed Implementation
[0010] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.
[0011] like Figure 1 As shown, a method for predicting the entrainment rate of gas well droplets based on big data includes the following steps: S1. Collect experimental data on the entrainment rate of gas well droplets, summarize the influencing factors considered under different experimental conditions, and form an initial dataset.
[0012] In an optional embodiment of the present invention, 18 sets of experimental data, totaling 930 experimental data points for vertical pipe annular flow, were collected. These data covered a wide range of experimental conditions, and the initial dataset included apparent gas flow velocity, gas density, gas Reynolds number, Weber number, apparent liquid flow number, liquid density, liquid dynamic viscosity, surface tension, liquid Reynolds number, pipe diameter, and pressure. The apparent liquid flow velocity was 0.0035–1 m / s, the apparent gas flow velocity was 0.8–116 m / s, the test pressure was 1–200 bara, and the pipe diameter was 5–127 mm.
[0013] S2. Perform co-validation and residual feedback preprocessing on the initial dataset to obtain a clean preprocessed dataset.
[0014] Based on the initial dataset, the box plot method is used to label the outlier variables in the initial dataset to obtain the first set of outlier variables; Based on the initial dataset, the Isolation Forest algorithm is used to identify initial outliers in the initial dataset; a pre-trained regression model is used to predict anomalies in the initial dataset, the relative error of each sample is calculated, and a residual threshold is set based on the relative error; data that exceed the residual threshold among the initial outliers are recorded as outlier variables, resulting in a second set of outlier variables. Obtain the intersection of the first set of outliers and the second set of outliers to get the set of high-confidence outliers. Remove the set of high-confidence outliers from the initial dataset to get a clean dataset.
[0015] An innovative threshold judgment step based on the prediction residual of the benchmark model is introduced. The key to this step is the introduction of a supervised feedback mechanism, which effectively removes noise while retaining the data features that are beneficial to model training to the maximum extent, resulting in a clean dataset.
[0016] In an optional embodiment of the present invention, the apparent airflow velocity is normalized by calculating the first quartile (Q1), the third quartile (Q1), and the lower boundary (Q1). Draw a box plot using 1.5×IQR and the upper boundary (Q3+1.5×IQR), as shown below. Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 As shown in the figure. Furthermore, the isolated forest algorithm and residual threshold supervision method were used to enhance outlier removal, resulting in the elimination of 98 outlier groups.
[0017] S3. Grouping fluid physics mechanism features based on clean datasets.
[0018] The apparent gas flow rate, gas density, gas Reynolds number, and Weber number in the clean dataset are classified as entrained dynamic features; the apparent liquid flow rate, liquid density, liquid dynamic viscosity, surface tension, and liquid Reynolds number in the clean dataset are classified as entrained impedance features; and the pipe diameter and pressure in the clean dataset are classified as system geometric and state features.
[0019] S4. Based on different feature groups, the droplet entrainment rate is predicted, and prediction results for multiple operating conditions are obtained.
[0020] To ensure the model strictly adheres to known physical laws even under conditions of limited training data and to improve its reliability across the entire operating range, a physical information neural network is used to process the entrainment dynamic characteristics. Basic fluid dynamics laws are introduced as regularization terms into the loss function by adding physical constraints. An empirical value cheat table for λ is generated by combining grid search with the percentage of violation of physical constraints to determine the coefficient weights of each physical constraint loss, thereby obtaining the predicted first droplet entrainment rate E. PINN The entrainment impedance characteristics are processed using symbolic regression based on genetic programming. Through natural selection and evolution, multiple heterogeneous physical quantities are transformed into a novel, physically meaningful dimensionless feature group, thereby obtaining the second droplet entrainment rate prediction result E. GP The system's geometry and state features contain only two characteristics, resulting in very low degrees of freedom. Based on Occam's razor principle in machine learning, a simple random forest algorithm is used to process the data and obtain the third droplet entrainment rate prediction result E. RF .
[0021] In an optional embodiment of the present invention, the physical information neural network processes the entrainment dynamic characteristics and introduces the following constraints: energy consumption inequality constraint: the work done by the airflow for entraining droplets cannot exceed the kinetic energy of the airflow itself; critical Weber number constraint: entrainment can only occur when the Weber number provided by the airflow is greater than the critical Weber number of the droplet; gas phase Reynolds number constraint: as the gas phase Reynolds number increases, the entrainment fraction tends to saturate, and a smoothing constraint is added to avoid violent oscillations that do not conform to physical laws.
[0022] In an optional embodiment of the invention, the ratio of training set size to validation set size to test set size is 7:2:1. Experimental data on droplet entrainment rates are typically valuable and limited. Using a 7:2:1 ratio ensures that the model is adequately trained while retaining sufficient data for effective fine-tuning and reliable unbiased evaluation.
[0023] The physical information neural network is a trained physical information neural network; the total loss function used during training includes a data loss function and a physical loss function, the physical loss function including an energy consumption loss function, a critical Weber number constraint loss function, and a meteorological Reynolds number constraint loss function; the expression for the total loss function is:
[0024] In the formula, Represents the total loss function. Represents the data loss function. Represents the physical loss function. This represents the energy consumption loss function. This represents the loss function constrained by the critical Weber number. Represents the meteorological Reynolds number-constrained loss function; , and All are searchable weight coefficients; The energy consumption loss function is as follows:
[0025] In the formula, For the sample size, For sample index, It is the ReLU activation function. For the first The entrainment dissipation power density of each sample. This represents the upper limit of the energy availability coefficient. For the first The gas density of each sample No. Apparent airflow rate of each sample; The critical Weber number constraint loss function is: ,in , In the formula, For the first The Weber number of a sample. The critical Weber number is... For surface tension, Pipe diameter; The meteorological Reynolds number-constrained loss function is: ,in , In the formula, For the first The entrainment score of each sample, with a value ranging from 0 to 1; The gas Reynolds number, This represents the gas viscosity.
[0026] Find the weight coefficient , and The specific method is as follows: The search ranges for the weight coefficients of the energy consumption loss function, the critical Weber number constraint loss function, and the meteorological Reynolds number constraint loss function are preset respectively; By traversing the combinations of various weight coefficients through grid search, a physical information neural network is trained using each combination to calculate the percentage of physical constraint violations of the energy consumption loss function, the critical Weber number constraint loss function, and the meteorological Reynolds number constraint loss function. Based on the percentage of physical constraint violations, the optimal combination of weight coefficients is selected, and a quick reference table of empirical values for weight coefficients is formed, as shown in Table 1. Table 1
[0027] Determine the weight coefficients of each loss function using a quick reference table of empirical weight coefficient values.
[0028] In this embodiment, it is determined , , The coefficient weights are 10, 0.1, and 1, respectively. Therefore, the total loss function is: .
[0029] Using genetic programming-based symbolic regression to predict droplet entrainment rate based on entrainment impedance characteristics, the predicted droplet entrainment rate is as follows:
[0030] In the formula, This represents the predicted entrainment rate of the second droplet, where ln represents the natural logarithm function. For the middle term, For the density of the liquid, For liquid viscosity, For liquid Reynolds number, Surface tension; intermediate item The expression is: .
[0031] In an optional embodiment of the present invention, the random forest algorithm is further used to process the geometric and state characteristics of the system, and the optimal hyperparameter combination is determined through key hyperparameter tuning and optimization methods to obtain E. RF ,like Figure 7 As shown.
[0032] S5. Based on the current operating conditions of the gas well, the prediction results of multiple operating conditions are dynamically weighted and fused to obtain the final droplet entrainment rate prediction result.
[0033] The specific method is as follows: A dynamic weight generator is constructed, with three nodes in its input layer, corresponding to three operating condition variables, including entrained dynamic characteristics, entrained impedance characteristics, and system geometry and state characteristics; and three nodes in its output layer, corresponding to the weight coefficients of different operating condition variables. Input the three operating condition variables corresponding to the current operating condition into the dynamic weight generator to obtain the dynamic weight coefficients A, B, and C of the three different operating condition variables; The final droplet entrainment rate prediction result is obtained by weighting and fusing the prediction results of the first droplet entrainment rate, the second droplet entrainment rate, the third droplet entrainment rate, and the dynamic weight coefficients of three different operating condition variables.
[0034] In an optional embodiment of the present invention, Gaussian Mixture clustering is performed on the feature values to obtain three working condition labels. The LightGBM classifier outputs a probability vector based on the working condition. The probability vector p is then mapped to weights A, B, and C using Softmax regression. Finally, the weighted fusion yields E.
[0035]
[0036]
[0037] .
[0038] To verify the beneficial effects of the present invention, the final droplet entrainment rate prediction value E of the model is compared with the calculated value of the existing droplet entrainment rate prediction method.
[0039] In an optional embodiment of the present invention, the test set is Wolf's experimental data from 2001, with a pipe diameter of 31.8 mm, a pressure of 2.38 bara, a gas density of 2.7946 kg / m³, a liquid density of 1000 kg / m³, a gas dynamic viscosity of 0.000018 Pa•s, a liquid dynamic viscosity of 0.001 Pa•s, and an apparent gas velocity of 25 m / s. The results output by the present invention are compared with the calculated values from the Oliemans, Ishii & Mishima, Zhang, Cioncolini & Thome, Aliyu, and Wang formulas as follows: Figure 8 As shown, the present invention has the highest accuracy, which is superior to the traditional empirical formula calculation method.
[0040] This invention integrates gas-liquid phase operating parameters, dimensionless parameters, and system geometric features as inputs to mine complex nonlinear relationships in high-dimensional data, achieving accurate and efficient prediction of droplet entrainment rate under a wide range of operating conditions.
Claims
1. A method for predicting the entrainment rate of gas well droplets based on big data, characterized in that, include: Collect experimental data on droplet entrainment rate in gas wells, summarize the influencing factors considered under different experimental conditions, and form an initial dataset. The initial dataset is preprocessed with co-validation and residual feedback to obtain a clean preprocessed dataset. Grouping of fluid physics mechanism features based on clean datasets; Droplet entrainment rate is predicted based on different feature groups, and prediction results for multiple operating conditions are obtained. The prediction results of multiple operating conditions are dynamically weighted and fused based on the current operating conditions of the gas well to obtain the final droplet entrainment rate prediction result.
2. The method according to claim 1, characterized in that, The initial dataset includes apparent gas flow rate, gas density, gas Reynolds number, Weber number, apparent liquid flow number, liquid density, liquid dynamic viscosity, surface tension, liquid Reynolds number, pipe diameter, and pressure.
3. The method according to claim 1, characterized in that, The specific preprocessing steps for co-validation and residual feedback on the initial dataset are as follows: Based on the initial dataset, the box plot method is used to label the outlier variables in the initial dataset to obtain the first set of outlier variables; Identify initial outliers in the initial dataset using the Isolation Forest algorithm; An anomaly prediction is performed on the initial dataset using a pre-trained regression model, the relative error of each sample is calculated, and a residual threshold is set based on the relative error. Data exceeding the residual threshold in the initial outliers are recorded as outlier variables, resulting in the second set of outlier variables; Obtain the intersection of the first set of outliers and the second set of outliers to get the set of high-confidence outliers. Remove the set of high-confidence outliers from the initial dataset to get a clean dataset.
4. The method according to claim 2, characterized in that, The specific grouping of fluid physics mechanism features based on a clean dataset is as follows: The apparent gas flow rate, gas density, gas Reynolds number, and Weber number in the clean dataset are classified as entrained dynamic features; the apparent liquid flow rate, liquid density, liquid dynamic viscosity, surface tension, and liquid Reynolds number in the clean dataset are classified as entrained impedance features; and the pipe diameter and pressure in the clean dataset are classified as system geometric and state features.
5. The method according to claim 4, characterized in that, The specific method for predicting droplet entrainment rate based on different feature groups and obtaining prediction results for multiple operating conditions is as follows: The first droplet entrainment rate prediction result is obtained by using a physical information neural network based on entrainment dynamic characteristics; the second droplet entrainment rate prediction result is obtained by using a symbolic regression based on genetic programming based on entrainment impedance characteristics; and the third droplet entrainment rate prediction result is obtained by using a random forest algorithm based on system geometry and state characteristics.
6. The method according to claim 5, characterized in that, The physical information neural network is a trained physical information neural network; the total loss function used during training includes a data loss function and a physical loss function, the physical loss function including an energy consumption loss function, a critical Weber number constraint loss function, and a meteorological Reynolds number constraint loss function; the expression for the total loss function is: In the formula, Represents the total loss function. Represents the data loss function. Represents the physical loss function. This represents the energy consumption loss function. This represents the loss function constrained by the critical Weber number. Represents the meteorological Reynolds number-constrained loss function; , and All are searchable weight coefficients; The energy consumption loss function is as follows: In the formula, For the sample size, For sample index, It is the ReLU activation function. For the first The entrainment dissipation power density of each sample. This represents the upper limit of the energy availability coefficient. For the first The gas density of each sample For the first Apparent airflow rate of each sample; The critical Weber number constraint loss function is: ,in , In the formula, For the first The Weber number of a sample. The critical Weber number is... For surface tension, Pipe diameter; The meteorological Reynolds number-constrained loss function is: ,in , In the formula, For the first The entrainment score of each sample, with a value ranging from 0 to 1; The gas Reynolds number, This represents the gas viscosity.
7. The method according to claim 6, characterized in that, Find the weight coefficient , and The specific method is as follows: The search ranges for the weight coefficients of the energy consumption loss function, the critical Weber number constraint loss function, and the meteorological Reynolds number constraint loss function are preset respectively; By traversing the combinations of various weight coefficients through grid search, a physical information neural network is trained using each combination to calculate the percentage of physical constraint violations of the energy consumption loss function, the critical Weber number constraint loss function, and the meteorological Reynolds number constraint loss function. Select the optimal combination of weight coefficients based on the percentage of physical constraint violations, and form a quick reference table of empirical values for weight coefficients. Determine the weight coefficients of each loss function using a quick reference table of empirical weight coefficient values.
8. The method according to claim 6, characterized in that, Using genetic programming-based symbolic regression to predict the droplet entrainment rate based on entrainment impedance characteristics, the predicted entrainment rate of the second droplet is as follows: In the formula, This represents the predicted entrainment rate of the second droplet, where ln represents the natural logarithm function. For the middle term, For the density of the liquid, For liquid viscosity, For liquid Reynolds number, Surface tension; intermediate item The expression is: 。 9. The method according to claim 5, characterized in that, The specific method for dynamically weighting and fusing prediction results from multiple operating conditions based on the current operating conditions of the gas well to obtain the final droplet entrainment rate prediction result is as follows: A dynamic weight generator is constructed, with three nodes in its input layer, corresponding to three operating condition variables, including entrained dynamic characteristics, entrained impedance characteristics, and system geometry and state characteristics. Its output layer consists of three nodes, corresponding to the weight coefficients of different working condition variables; Input the three operating condition variables corresponding to the current operating condition into the dynamic weight generator to obtain the dynamic weight coefficients of the three different operating condition variables; The final droplet entrainment rate prediction result is obtained by weighting and fusing the prediction results of the first droplet entrainment rate, the second droplet entrainment rate, the third droplet entrainment rate, and the dynamic weight coefficients of three different operating condition variables.