CN113670384B - Multi-variable time sequence diagram convolution multiphase flow virtual metering method and system - Google Patents

Abstract

The invention discloses a multivariable time sequence chart convolution multiphase flow virtual metering method, which comprises the steps of firstly obtaining n groups of corresponding pressure, differential pressure and temperature, and calculating the association degree between the input and the output of the n groups of corresponding pressure, differential pressure and temperature, and then carrying out data enhancement on the corresponding pressure, differential pressure and temperature according to the association degree, training and testing a virtual metering model to obtain a flow metering model. The corresponding system comprises a data exploratory analysis module, an algorithm module and a predictive analysis output value module. The method has the remarkable effects that the liquid quantity measurement precision of the virtual flowmeter is improved through the graph neural network iterative algorithm when the multiphase fluid is measured, the measurement precision is higher than that of the impedance water-containing meter under the gas-containing working condition, the debugging is convenient, the purchasing and maintenance cost of the multiphase flowmeter of the oil field can be effectively reduced, the continuous real-time metering of the multiphase flow is realized, and the method has higher data reference value for the production dynamic monitoring, the flow assurance and the oil reservoir management of the oil field.

Description

A multi-variable time series diagram convolution multi-phase flow virtual metering method and system

Technical field

The invention relates to oil and gas engineering, and in particular to virtual measurement technology in petroleum engineering.

Background technique

With the development and application of artificial intelligence technology and big data algorithms, and affected by the downturn in global oil prices and the safety control of radioactive sources, the domestic and foreign oilfield markets have urgent and strong demand for low-cost, non-discharge type multiphase flow meters. Most current online multiphase flow meter (MPFM) measurements are based on overall flow using radioactive, electromagnetic electronics, ultrasonic and other technologies combined with phase fraction measurement. However, these phase fraction ratio measurement technologies have certain limitations. On the one hand, although radioactive technology is still the most accurate and reliable multi-phase flow measurement method so far, many countries around the world have strict restrictions on the use of radioactive source technology; on the other hand, On the technical level, when the gas content in the medium reaches a certain limit, the measurement error of the flow meter will deteriorate sharply, even if it adopts a special design. Special low-cost design solutions can therefore be used to meet the flow meter measurement range.

Contents of the invention

In view of this, the present invention provides a mathematical algorithm measurement model based on a software system, which can input feature and predicted value data to learn non-linear relationships to obtain measurement results.

The technical solution is as follows:

A multi-variable time series diagram convolution multi-phase flow virtual measurement method, which is characterized by following the following steps:

Step 1: Obtain historical data related to flow, including n sets of corresponding seven columns of data including pressure P, differential pressure DP, temperature T, liquid flow L, water flow W, oil flow O, and air flow G;

Step 2: Take out each column _of data For a column of data Y _i , rX _→Y ;

in:

n is the total amount of data in each column;

is the arithmetic mean of the corresponding column data;

is the arithmetic mean of the corresponding column data;

Step 3: Evaluate the importance of each column of data in _{X i} to each column of data in Y _i based on the magnitude of the correlation coefficient r Coefficients a, b, c;

Step 4: Multiply each data in pressure P, differential pressure DP, and temperature T by its corresponding weight coefficient to obtain revised pressure a*P, revised differential pressure b*DP, and revised temperature c*T;

Step 5. Use the revised pressure a*P, revised differential pressure b*DP, and revised temperature c*T as inputs, and use the liquid flow rate L, the water flow rate W, the oil flow rate O, and the air flow rate G as the outputs to train and test the virtual metering model. , get the flow measurement model: L, W, O, G = F (P, T, DP).

At the same time, the present invention provides a multi-variable time sequence diagram convolution multi-phase flow virtual metering system, the key points of which are: including a data exploratory analysis module, an algorithm module and a predictive analysis output value module set in sequence;

The data exploratory analysis module is used to obtain historical data related to traffic, the algorithm module is used to calculate the correlation coefficient r _X→Y , and the predictive analysis output value module is used to train and test the virtual metering model.

Description of the drawings

Figure 1 is a schematic diagram related to the present invention;

Figure 2 is a flow chart related to the present invention.

Detailed ways

The present invention will be further described below in conjunction with the examples and drawings.

A multi-variable time series diagram convolution multi-phase flow virtual measurement method, which is carried out according to the following steps:

Step 1: Obtain historical data related to flow, including n sets of corresponding seven columns of data including pressure P, differential pressure DP, temperature T, liquid flow L, water flow W, oil flow O, and air flow G, and check each data separately. Continuity of column data, and then serialize discrete data and continuous data respectively. Finally, the data is read in the form of a histogram to make a preliminary qualitative analysis, and the inter-class distance and quantitative difference indicators of the data are analyzed;

Step 2: Take out each column _of data For a column of data Y _i , rX _→Y ;

in:

n is the total amount of data in each column;

is the arithmetic mean of the corresponding column data;

is the arithmetic mean of the corresponding column data;

Step 3: Evaluate the importance of each column of data in _{X i} to each column of data in Y _i based on the magnitude of the correlation coefficient r Coefficients a, b, c;

Step 4: Multiply each data in pressure P, differential pressure DP, and temperature T by its corresponding weight coefficient to obtain revised pressure a*P, revised differential pressure b*DP, and revised temperature c*T;

Step 5. Use the revised pressure a*P, revised differential pressure b*DP, and revised temperature c*T as inputs, and use the liquid flow rate L, the water flow rate W, the oil flow rate O, and the air flow rate G as the outputs to train and test the virtual metering model. , get the flow measurement model: L, W, O, G = F (P, T, DP).

The virtual metering model can choose from a variety of existing models. In this case, the graph convolutional neural network model is taken as an example: first, according to the revised pressure a*P, revised differential pressure b*DP, revised temperature c*T, liquid flow rate L, The water flow rate W, the oil flow rate O, and the air flow rate G are read into the adjacency matrix, and then calculated downward through the graph convolutional neural network to realize data propagation between graph networks. Finally, the correlation coefficient is output through full connection regression, thereby training to obtain flow measurement. Model: L, W, O, G = F (P, T, DP).

Example 2:

A multi-variable time series diagram convolution multi-phase flow virtual metering system based on Embodiment 1, including a data exploratory analysis module, an algorithm module and a predictive analysis output value module set in sequence;

The data exploratory analysis module is used to obtain historical data related to traffic, the algorithm module is used to calculate the correlation coefficient r _X→Y , and the predictive analysis output value module is used to train and test the virtual metering model.

Example 3:

The graph convolutional neural network model is divided into three modules: the spatial correlation graph learning module, the temporal correlation graph multi-step Transformer module, the loss function and the fully connected module. The core idea is to implement the nonlinear mapping weighting of the adjacency matrix and the coefficient matrix. The edge random walk finds the best prediction point to construct an effective graph and then outputs the prediction value through full connection. The graph learning module is usually designed to learn an adjacency matrix to adaptively obtain the temporal and spatial changing relationships between variables in time series data.

Since the graph convolution network is based on the adjacency matrix to describe the relationship between edges and vertices, the core of the calculation is the following frequency response matrix.

X＝σ(v·diag[θ _N ]v ^T x)

In the above formula, v represents a point, diag represents a diagonal matrix, X represents the output layer feature map, x represents the input layer feature map, [θ _N ] represents the edge feature input, and v ^T represents the point transpose matrix.

The mathematical principle of graph learning is the first choice. We convert the read data into a matrix of m rows and n columns. The convolution kernel h of the graph learning module is calculated as follows:

The multi-step Transformer calculates the obtained graph convolution features through the graph convolution module

After Laplace transform

In the above formula is the Laplace transformation matrix eigenvalue, y is the predicted value to be output, and H outputs the layer normalization matrix.

The deep regression tree is designed using the neural network in the artificial intelligence algorithm, and the standard residual sum of squares and the minimum residual sum of squares are selected to control the output and growth process of the tree for the deep learning regression tree model. There are weak links between random forests, and this method can avoid errors caused by greedy algorithms during the tree growth process. First, we define the loss function as L, where y is the predicted value, use a virtual econometric analysis model and use MSE (mean squared error) to evaluate the loss error.

Let the derivative be 0 to minimize the mean square error.

Based on the above calculations, the predicted y can be obtained through the fully connected layer with the smallest mean square error. is the cosine loss value.

Time series forecasting tasks can be performed following different methods. The most classic methods are based on statistics and autoregression. To be more accurate with augmentation and ensemble based algorithms, we have to use rolling cycles to generate a large number of useful handcrafted features. On the other hand, we can use neural network models that provide more freedom during development, providing customizable features for sequential modeling.

Recurrent and convolutional structures have achieved great success in time series forecasting. Interesting approaches in this area are by employing Transformers and Attention architectures that are originally native to sequence data.

The use of structures seems unusual, in a graph structure we have a network of different nodes that are related to each other by some kind of link. What we try to do is use graphical representations of time series to generate future forecasts.

We exploit graph convolutional neural networks to explore the nested structure of the data. Graph neural networks are employed in uncommon situations such as time series forecasting. In our deep learning model, graph dependencies are combined with the recursive part in an attempt to provide more accurate predictions. This approach is well suited to our problem because we can emphasize the underlying hierarchical structure in the data and encode it with a correlation matrix.

The graph learning module first reads the cleaned data into memory through GCN and then uses GNN to convert the input features and output prediction values into a sparse graph adjacency matrix. By analyzing the coefficient matrix between the adjacency matrices of the sparse graph, we can get Is the relationship between parameters linear or non-linear and visualized? In order to reduce the amount of calculations in this process, we use a single-direction calculation graph to implement the calculation of the graph learning module to alleviate the possible over-fitting of over-smoothing and thereby reduce The calculation amount is stable and the graph learning module is realized. This module mainly completes the mapping of feature layers from data input through hidden layers.

In the graph convolution module, our graph learning module is the first choice to analyze the input feature relationship, and then assign the input features to the structure in the network through the encoder to learn the feature weights and map them to a low-dimensional space. This process is implemented through a feature embedding matrix. This The learning process ensures the learning efficiency of nodes by optimizing the parameters of the encoder and implements a graph convolution module. This module mainly analyzes the weight of the wandering edges by analyzing the features of the feature layer through graph convolution.

Example 4:

This invention uses a sequential graph convolutional neural network algorithm. Through this algorithm, we design the following systematic method to solve the problem. This system can better handle pressure, differential pressure Dp, temperature Temp or other high-energy, low-energy or even multiple input independent variable features for us, and analyze the relationship between features and the non-linear mapping between independent variables and dependent variables to predict Liquid output, oil volume, gas volume.

(1) Step one: Use read__sql_query to read the real liquid content, oil content, water content, gas content, pressure, pressure difference, and temperature as a training set. After slicing, extract the quarterly, monthly, and daily data that need to be analyzed and then discrete The data must be filled with random numbers and serialized before being read into the data EDA analysis module.

(2) The second step: scale the error by retaining six decimal points before use, and then distribute the pressure, differential pressure, and temperature respectively. You can analyze the qualitative issues of trend changes and remove the range of noise point data, and then analyze Pearson correlation coefficient between independent variables.

(3) The third step: Use pyheatmap to visualize the heat map analysis through the Pearson correlation coefficient between the independent variables pressure, pressure difference, temperature and the dependent variable. After the above analysis, directly analyze the relationship between the independent variables and the dependent variable qualitatively through the heat map. Data relationship type and correlation strength.

(4) Step 4: After the first three steps, we obtained complete serialized data, divided the data into training set, test set, and verification set, and obtained the training set and test set data through the graph convolutional neural network model. Training weights file.

(5) Step 5: Pass the trained weight file through the model loading part. Test the optimized model on the validation set data and deploy it through the online deployment system. Then, the independent variable features read through each different well are output through the output value module. Predicted values of liquid content, water content, gas content, and oil content.

Example 5:

A learning and training method for a virtual metering model. The first step is to use a traditional flow meter system to measure data at a certain well plant for a period of time, and then transmit the data to the cloud through a data transmission device or a private cloud or directly transmit the data. to edge computing platforms.

Step 2: Convert the data obtained in the first step into sequence data through data preprocessing and cleaning. The sequence data is divided into discrete and continuous types.

Step 3: Conduct exploratory data analysis on the data obtained in Step 2 to obtain the best predictors and principal component factors affected by changes.

Step 4: Perform model training and transfer learning on the feature factors and predicted values to obtain the weight file of the model, and then deploy the weight file obtained by the model as an online algorithm system.

Step 5: Analyze, verify, optimize and update the model through the data obtained from the prediction calculation in step 4.

Verified by the implementation of oil field projects, statistical calculations show that the virtual measurement error obtained by the present invention is very small, and the mean square error (MSE) is less than 0.05.

Beneficial effects: The technical solution of the present invention is used to improve the liquid volume measurement accuracy of the virtual flow meter when measuring multi-phase fluids through the graph neural network iterative algorithm. It has higher measurement accuracy than the impedance water meter under gas-containing conditions and is easy to debug. It can effectively reduce the purchase and maintenance costs of multi-phase flow meters in oil fields, realize continuous real-time measurement of multi-phase flow, and has high data reference value for dynamic monitoring of oil field production, flow assurance and reservoir management.

Finally, it should be noted that the above descriptions are only preferred embodiments of the present invention. Under the inspiration of the present invention, those of ordinary skill in the art can make a variety of similar embodiments without violating the purpose and claims of the present invention. It means that such transformations fall within the protection scope of the present invention.

Claims (4)

1. A multi-variable time series diagram convolution multi-phase flow virtual metering method, which is characterized by following the following steps: Step 1: Obtain historical data related to flow, including n sets of corresponding seven columns of data including pressure P, differential pressure DP, temperature T, liquid flow L, water flow W, oil flow O, and air flow G; Step 2: Take out each column _of data For a column of data Y _i , rX _→Y ; in: n is the total amount of data in each column; is the arithmetic mean of the corresponding column data; is the arithmetic mean of the corresponding column data; Step 3: Evaluate the importance of each column of data in _{X i} to each column of data in Y _i based on the magnitude of the correlation coefficient r Coefficients a, b, c; Step 4: Multiply each data in pressure P, differential pressure DP, and temperature T by its corresponding weight coefficient to obtain revised pressure a*P, revised differential pressure b*DP, and revised temperature c*T; Step 5. Use the revised pressure a*P, revised differential pressure b*DP, and revised temperature c*T as inputs, and use the liquid flow rate L, the water flow rate W, the oil flow rate O, and the air flow rate G as the outputs to train and test the virtual metering model. , get the flow measurement model: L, W, O, G = F (P, T, DP). 2. A multi-variable time series diagram convolution multi-phase flow virtual metering method according to claim 1, characterized in that: in the step one, after obtaining the historical data related to the flow rate, the continuous data of each column of data is separately checked. properties, and then serialize the discrete data and continuous data respectively, and finally read the data in the form of a histogram to make a preliminary qualitative analysis, and analyze the inter-class distance and quantitative difference indicators of the data; then proceed to step two. 3. A multi-variable time sequence diagram convolution multi-phase flow virtual metering method according to claim 1, characterized in that: in the step five, first, according to the revised pressure a*P, the revised differential pressure b*DP, the revised Temperature c*T, liquid flow rate L, water flow rate W, oil flow rate O, and air flow rate G are read into the adjacency matrix, and then calculated downward through the graph convolutional neural network to realize data propagation between graph networks, and finally through fully connected regression. Output the correlation coefficient to train the flow measurement model: L, W, O, G = F (P, T, DP). 4. A system based on the multi-variable time series diagram convolution multi-phase flow virtual metering method according to claim 1, characterized in that it includes a data exploratory analysis module, an algorithm module and a predictive analysis output value module set in sequence; The data exploratory analysis module is used to obtain historical data related to traffic, the algorithm module is used to calculate the correlation coefficient r _X→Y , and the predictive analysis output value module is used to train and test the virtual metering model.