Oil-gas-water three-phase flow process data two-stage amplification and flow characteristic modeling method

Abstract

The invention discloses an oil-gas-water three-phase flow process data two-stage amplification and flow characteristic modeling method. The method comprises the following steps: step 1) establishing a probability gating circulation unit model of a multiphase flow pump flow characteristic; 2) predicting an input sample of each working condition by using a PGRU model; step 3) utilizing probability characteristics provided by the PGRU model to identify an area needing data amplification; step 4) based on an active noise elimination method, executing sample amplification on the identified area; (5) repeating the steps (2)-(4), identifying a to-be-amplified sample area of each working condition, and supplementing samples which are consistent with original data in distribution; 6) collecting all data from the steps 2-5 as original data of the TimeVAE model, and performing second-stage data amplification; and 7) training a multiphase flow pump flow characteristic prediction model by using all data obtained in the step 6). According to the method, two-stage amplification and flow characteristic prediction of three-phase flow process data can be realized under a complex and changeable oil-gas-water three-phase mixed transportation working condition.

Description

Technical Field

[0001] This invention relates to the field of multiphase flow pump and valve design technology, specifically to a two-stage amplification and flow characteristic modeling method for three-phase flow data of oil, gas and water. Background Technology

[0002] Multiphase flow pumps are key equipment for the closed-loop transportation of oil, gas, and water three-phase mixtures in crude oil extraction. Under dynamic operating conditions, the interaction between complex interfacial dynamics and irregular valve movements causes the oil-gas-water three-phase flow to exhibit unsteady, non-uniform, multi-scale, and multi-physics field coupling characteristics. The resulting gas-liquid cavitation and oil-water emulsification phenomena not only reduce pump efficiency but also lead to various malfunctions. Establishing the relationship between the dynamic characteristics of multiphase flow pumps, such as flow rate and phase content, and transportation operating parameters is of great significance for optimizing pump structure design and real-time control of multiphase flow conditions.

[0003] Currently, research on multiphase flow characteristics mainly employs mechanistic models, computational fluid dynamics (CFD) models, and experimental measurement methods. Due to limited understanding of the physical mechanisms of phase transitions, interfaces, and flow instabilities, most existing mechanistic models can only describe steady-state gas-liquid two-phase flow, failing to capture the complex dynamic characteristics exhibited in oil-gas-water stratification and oil-water emulsion flows. Current CFD models typically rely on empirically selected single turbulence models, unable to fully describe the dynamic processes of three-phase flow involving multiple flow modes. Furthermore, effectively capturing interphase interactions at different time scales usually requires high-resolution meshes, increasing computational time and resource consumption. Moreover, given the limited multiphase flow testing techniques that struggle to obtain large amounts of reliable experimental data, measurement methods combining traditional sensors with advanced information processing technologies have received considerable attention. However, existing experimental methods are mostly applied to gas-liquid two-phase flow in pipelines, with research on describing the characteristics of three-phase flow within multiphase pumps remaining limited. Particularly for high-viscosity and corrosive crude oil, installing a sufficient number of sensors within the pump chamber to obtain adequate data is also quite challenging.

[0004] In recent years, some data-driven empirical models, which do not require substantial understanding of the complex internal mechanisms of multiphase flow, have gradually been used to solve some problems related to multiphase flow. However, the quality of purely data-driven empirical models depends on the availability of a large amount of reliable modeling data. Therefore, this invention proposes a two-stage data augmentation and flow characteristic modeling method that integrates knowledge of three-phase flow processes and data probabilistic characteristics. Furthermore, a literature search revealed no reports of such a two-stage data augmentation and flow characteristic modeling method for oil-gas-water three-phase flow processes. Summary of the Invention

[0005] This invention aims to overcome the aforementioned shortcomings of existing technologies by providing a two-stage amplification and flow characteristic modeling method for oil-gas-water three-phase flow process data. Based on a limited sample, this method can learn the complex flow characteristics of oil-gas-water three-phase flow within a multiphase pump and the relationship between these characteristics and mixed-transport conditions. This assists engineers in better understanding the flow mechanism of oil-gas-water three-phase flow within the pump, thereby optimizing pump structure design and real-time control of the multiphase flow state.

[0006] The technical solution of the present invention is as follows:

[0007] A two-stage amplification and flow characteristic modeling method for three-phase flow data of oil, gas and water includes the following steps:

[0008] 1) Collect input and output training samples under different three-phase oil, gas and water mixed transportation conditions, and establish a probabilistic gated cyclic unit (PGRU) model of the flow characteristics of multiphase flow pumps;

[0009] First, based on the design experience of three-phase flow mixed transportation of oil, gas and water, the variables affecting the flow characteristics of multiphase flow pumps are analyzed, and the input matrix X and output matrix Y of the PGRU model are determined.

[0010] Secondly, training samples of multiphase flow pump input and output under different mixed-transport conditions were collected. Where S k For the sample set of the k=1,…,K working conditions, x m For the m-th input sample, y m Let m be the m-th output sample, and M be the total number of samples.

[0011] Finally, a PGRU model was built using the collected samples.

[0012] The implementation method is illustrated using the example of establishing a PGRU model by inserting Monte Carlo dropout (MC dropout) algorithm layers between the recurrent layers of the GRU model. The final hidden state of the PGRU model after performing MC dropout sampling for n=1,…,N times is described as follows:

[0013]

[0014] In the formula, They are sample x m After input, the PGRU model's reset gate, update gate, candidate hidden state, and final hidden state are obtained after the nth MC dropout sampling. N is the total number of MC dropout samplings, w r w u w c and b r b u b cThese are the weight matrix and bias vector of the PGRU model, respectively. σ(·) is the sigmoid activation function, tanh is the hyperbolic tangent function, and * denotes element-wise multiplication. The weight matrix and bias vector of the PGRU model are updated using backpropagation.

[0015] 2) Use the PGRU model to predict the input samples for each working condition;

[0016] Working condition S k The k-th input sample x k For example, after performing the nth MC dropout sampling, the corresponding prediction output of the PGRU model is... for:

[0017]

[0018] In the formula, w0 and b0 are the weight matrix and bias vector of the fully connected layer, respectively. It is sample x k The final hidden state of the PGRU model after input and nth MC dropout sampling.

[0019] 3) Utilize the probabilistic characteristics provided by the PGRU model to identify regions that require data augmentation;

[0020] With input sample x k For example, the PGRU model predicts x k The probability characteristic σ(x) k It can be described as:

[0021]

[0022] In the formula, After performing the nth MC dropout sampling, x k The prediction variance. If σ(x) k If ρ = constant, then x k Samples need to be amplified in the nearby area.

[0023] 4) Based on active noise cancellation, perform sample amplification for the identified regions;

[0024] In x k Taking the addition of samples in the vicinity as an example, firstly, ensure that the J initial samples to be added all conform to... Distribution, j = 1, ..., J and 0 < τ k ≤1;

[0025] Secondly, the Kolmogorov-Smirnov test is used to assess the consistency between the distributions of the added samples and the original samples. The test method is described below:

[0026]

[0027] In the formula, It is |F o (x k )-F a (x j The maximum value of F | o (x k ) and F a (x j ) are the cumulative distribution functions of the original data and the augmented data, respectively. If Then adjust τ m The size, until until.

[0028] 5) Repeat steps 2-4 to identify the sample regions to be expanded for each working condition and supplement them with samples that are consistent with the original data distribution;

[0029] 6) Collect all data from steps 2-5 as the raw data for the TimeVAE model and perform the second-stage data augmentation.

[0030] First, collect all the data.

[0031] Train the TimeVAE model for i = 1, ..., I;

[0032] Taking a three-layer temporal convolutional network (TCN) as the encoder and a three-layer deconvolutional network as the decoder as an example, the loss function L(φ,ψ) for training the TimeVAE model can be described as follows:

[0033]

[0034] In the formula, Φ and ψ are the parameters of the encoder and decoder, respectively, and q φ (z|x) and p ψ (x|z) represent the latent distributions of the encoder and decoder, respectively, and Eq φ (z|x)[] represents the expression in q φ The expected value is calculated on (z|x), where p(z) is the prior distribution, z is the latent variable of the encoder, and γ is the weight of the Kullback-Leibler (KL) divergence. The parameters Φ, ψ, and γ are optimized using the backpropagation method until formula (5) satisfies L(φ,ψ)≤ζ=constant.

[0035] Finally, new samples {x′,y′} are generated using the trained TimeVAE.

[0036] x′ can be obtained from the fully connected layer (FCL) of the TimeVAE decoder, i.e. FCL stands for Fully Connected Layer Functions. The outputs come from the third, second, and first layer decoders, respectively. The latent variable z is reshaped, i.e., z = Reshape(z). S1 = {x, y} is therefore expanded to S2 = {x ∪ x′, y ∪ y′}.

[0037] 7) Using all the data obtained in step 6, train a prediction model for the flow characteristics of a multiphase flow pump.

[0038] Taking the construction of a GRU online prediction model as an example, using all the data S2 obtained in step 6, the GRU model is trained using formula (1) and N=0 (i.e., the MC dropout layer is canceled), and formula (2) is used to train the L input sample sets of the new working condition. Traffic can be predicted online.

[0039] The design concept of this invention is as follows:

[0040] This invention proposes a two-stage data augmentation and flow characteristic modeling method that integrates knowledge of three-phase flow processes and data probabilistic characteristics. In the first stage, the regions requiring data augmentation for each operating condition are identified, and samples are added for each condition one by one based on active noise cancellation. Next, all data from the first stage is collected as raw data for a Time Variational Autoencoder (TimeVAE) for the second stage of data augmentation. Finally, using all samples after the two-stage data augmentation, the relationship between the flow characteristics of the oil-gas-water three-phase flow and the parameters of the transport operating conditions is constructed, enabling real-time prediction of the flow characteristics of multiphase flow pumps under different transport conditions. This method, which utilizes local data augmentation in the first stage and global augmentation in the second stage, can provide sufficient modeling information to describe the complex characteristics of three-phase flow within multiphase flow pumps based on limited experimental samples.

[0041] The beneficial effects of this invention are as follows:

[0042] This invention proposes a two-stage data augmentation and flow characteristic modeling method for oil-gas-water three-phase flow processes. Based on limited samples, this method effectively identifies complex characteristic regions of oil-gas-water three-phase flow and supplements sample data characterizing multiphase flow process properties. Compared to theoretical models, CFD simulations, and experimental methods, this method offers a practical and rapidly implementable approach. Compared to conventional one-step data augmentation schemes, this method integrates process knowledge and data probabilistic characteristics. By implementing two-stage data augmentation, it can simultaneously capture the local and global characteristics of oil-gas-water three-phase flow processes, providing sufficient modeling information to describe the complex characteristics of three-phase flow processes within multiphase pumps and reducing the need for large amounts of labeled data. Attached Figure Description

[0043] Figure 1 This is a flowchart of the method of the present invention;

[0044] Figure 2 This is the data amplification sample curve under the first stage of the present invention in operating condition S1;

[0045] Figure 3 This is the data amplification sample curve for the first stage of operating condition S2 of this invention;

[0046] Figure 4 This is the data amplification sample curve for the first stage of operating condition S3 of this invention;

[0047] Figure 5 This is the data amplification sample curve for the first stage of operating condition S4 of this invention;

[0048] Figure 6 This is the data amplification sample curve for the first stage of operating condition S5 of this invention;

[0049] Figure 7 This is the data amplification sample curve for the first stage of operating condition S6 of this invention;

[0050] Figure 8 Comparison of the prediction results of the TimeVAE-GRU-II, TimeVAE-GRU-I and the original GRU model for the flow characteristics of multiphase flow pumps in this invention Figure 1 ;

[0051] Figure 9 Comparison of the prediction results of the TimeVAE-GRU-II, TimeVAE-GRU-I and the original GRU model for the flow characteristics of multiphase flow pumps in this invention Figure 2 ;

[0052] Figure 10 Comparison of the prediction results of the TimeVAE-GRU-II, TimeVAE-GRU-I and the original GRU model for the flow characteristics of multiphase flow pumps in this invention Figure 3 . Detailed Implementation

[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0054] like Figure 1 As shown, the two-stage amplification and flow characteristic modeling method for three-phase flow data of oil, gas and water includes the following steps:

[0055] 1) Collect input and output training samples under different three-phase oil, gas and water mixed transportation conditions, and establish a probabilistic gated cyclic unit (PGRU) model of the flow characteristics of multiphase flow pumps;

[0056] First, taking a reciprocating three-phase mixed-transfer pump of oil, gas, and water as the experimental object, based on its engineering design experience, it can be known that the measurable variable affecting the flow characteristics of the reciprocating mixed-transfer pump is the pump inlet pressure P. s β of the imported gas content s Export pressure P d The crank angle θ = ωt. Therefore, the input variables for the PGRU model are determined as follows: The output variable is Where x m For the m-th input sample, y m For the m-th output sample, Q m Let M be the output flow rate of the m-th sample;

[0057] Secondly, six sample sets S1, ..., S6, totaling 720 samples, were collected from six different mixed-transportation conditions for training the PGRU model. Each condition has the same inlet pressure P. s β of the imported gas content s Export pressure P d And different crank angles θ. Because the flow characteristics of a reciprocating mixed-transfer pump under each operating condition are periodic (θ = 0°-360° is one cycle), and its discharge flow characteristics (θ = 180°-360°) are more complex than its suction flow characteristics (θ = 0°-180°) within one cycle, this is more practically significant for optimizing pump structure design and real-time control of the mixed-transfer pump's flow state. Therefore, this invention only collects a sample set during the discharge process of the mixed-transfer pump. That is, the operating condition corresponding to S1 is P. s =0.2 MPa, β s =0.4, P d =3.0 MPa, θ = 180°-360°, the working condition corresponding to S2 is P s =0.3 MPa, β s =0.2, P d =1.0 MPa, θ = 180°-360°, the working condition corresponding to S3 is P s =0.4 MPa, β s =0.6, P d =3.0 MPa, θ = 180°-360°, the working condition corresponding to S4 is P s =0.3 MPa, β s =0.5, P d =2.5MPa, θ =180°-360°, the working condition corresponding to S5 is P s =0.4 MPa, β s =0.1, P d =1.5MPa, θ =180°-360°, the working condition corresponding to S6 is P s =0.4 MPa, β s =0.8, Pd =3.0 MPa, θ = 180°-360°. The training process of the PGRU model is as follows: Given any set of w r w u w c and b r b u b c For the initial value, taking MC dropout sampling times N=12 as an example, the PGRU model is trained using formula (1) with S1,…,S6.

[0058] 2) Use the PGRU model to predict the input samples for each working condition;

[0059] Taking the S1 working condition as an example, the predicted output of its 120 input samples after 1 to 12 MC dropout samplings is calculated using formula (2).

[0060] 3) Utilize the probabilistic characteristics provided by the PGRU model to identify regions that require data augmentation;

[0061] The probability characteristics of each input sample are calculated using formula (3). This method sets a constant ρ = 0.15, taking the 30th input sample x under condition S1 as an example. 30 For example, calculate its probability characteristic σ(x) 30 If )=0.25>ρ=0.15, then x 30 The surrounding area was identified as a region requiring sample amplification. Repeating the above steps, a total of 6 regions requiring sample amplification were identified in S1 mode.

[0062] 4) Based on active noise cancellation, perform sample amplification for the identified regions;

[0063] In x 30 Taking the addition of samples in the nearby area as an example, this method sets τ 30 The initial value is 0.1. First, let the 5 additional initial samples x be... 1-5 All obey distributed;

[0064] Secondly, calculate the cumulative distribution function of the five additional samples and the original data x. 30 The cumulative distribution function is calculated one by one using formula (4). This method sets η = 0.05, and calculations show that... The values ​​were 0.01, 0.02, 0.02, 0.03, and 0.03 respectively, therefore the five additional sample distributions all met the requirements. Repeating the above steps, a total of 25 sample distributions were added for the S1 operating condition, resulting in 25 extended flow characteristic curves consistent with the S1 flow characteristics.

[0065] 5) Repeat steps 2-4 to identify the sample regions to be expanded for each working condition and supplement them with samples that are consistent with the original data distribution;

[0066] For each operating condition, S1, ..., S6, a total of 164 sample distributions were amplified, resulting in 164 expanded flow characteristic curves. As an example, Figure 2-7 The three expanded flow characteristic curves for each operating condition are shown.

[0067] 6) Collect all data from steps 2-5 as the raw data for the TimeVAE model and perform the second-stage data augmentation.

[0068] The raw experimental data from S1, ..., S6 and the 164 flow characteristic curves amplified in steps 2-5 were collected to train the TimeVAE model.

[0069] Arbitrarily initialize parameters Φ, ψ, and γ, and calculate the loss function L(φ,ψ) of the TimeVAE model using formula (5). In this invention, ζ = 0.05 is set. If L(φ,ψ) > ζ, the backpropagation method is used to optimize parameters Φ, ψ, and γ until L(φ,ψ) ≤ ζ.

[0070] Finally, 164 flow characteristic curves were amplified using the trained TimeVAE. Therefore, after steps 1-6, the 6 sample curves for 6 operating conditions were amplified into 170 sample curves.

[0071] 7) Using all the data obtained in step 6, train a prediction model for the flow characteristics of a multiphase flow pump.

[0072] All samples are obtained from 170 sample curves. Using formula (1) and setting N=0, the GRU model is trained. Then, formula (2) is used to train the L input sample sets of the new working condition. Traffic can be predicted online.

[0073] The GRU model (TimeVAE-GRU-II) constructed in step 7 of this method was tested on three test samples (corresponding to operating conditions P). s =0.4 MPa, β s =0.5, P d =2.0Mpa, θ=180°-360°; P s =0.3 MPa, β s =0.3, P d =1.5Mpa, θ=180°-360°; P s =0.4 MPa, β s =0.7, P d=3.0 MPa, θ = 180°-360°. The predicted values ​​of the flow characteristics are compared with the predicted and experimental values ​​of TimeVAE-GRU-I (without using the first-stage data augmentation method, only using step 6 to augment the experimental samples of 6 operating conditions, and using step 7 to construct the GRU model), the original GRU model (without using the data augmentation method, only using the experimental samples of 6 operating conditions to construct the GRU model), and the original GRU model. See [link to relevant documentation]. Figure 8 , Figure 9 and Figure 10 As shown.

[0074] Choose the maximum absolute error (MAE), mean squared error (RMSE), and R-squared (R²). 2 Three indicators serve as evaluation criteria. The three criteria are defined as follows:

[0075]

[0076] In the formula, Q l It is the l-th test sample x l The experimental values, and It is Q l The average and predicted values. For both ME and RMSE indices, smaller values ​​indicate better model predictive performance; R... 2 The higher the value of the indicator, the better the model's predictive performance. The comparison results are shown in Table 1.

[0077] Table 1. Predictive performance of the method of the present invention on the flow characteristics of the test sample set (minimum value highlighted in bold).

[0078]

[0079] As shown in Table 1, compared with the original GRU model, the prediction performance of the TimeVAE-GRU-I and TimeVAE-GRU-II models established using the data-amplified samples was improved, demonstrating the effectiveness of data augmentation based on TimeVAE. Compared with the TimeVAE-GRU-I model, which lacks the first-stage data augmentation samples, the two-stage data augmentation method proposed in this invention, which integrates three-phase flow process knowledge and data probability characteristics, provides more effective data for establishing a three-phase flow flow characteristic prediction model, and the prediction performance of all three test samples was further improved.

[0080] Therefore, the two-stage amplification and flow characteristic modeling method for three-phase flow processes of oil, gas, and water proposed in this invention has universality and versatility, and can accurately predict the flow characteristics of multiphase flow pumps based on limited modeling samples. Furthermore, compared to conventional one-step global data amplification methods, this method integrates three-phase flow process knowledge into a data-driven empirical model with probabilistic characteristics. Utilizing a two-stage data amplification mode, it captures both local and global characteristics of the three-phase flow process, improves the generalization ability of the prediction model, and reduces the need for large amounts of labeled data.

[0081] Its simple and reliable segmented implementation method can reduce design complexity, lower design costs, and save modeling time, providing an effective auxiliary means for the design of current multiphase flow pumps and the real-time adjustment of flow state.

[0082] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A two-stage amplification and flow characteristic modeling method for three-phase flow data of oil, gas and water, characterized in that, Includes the following steps: Step 1) Collect input and output training samples under different three-phase oil, gas and water mixed transportation conditions, and establish a probabilistic gated cyclic unit (PGRU) model of the flow characteristics of multiphase flow pumps. Step 2) Use the PGRU model to predict the input samples for each working condition; Step 3) Use the probabilistic characteristics provided by the PGRU model to identify the regions that need data augmentation; Step 4) Perform sample amplification for the identified regions based on active noise cancellation; Step 5) Repeat steps 2)-4) to identify the sample regions to be amplified for each working condition and supplement them with samples that are consistent with the original data distribution. Step 6) Collect all data from Steps 2)-5) as the raw data for the TimeVAE model and perform the second-stage data augmentation. Step 7) Use all the data obtained in Step 6) to train a prediction model for the flow characteristics of a multiphase flow pump.

2. The method for two-stage amplification and flow characteristic modeling of three-phase flow data of oil, gas and water according to claim 1, characterized in that, The specific process of step 1) is as follows: First, based on the design experience of three-phase flow mixed transportation of oil, gas and water, the variables affecting the flow characteristics of multiphase flow pumps are analyzed, and the input matrix X and output matrix Y of the PGRU model are determined. Secondly, training samples of multiphase flow pump input and output under different mixed-transport conditions were collected. Where S k For the sample set of the k=1,…,K working conditions, x m For the m-th input sample, y m Let M be the m-th output sample, where M is the total number of samples; finally, a PGRU model is built using the collected samples. A Monte Carlo (MC) dropout algorithm layer is inserted between the recurrent layers of the GRU model to build the PGRU model; The final hidden state of the PGRU model after performing the n=1,…,Nth MC dropout sampling is described as follows: In the formula, They are sample x m After input, the PGRU model's reset gate, update gate, candidate hidden state, and final hidden state are obtained after the nth MC dropout sampling; N is the total number of MC dropout samplings, w r w u w c and b r b u b c These are the weight matrix and bias vector of the PGRU model, respectively; σ(·) is the sigmoid activation function, tanh is the hyperbolic tangent function, and * indicates element-wise multiplication; the weight matrix and bias vector of the PGRU model are updated using the backpropagation method.

3. The two-stage amplification and flow characteristic modeling method for oil-gas-water three-phase flow process data according to claim 2, characterized in that, The specific process of step 2) is as follows: Operating Condition S k The k-th input sample is x k At that time, after the nth MC dropout sampling, the corresponding prediction output of the PGRU model is... for: In the formula, w0 and b0 are the weight matrix and bias vector of the fully connected layer, respectively. It is sample x k The final hidden state of the PGRU model after input and nth MCdropout sampling.

4. The method for two-stage amplification and flow characteristic modeling of three-phase flow data of oil, gas and water according to claim 3, is characterized in that, The specific process of step 3) is as follows: Input sample is x k At that time, the PGRU model predicts x k The probability characteristic σ(x) k It can be described as: In the formula, After performing the nth MC dropout sampling, x k The prediction variance; if σ(x) k If ρ = constant, then x k Samples need to be amplified in the nearby area.

5. The method for two-stage amplification and flow characteristic modeling of three-phase flow data of oil, gas and water according to claim 4, characterized in that, The specific process of step 4) is as follows: In x k When adding samples to the nearby area, first ensure that the J initial samples being added all conform to the following rules. Distribution, j = 1, ..., J and 0 < τ k ≤1; Secondly, the Kolmogorov-Schwarz test is used to assess the consistency between the distributions of the added samples and the original samples. The test method is described below: In the formula, Yes | F o (x k )-F a (x j The maximum value of F | o (x k ) and F a (x j ) are the cumulative distribution functions of the original data and the augmented data, respectively; if D xj If η is a constant, then adjust τ. m The size, until until.

6. The method for two-stage amplification and flow characteristic modeling of three-phase flow data of oil, gas and water according to claim 5, is characterized in that, The specific process of step 6) is as follows: First, collect all the data. Train the TimeVAE model for i = 1, ..., I; When a three-layer temporal convolutional network (TCN) is used as the encoder of a TimeVAE and a three-layer deconvolutional network is used as the decoder, the loss function L(φ,ψ) for training the TimeVAE model is described as follows: In the formula, Φ and ψ are the parameters of the encoder and decoder, respectively, and q φ (z|x) and p ψ (x|z) represent the latent distributions of the encoder and decoder, respectively, and Eq φ (z|x)[] represents the expression in q φ The expected value is calculated on (z|x), where p(z) is the prior distribution, z is the latent variable of the encoder, and γ is the weight of the KL divergence. The parameters Φ, ψ, and γ are optimized using the backpropagation method until formula (5) satisfies L(φ,ψ)≤ζ=constant; Finally, new samples {x′,y′} are generated using the trained TimeVAE. x′ is obtained from the fully connected layer FCL of the TimeVAE decoder, i.e. FCL stands for Fully Connected Layer Functions; The outputs from the third, second, and first layer decoders, respectively; The latent variable z is reshaped, i.e., z = Reshape(z); S1 = {x, y} is therefore expanded to S2 = {x Ux′, y Uy′}.

7. The method for two-stage amplification and flow characteristic modeling of three-phase flow data of oil, gas and water according to claim 6, is characterized in that, The specific process of step 7) is as follows: A GRU online prediction model is constructed. Using all the data S2 obtained in step 6), the GRU model is trained using formula (1) and N=0. Then, formula (2) is used to train the model on the L input sample sets of the new working condition. Traffic can be predicted online.

Images