An interwell connectivity analysis method for asymmetric time data alignment

Abstract

The application provides an interwell connectivity analysis method for asymmetric time data alignment, comprising the following steps: S1, constructing a preliminary representation of an original one-dimensional injection-production time sequence based on graph convolution, and building a graph structure framework of an interwell connectivity analysis model; S2, remodeling the one-dimensional injection-production time sequence structure into a two-dimensional injection-production time sequence structure in the spatial domain according to different time window settings, so as to extract the interwell connectivity features; S3, performing interwell correlation degree representation of the two-dimensional injection-production time sequence data in the frequency domain through Fourier transform; and S4, training the graph neural network interwell connectivity model after feature extraction, and performing interwell connectivity representation by using the reconstruction error of the injection-production time sequence rule. The application can more accurately capture the complex and subtle time sequence dependent relationship in the asymmetric injection-production time sequence data, can independently describe and analyze the information carried by each frequency component, and can effectively avoid the inaccuracy and misdirection possibly caused by noise smoothing and overfitting problems.

Description

Technical Field

[0001] The present invention belongs to the field of petroleum engineering, and in particular relates to an inter-well connectivity analysis method for asymmetric time data alignment. Background Art

[0002] When a reservoir becomes energy-deficient after a period of natural energy development and enters a second phase of development, water injection is an important way to increase oil recovery. However, in order to ensure that oil is transferred from injection wells to production wells, managers must analyze the connectivity of the wells. In water-flooded reservoirs, underground flow is invisible and is affected by heterogeneous geophysical properties (such as porosity, compressibility, and especially permeability). Interwell connectivity analysis is an important part of reservoir characterization, aiming to quantify the contribution of injection wells to production wells, thereby reflecting the relative permeability strength of the flow channel. Based on the analysis of interwell connectivity, oil fields perform hydrodynamic adjustments such as water plugging, profile adjustment, and well network optimization.

[0003] Traditional connectivity identification methods include well testing, tracer methods, and numerical simulation methods. These traditional methods have drawbacks such as high cost, limited application, and long time consumption. Oil fields have a large amount of production and water injection data, which are inherently related to interwell connectivity. Data-driven methods provide a useful solution to the problem of calculating interwell connectivity characteristics. They aim to calculate point-by-point similarity without considering the impact of time series. However, the processing of multi-well interference remains a difficult problem. CRM is a common method for dealing with the impact of history on the current timestamp. It is derived from the material balance equation and the linear productivity model. The current interwell connectivity analysis has the problem of poor accuracy. Summary of the Invention

[0004] An embodiment of the present application provides an inter-well connectivity analysis method with asymmetric time data alignment, which can independently depict and analyze the information carried by each frequency component, thereby effectively avoiding inaccuracies and misleading information caused by noise smoothing and overfitting problems in inter-well connectivity analysis.

[0005] The present application provides a method for analyzing well connectivity using asymmetric time data alignment, including:

[0006] S1, based on graph convolution, constructs a preliminary representation of the original one-dimensional injection and production time series and builds a graph structure framework for the well connectivity analysis model;

[0007] S2, in the spatial domain, according to different time window settings, the one-dimensional injection-production time series structure is reshaped into a two-dimensional injection-production time series structure to extract the inter-well connectivity relationship features;

[0008] S3, characterize the inter-well correlation degree of 2D injection-production time series data in the frequency domain by Fourier transform;

[0009] S4, the graph neural network inter-well connectivity model after training feature extraction, uses the reconstruction error of the injection-production time sequence law to characterize the inter-well connectivity.

[0010] Wherein, step S1 includes:

[0011] S1.1. Obtain injection and production dynamic data of the block to be studied from the monitoring system of the reservoir block to construct original one-dimensional injection and production time series data with asymmetry in the time dimension; obtain geological data of the block to be studied;

[0012] S1.2, when using a graph convolutional network for well connectivity analysis, feature embedding is performed on the original one-dimensional time series data. Through the combined effects of graph convolutional mapping and position encoding, the original one-dimensional injection and production time series, which is asymmetric in the time dimension, is input into the embedding module for processing and feature conversion, thereby obtaining a preliminary representation of the injection and production data time series.

[0013] S1.3, building a graph structure framework for the inter-well connectivity analysis model; based on the geological data obtained in step S1.1, establishing a graph neural network inter-well connectivity analysis model with different nodes representing each production well and each water injection well; wherein, when the first node represents the first production well, the first node is characterized by the liquid production of the first production well stratum; when the second node represents the second water injection well, the second node is characterized by the water injection volume of the second water injection well stratum; the edges of the graph neural network model are used to represent the distance and relative position between the third well point and other well points; the third well point is a production well or a water injection well, and the other well points are production wells or water injection wells;

[0014] The input end of the graph neural network well-to-well connectivity analysis model is used to input the liquid production of each production well and the water injection volume of each water injection well; the output end of the graph neural network well-to-well connectivity analysis model is used to output the predicted liquid production of each production well.

[0015] Among them, the geological data include the number of injection wells, the number of production wells, the distance between each production well and each injection well, the distance between each production well and each production well, the distance between each injection well and each injection well, the relative position between each production well and each injection well, the relative position between each production well and each production well, and the relative position between each injection well and each injection well, as well as perforation parameters, layer depth, effective thickness, permeability and porosity.

[0016] Wherein, step S2 includes:

[0017] S2.1, based on the acquisition frequency characteristics of the production and injection data of the block to be studied, select multiple different time windows as the period P of the two-dimensional injection-production time series structure to be reshaped;

[0018] S2.2. For each selected period P, the original one-dimensional injection-production time series data is segmented with a time length of P and reshaped into a two-dimensional matrix of P×S, where S represents the number of periods contained in this time length. In this way, the inter-well connectivity relationship features in the spatial domain are extracted.

[0019] The period P is 1 hour, 6 hours or 12 hours.

[0020] Wherein, step S3 includes:

[0021] S3.1, decompose the two-dimensional injection-production time series structure in the frequency domain by Fourier transform to obtain a series of two-dimensional injection-production time series representations with different frequency characteristics;

[0022] S3.2, characterize the inter-well correlation degree of the one-dimensional structural Fourier component under each independent frequency component, and integrate the multi-scale information through fusion strategy;

[0023] S3.3, perform inverse Fourier transform on the data after information fusion in step S3.2 to restore the data to a two-dimensional structure. The formula is as follows:

[0024] X 2D * =iFFT(X 2D ')∈R P×S×c

[0025] Where, X 2D * is the reshaped two-dimensional original data obtained in step S2.2, iFFT is the inverse Fourier transform process, X 2D ′ is the data after information fusion in step S3.2;

[0026] S3.4, based on the multiple reconstructed two-dimensional data sets obtained, these two-dimensional injection and acquisition data structures based on different time patterns are restored one by one to their original one-dimensional states, and through fusion, they are integrated into new one-dimensional injection and acquisition time series data, and the new one-dimensional injection and acquisition time series data conforms to a symmetrical relationship in space.

[0027] Among them, step S3.1 includes: applying Fourier transform technology to the two-dimensional injection-production time series structure obtained in step S2.2 to realize Fourier decomposition of the signal, thereby obtaining a series of different Fourier components, which contain the detailed details and overall trend characteristics of the original sequence.

[0028] Wherein, step S4 includes:

[0029] S4.1, using the new one-dimensional injection-production time series data obtained in step S3.4, training a graph neural network well connectivity analysis model;

[0030] S4.2, constructing a learning criterion for model training based on the reconstruction error between the new one-dimensional injection-production time series data and its original injection-production data obtained in step S3.4, and performing gradient updates on the parameters in the model until the training error meets the preset conditions; if the preset conditions for the model training error are not met, returning to step 2 and retraining until the requirements are met;

[0031] S4.3, after model training is completed, the well connectivity characterization results guided by the reconstruction error after data alignment are obtained, and the liquid production dynamic data of each production well are output.

[0032] In step S4.2, a learning criterion for model training is constructed, and the parameters in the model are gradient updated until the training error is less than 10 -2 If the model training error is less than 10 -2 If the requirement is not met, return to step 2 and retrain until the requirement is met.

[0033] The well connectivity analysis method using asymmetric time data alignment in the embodiment of the present application has the following beneficial effects:

[0034] This application reconstructs the original one-dimensional injection-production time series data into two-dimensional injection-production time series data through the perspective of multiple time windows, enabling well connectivity analysis to incorporate diverse temporal patterns. By deeply characterizing and analyzing such two-dimensional data, we can more accurately capture the complex and subtle temporal dependencies in asymmetric injection-production time series data.

[0035] This application uses frequency-domain analysis and Fourier decomposition technology to finely divide raw reservoir injection and production data into multiple components with distinct frequency characteristics. This allows for independent characterization and analysis of the information carried by each frequency component, effectively avoiding inaccuracies and misleading information caused by noise smoothing and overfitting in well connectivity analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of the well connectivity analysis method using asymmetric time data alignment according to an embodiment of the present application;

[0037] Figure 2 Another flowchart of the well connectivity analysis method using asymmetric time data alignment according to an embodiment of the present application is provided;

[0038] Figure 3 Schematic diagram of the overall structure of the asymmetric time-aligned well connectivity analysis method. DETAILED DESCRIPTION

[0039] The present application will be further described below with reference to the accompanying drawings and embodiments.

[0040] In the following description, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. The following description provides multiple embodiments of the present invention, and different embodiments can be replaced or combined, so this application can also be considered to include all possible combinations of the same and / or different embodiments described. Therefore, if one embodiment includes features A, B, and C, and another embodiment includes features B and D, then this application should also be considered to include embodiments containing one or more of all other possible combinations of features A, B, C, and D, even though such embodiments may not be explicitly described in the following text.

[0041] Example 1

[0042] The complex process of oil reservoir production is closely interconnected, forming a vast data network. Injection and production dynamics data is a crucial component of this network. During the injection and production process, various devices and sensors record a vast amount of data in real time, reflecting the dynamic changes within the reservoir. This data, including injection volume, production volume, pressure changes, and temperature fluctuations, varies continuously over time, forming a series of time series. These time series data are more than just a collection of numbers; they contain a wealth of information. First, they provide a visual representation of the reservoir's production status at different points in time, providing decision-making support for engineers and technicians. Second, this data is inherently closely related to interwell connectivity. Therefore, the time series within injection and production dynamics data should be fully considered and utilized in connectivity analysis.

[0043] When injection and production data are formed into time series, they often exhibit complex multi-temporal patterns, which lead to data asymmetry in the temporal dimension. Specifically, different types of injection and production data may have different acquisition frequencies and patterns. For example, liquid production data is typically collected on a daily basis, as a daily acquisition frequency is sufficient for daily reservoir management. However, water injection data is different. Because water injection operations require more sophisticated control and monitoring, data collection frequencies are often higher, sometimes even every half hour. When attempting to apply these time series data to connectivity analysis, this asymmetry directly affects the accuracy and reliability of the analysis and processing. Furthermore, data with multi-temporal patterns may complicate analytical models, reducing the efficiency and real-time nature of the analysis.

[0044] In order to process these injection and production data that are asymmetric in time dimension, this paper proposes an inter-well connectivity analysis method based on asymmetric time data alignment, such as Figure 1As shown in the figure, it includes: S1, constructing a preliminary representation of the original one-dimensional injection-production time series based on graph convolution, and building a graph structure framework for the inter-well connectivity analysis model; S2, reshaping the one-dimensional injection-production time series structure into a two-dimensional injection-production time series structure according to different time window settings in the spatial domain to extract the inter-well connectivity relationship features; S3, characterizing the inter-well correlation degree of the two-dimensional injection-production time series data in the frequency domain through Fourier transform; S4, training the graph neural network inter-well connectivity model after feature extraction, and using the reconstruction error of the injection-production time series law to characterize the inter-well connectivity.

[0045] This application uses frequency-domain analysis and Fourier decomposition technology to finely divide raw reservoir injection and production data into multiple components with distinct frequency characteristics. This allows the information carried by each frequency component to be independently depicted and analyzed, effectively avoiding inaccuracies and misleading information caused by noise smoothing and overfitting in well connectivity analysis.

[0046] Example 2

[0047] During the exploration and development of a certain oil reservoir block, accurate assessment of interwell connectivity requires analysis of the block's production and water injection data. Considering that production data is collected every 10 minutes, while water injection data is collected daily, these two different data collection frequencies present challenges for analysis. This example uses a method based on graph convolution, multi-scale time window reconstruction, and Fourier transform to extract interwell connectivity features from the production and injection data, ultimately generating a representation of interwell connectivity.

[0048] Step 1: Construct a preliminary representation of the original one-dimensional injection-production time series based on graph convolution, and build a graph structure framework for the inter-well connectivity analysis model.

[0049] Step 1.1: Obtain injection and production dynamic data for the block under study from the reservoir monitoring system. Because the injection and production sampling intervals differ, this constitutes original one-dimensional injection and production time series data that is asymmetric in the temporal dimension. Simultaneously, obtain geological data for the block under study; this data includes the number of injection wells, the number of production wells, the distances between each production well and each injection well, the distances between each production well and each production well, the distances between each injection well and each injection well, the relative positions between each production well and each injection well, the relative positions between each production well and each production well, and the relative positions between each injection well, as well as perforation parameters, layer depth, effective thickness, permeability, and porosity.

[0050] In step 1.2, when using graph convolutional networks (GCNs) for well connectivity analysis, feature embedding is performed on the original one-dimensional time series data. Through the combined action of graph convolutional mapping and position encoding, the original one-dimensional injection and production time series, which is asymmetric in the time dimension, is input into the embedding module for processing and feature transformation, thereby obtaining a preliminary representation of the injection and production data time series. Here, we assume that the convolution kernel of the GCN is k and the output dimension is d. At the same time, position encoding is introduced to handle the asymmetry in the time dimension. Through the combined action of GCN and position encoding, the original one-dimensional time series data is input into the embedding module, feature transformation is performed, and a preliminary time series representation is obtained.

[0051] Step 1.3, build a graph structure framework of the well-to-well connectivity analysis model; based on the geological data obtained in step 1.1, establish a graph neural network well-to-well connectivity analysis model with different nodes representing each production well and each water injection well; wherein, when the node represents a production well, the characteristic of the node is the liquid production of the production well in the formation; when the node represents a water injection well, the characteristic of the node is the water injection volume of the water injection well in the formation; the edges of the graph neural network model are used for the distance and relative position between a well point and other well points; a well point is a production well or a water injection well, and the other well points are production wells or water injection wells.

[0052] The input end of the graph neural network well-to-well connectivity analysis model is used to input the liquid production of each production well and the water injection volume of each water injection well; the output end of the graph neural network well-to-well connectivity analysis model is used to output the predicted liquid production of each production well.

[0053] Step 2: Based on different time window settings in the spatial domain, the one-dimensional injection-production time series structure is reshaped into a two-dimensional injection-production time series structure to extract the inter-well connectivity relationship features.

[0054] In step 2.1, based on the acquisition frequency of production and injection data for the block under study, select multiple different time windows as the period P of the reshaped 2D injection-production time series structure. For example, you can choose 1 hour, 6 hours, 12 hours, etc.

[0055] In step 2.2, for each selected period P, the original one-dimensional injection-production time series data is segmented into time periods P and reshaped into a two-dimensional matrix of P × S. S represents the number of periods within that time period. This method extracts inter-well connectivity features in the spatial domain.

[0056] This structure is designed to simultaneously display the changing trends of data within a single cycle, as well as the changes and relationships between data across S cycles. This conversion method not only preserves the temporal characteristics of time series data, but also, through spatial display, allows injection and production data with different periodic attributes to blend and interact in two-dimensional space, further exploring and revealing the deep-level information of injection and production time series data.

[0057] Step 3: Characterize the inter-well correlation degree of two-dimensional injection-production time series data in the frequency domain through Fourier transform.

[0058] In step 3.1, the two-dimensional injection-production time series structure is decomposed in the frequency domain by Fourier transform to obtain a series of two-dimensional injection-production time series representations with different frequency characteristics.

[0059] Specifically, the Fourier transform technique is applied to the two-dimensional injection-production time series structure obtained in step 2.2 to perform Fourier decomposition of the signal, thereby obtaining a series of different Fourier components. These components contain the detailed features and overall trend characteristics of the original sequence, and can accurately capture the local changes in the signal in both time and frequency dimensions.

[0060] The specific working principle of the Fourier transform is to first perform a one-dimensional fast Fourier transform operation on each row of the two-dimensional data to obtain the low-frequency component L and high-frequency component H in the horizontal direction of each row. Then, a one-dimensional fast Fourier transform operation is performed again on each column of the row-transformed data to obtain four different frequency component combinations.

[0061] In step 3.2, the inter-well correlation is characterized by the one-dimensional structural Fourier transform component at each independent frequency component, and the multi-scale information is integrated through a fusion strategy. This helps to more fully understand the inter-well connectivity characteristics contained in the injection and production data.

[0062] Step 3.3: Perform inverse Fourier transform on the data after information fusion in step 3.2 to restore the data to a two-dimensional structure. The formula is as follows:

[0063] X 2D * =iFFT(X 2D ')∈R P×S×c

[0064] Where, X 2D * is the reshaped two-dimensional original data obtained in step 2.2, iFFT is the inverse Fourier transform process, X 2D ' is the data after information fusion in step 3.2.

[0065] In step 3.4, based on the multiple reconstructed 2D data sets, these 2D injection and acquisition data structures based on different time patterns are restored one by one to their original 1D state. Through fusion, these data are integrated into a complete and more meaningful new 1D injection and acquisition time series. At this point, the injection and acquisition data are spatially symmetrical. This involves feature extraction and integration, which aims to better infer the correlations between the data.

[0066] Step 4: Train the graph neural network inter-well connectivity model after feature extraction, and use the reconstruction error of the injection-production time sequence to characterize the inter-well connectivity.

[0067] Step 4.1: Use the new one-dimensional injection-production time series data obtained in step 3.4 to train the graph neural network well connectivity analysis model.

[0068] Step 4.2: Based on the reconstruction error between the new one-dimensional injection-production time series data and its original injection-production data obtained in step 3.4, a learning criterion for model training is constructed, and the parameters in the model are gradient updated until the training error is less than 10 -2 If the model training error is less than 10 -2 If the requirements are not met, return to step 2 to reselect the time scale, reconstruct the data, and retrain until the requirements are met.

[0069] In step 4.3, after model training is completed, the well connectivity characterization results guided by the reconstruction error after data alignment are obtained, and the production dynamic data of the liquid production volume of each production well are output.

[0070] This example combines graph convolution, multi-scale time window reconstruction, and Fourier transform methods to successfully process production and injection data with different sampling frequencies and characterize interwell connectivity. This approach not only improves understanding of fluid flow within reservoir blocks but also provides strong technical support for exploration and development of these blocks.

[0071] This application reconstructs the original one-dimensional injection-production time series data into two-dimensional injection-production time series data through the perspective of multiple time windows, enabling well connectivity analysis to incorporate diverse temporal patterns. By deeply characterizing and analyzing such two-dimensional data, we can more accurately capture the complex and subtle temporal dependencies in asymmetric injection-production time series data.

[0072] This application uses frequency-domain analysis and Fourier decomposition technology to finely divide raw reservoir injection and production data into multiple components with distinct frequency characteristics. This allows for independent characterization and analysis of the information carried by each frequency component, effectively avoiding inaccuracies and misleading information caused by noise smoothing and overfitting in well connectivity analysis.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for analyzing well connectivity using asymmetric time data alignment, characterized in that: include: S1, based on graph convolution, constructs a preliminary representation of the original one-dimensional injection-production time series and builds a graph structure framework for the well connectivity analysis model; including: S1.

1. Obtain injection and production dynamic data of the block to be studied from the monitoring system of the reservoir block to construct original one-dimensional injection and production time series data with asymmetry in the time dimension; obtain geological data of the block to be studied; S1.2, when using a graph convolutional network for well connectivity analysis, feature embedding is performed on the original one-dimensional time series data. Through the combined effects of graph convolutional mapping and position encoding, the original one-dimensional injection and production time series, which is asymmetric in the time dimension, is input into the embedding module for processing and feature conversion, thereby obtaining a preliminary representation of the injection and production data time series. S1.3, building a graph structure framework for the inter-well connectivity analysis model; based on the geological data obtained in step S1.1, establishing a graph neural network inter-well connectivity analysis model with different nodes representing each production well and each water injection well; wherein, when the first node represents the first production well, the first node is characterized by the liquid production of the first production well stratum; when the second node represents the second water injection well, the second node is characterized by the water injection volume of the second water injection well stratum; the edges of the graph neural network model are used to represent the distance and relative position between the third well point and other well points; the third well point is a production well or a water injection well, and the other well points are production wells or water injection wells; The input end of the graph neural network well connectivity analysis model is used to input the liquid production of each production well and the water injection rate of each water injection well; the output end of the graph neural network well connectivity analysis model is used to output the predicted liquid production of each production well; S2, based on different time window settings in the spatial domain, reshapes the one-dimensional injection-production time series structure into a two-dimensional injection-production time series structure to extract inter-well connectivity relationship features; including: S2.1, based on the acquisition frequency characteristics of the production and injection data of the block to be studied, select multiple different time windows as the period P of the two-dimensional injection-production time series structure to be reshaped; S2.2, for each selected period P, the original one-dimensional injection-production time series data is segmented into P-time segments and reshaped into a P×S two-dimensional matrix, where S represents the number of periods contained in that time segment. This method is used to extract the inter-well connectivity feature in the spatial domain. S3, characterizes the inter-well correlation of two-dimensional injection-production time series data in the frequency domain through Fourier transform; including: S3.1, decompose the two-dimensional injection-production time series structure in the frequency domain by Fourier transform to obtain a series of two-dimensional injection-production time series representations with different frequency characteristics; S3.2, characterize the inter-well correlation degree of the one-dimensional structural Fourier component under each independent frequency component, and integrate the multi-scale information through fusion strategy; S3.3, perform inverse Fourier transform on the data after information fusion in step S3.2 to restore the data to a two-dimensional structure. The formula is as follows: Where, is the reshaped two-dimensional original data obtained in step S2.2, is the inverse Fourier transform process, is the data after information fusion in step S3.2; S3.4, based on the obtained multiple reconstructed two-dimensional data sets, restoring the two-dimensional injection-acquisition data structures based on different time patterns to their original one-dimensional states one by one, and integrating them into new one-dimensional injection-acquisition time series data by fusion, wherein the new one-dimensional injection-acquisition time series data conforms to a spatially symmetric relationship; S4, the graph neural network inter-well connectivity model after training feature extraction, uses the reconstruction error of the injection-production time sequence law to characterize the inter-well connectivity.

2. The well connectivity analysis method based on asymmetric time data alignment according to claim 1, characterized in that: The geological data include the number of injection wells, the number of production wells, the distance between each production well and each injection well, the distance between each production well and each production well, the distance between each injection well and each injection well, the relative position between each production well and each injection well, the relative position between each production well and each production well, and the relative position between each injection well and each injection well, as well as perforation parameters, layer depth, effective thickness, permeability and porosity.

3. The well connectivity analysis method based on asymmetric time data alignment according to claim 1, characterized in that: The period P is 1 hour, 6 hours or 12 hours.

4. The well connectivity analysis method based on asymmetric time data alignment according to claim 1, characterized in that: Step S3.1 includes: applying Fourier transform technology to the two-dimensional injection-production time series structure obtained in step S2.2 to realize Fourier decomposition of the signal, thereby obtaining a series of different Fourier components, which contain detailed features and overall trend characteristics of the original sequence.

5. The well connectivity analysis method based on asymmetric time data alignment according to claim 4, characterized in that: Step S4 includes: S4.1, using the new one-dimensional injection-production time series data obtained in step S3.4, training a graph neural network well connectivity analysis model; S4.2, constructing a learning criterion for model training based on the reconstruction error between the new one-dimensional injection-production time series data and its original injection-production data obtained in step S3.4, and performing gradient updates on the parameters in the model until the training error meets the preset conditions; if the preset conditions for the model training error are not met, returning to step 2 and retraining until the requirements are met; S4.3, after model training is completed, the well connectivity characterization results guided by the reconstruction error after data alignment are obtained, and the liquid production dynamic data of each production well are output.

6. The well connectivity analysis method based on asymmetric time data alignment according to claim 5, characterized in that: In step S4.2, the learning criteria for model training are constructed and the parameters in the model are updated with gradients until the training error is less than 10 -2 If the model training error is less than 10 -2 If the requirement is not met, return to step 2 and retrain until the requirement is met.

Images