CN117540277B - A lost circulation warning method based on WGAN-GP-TabNet algorithm - Google Patents

Abstract

The invention discloses a lost circulation early warning method based on WGAN-GP-TabNet algorithm, and belongs to the technical field of lost circulation prediction. Comprising the following steps: collecting field data; screening characteristic parameters with strong correlation with the leakage flow and deleting parameters with weak correlation in the field data so as to form initial parameters; classifying the initial parameters according to the leakage flow, and classifying each class into a majority class and a minority class according to the number of each class of parameters after classification; inputting the initial parameters into WGAN-GP model to generate minority class data; training and evaluating the lost circulation early warning model TabNet by using the initial parameters and the generated minority class data; site data is collected and the extent of leakage is predicted using a TabNet model that is trained. The invention effectively solves the problems of low recall rate and few kinds of prediction precision caused by unbalanced kinds in the deep learning leakage prevention and blocking, and has the advantages of stability, reliability, high accuracy, convenient operation, high reaction speed, strong mobility and the like.

Description

A lost circulation warning method based on WGAN-GP-TabNet algorithm

Technical Field

The present invention relates to the technical field of well leakage prediction during drilling, and also to the technical field of data-driven deep learning, and specifically to a well leakage warning method based on a WGAN-GP-TabNet algorithm.

Background technique

In the digital information age, it is becoming more and more a trend to use computers and automation technology to deal with various problems encountered in oil and gas drilling. In the face of downhole leakage problems, improper treatment measures will lead to low plugging success rate, continuous leakage of drilling fluid, increased loss of on-site working time, and even well abandonment. Frequent well leakage problems consume a lot of construction time, plugging construction will increase the drilling cycle, greatly increase drilling costs, and cannot meet the strategic needs of low-cost development.

Data-driven machine learning or deep learning methods seem to be a solution. Data-driven machine learning models rely on various parameters collected during oil field drilling, including drilling parameters, geological parameters, engineering parameters, drilling fluid parameters, etc. Through data processing, feature extraction, model training, model evaluation and other steps, a prediction model for lost formations is established. Deep learning models have advantages over machine learning models when processing large amounts of data, because deep learning models can automatically learn high-level feature representations of data, which means they can extract useful features from raw data without manual feature engineering. Deep learning models are usually composed of multiple layers and can handle a large number of parameters, which enables them to adapt to large-scale data and learn complex relationships. Machine learning models may become inflexible or insufficient to capture patterns in the data when faced with large-scale data, and deep learning models gradually extract abstract information from the data through multi-level representations.

However, in recent years, deep learning models (such as RNN, LSTM, one-dimensional CNN, etc.) have not been as effective as traditional machine learning models in leak prediction. The time when severe leaks occur during drilling is much less than the time when no leaks occur, that is, the number of severe leak categories is much smaller than the number of no leaks, resulting in low accuracy and recall rates for predicting medium and severe leak categories in deep learning models.

Summary of the invention

In order to solve the above problems, the present invention provides a well leakage warning method based on the WGAN-GP-TabNet algorithm, which adopts a multi-learning model TabNet to judge the leakage level, and enriches the minority class data for training and testing the TabNet model based on the WGAN-GP model. The obtained WGAN-GP-TabNet model has a good effect in predicting well leakage.

The specific scheme of the present invention is as follows:

A well leakage early warning method based on WGAN-GP-TabNet algorithm includes the following steps:

Step 1: Collect drilling engineering data from the site, establish a large database for leak prevention and plugging, and pre-process the data;

Step 2: based on the correlation between each characteristic parameter in the preprocessed data and the leakage flow, the characteristic parameters with strong correlation are selected, the characteristic parameters with strong correlation in the preprocessed data are used as initial parameters, and LOESS denoising is performed on the initial parameters;

Step 3: Classify the initial parameters after LOESS noise reduction according to the leakage flow, and classify each level into majority class and minority class according to the number of initial parameters in each level after classification;

Step 4: Input the data processed in step 3 into the WGAN-GP model to generate minority class data;

Step 5: Use the initial parameters after LOESS denoising in step 2 and the minority class data generated in step 4 to train and evaluate the lost circulation warning model TabNet. If the training fails, return to any step of steps 2-4 to modify the relevant parameters and continue training. If the training passes, determine the TabNet model and proceed to the next step.

Step 6: Collect field data and predict its loss degree through TabNet model, wherein the field data includes the characteristic parameters with strong correlation.

As a specific implementation of the present invention, classifying each level according to the number of initial parameters in each level after grading specifically includes: counting the proportion of data of each missing level in the initial parameters after LOESS processing, taking the missing level with the highest proportion of missing data as the benchmark, for all missing levels, if the ratio of its proportion to the proportion of the missing level with the highest proportion is lower than the set classification threshold, then the missing level belongs to the minority class, otherwise it belongs to the majority class.

Furthermore, the classification threshold is set to 20%.

As a specific implementation of the present invention, screening characteristic parameters with strong correlation includes the following steps:

S1. Use the Spearman correlation coefficient, mutual information method and LightGBM models to process the preprocessed data and determine the correlation between each characteristic parameter and the loss flow in each model;

S2. Determine the importance of each characteristic parameter in different models according to the correlation between each characteristic parameter and the leakage flow;

S3. Determine the comprehensive importance of the feature parameter by comprehensively considering the importance of the same feature parameter in different models, and select feature parameters with strong comprehensive correlation.

Further, step S2 includes sorting the characteristic parameters of the same model in order of increasing relevance, and the score of each characteristic parameter is equal to the value of its rank. Step S3 includes calculating the total score of each characteristic parameter, and determining the comprehensive importance of each characteristic parameter according to the total score ranking;

Among them, the calculation formula for the total score of each characteristic parameter is as follows:

In the formula, _Dj is the total score of the jth characteristic parameter, n is the total number of models representing correlation, _fi is the weight of the score of the i-th correlation model in the total score; _Dij is the score of the j-th characteristic parameter in the i-th correlation model.

As a specific implementation of the present invention, the WGAN-GP model is as follows: the generator has 4 hidden layers, and the number of neurons in the hidden layer is 256, 128, 64, and 64; the discriminator has 5 hidden layers, and the number of neurons in the hidden layer is 256, 128, 64, 64, and 32, and the discriminator output layer is 1 neuron; a Dropout layer is set after the hidden layer of the generator and the discriminator, and the dropout rate is 0.25. The discriminator output layer has a node for judging the authenticity of the input sample, and Adam is used as the optimization function.

As a specific implementation of the present invention, step 4 includes the following steps:

When the generator inputs random noise, a category label is added for category guidance, and the category to which the data belongs is determined at the output of the discriminator to generate data by category.

The real samples and generated samples are input into the discriminator together, so that it can learn to capture the characteristics of data distribution, so as to make the generated samples closer to the real samples; while the data distribution is gradually approaching, hierarchical tasks are introduced to further train the discriminator to distinguish between real samples and generated samples, and ensure that the generated samples also have good performance in classification tasks; during the iterative process, the Adam method is used to update the parameters of the discriminator and generator in the WGAN-GP network in turn with the loss value of the discriminator and the loss value of the generator in the current iteration until the generator and discriminator errors.

As a specific implementation of the present invention, the number of minority class data generated in step 4 satisfies the following conditions: after the initial parameters after LOESS denoising and the minority class data generated in step 4 are merged, the ratio of the number of data at the level with the largest number of data to the number of data at each level in the minority class is 5:1.

As a specific implementation of the present invention, the TabNet model is composed of a plurality of decision steps stacked, each of which is composed of a Feature transformer and an Attentive transformer, a Mask layer, a Split layer and a ReLU; if the input sample features have discrete features, TabNet first uses a training embedding method to map the discretized features into continuous numerical features, and then ensures that the data input form of each decision step is a B×D matrix, where B represents the size of the batch size and D represents the dimension of the well leakage parameter; the features of each decision step are output by the Attentive transformer of the previous decision step, and finally the output results of the decision step are integrated into the overall decision;

Compared with the existing technology, it has the following advantages:

(1) The present invention proposes a method for predicting well leakage based on the WGAN-GP-TabNet model. The method is based on deep learning of drilling data and has high accuracy.

(2) To address the imbalance of missing data in drilling data, the present invention uses a generative data model to balance sample distribution and enhance data features.

(3) The present invention proposes a combined extraction method for well leakage characteristic parameters based on correlation coefficient, mutual information and LightGBM. Correlation includes linear correlation and nonlinear correlation, and nonlinear relationship is much more complex and more difficult to describe than linear relationship. At present, it is difficult for a single feature screening method to accurately describe all the correlations between variables, and each method has different indicators for measuring correlation. Therefore, we consider using three methods including linear and nonlinear feature screening methods for comprehensive selection.

(4) The present invention adopts multiple methods (LOESS noise reduction, balanced leakage sample distribution) to enhance the robustness of the model, which has the advantages of stability, reliability, high accuracy, easy operation, fast response speed, and strong portability. It has a positive impact on enabling field engineering personnel to take relevant plugging measures based on leakage warning, maintain the safety of drilling personnel and improve the efficiency of the drilling process.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of the WGAN-GP-TabNet lost circulation prediction system;

Fig. 2 is a feature screening flow chart;

Figure 3 is the loss function change curve of the WGAN-GP generator and discriminator;

Figure 4 is a flow chart of TabNet structure;

Figure 5 is the performance graph of WGAN-GP-TabNet with different ratios of generated data.

Detailed ways

The present invention will be further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

Example 1

FIG1 is a flow chart of the WGAN-GP-TabNet lost circulation prediction system. The specific implementation method will be described in conjunction with this figure, as shown below:

Step 1: Collect drilling engineering data from the site, establish a leak prevention and plugging database, and pre-process the data. This embodiment collects data from a well in the Southwest Oil and Gas Field, which has more than 30 characteristic parameters. The data pre-processing of this embodiment is data cleaning, which includes missing value processing, data interpolation, outlier "3σ" rule detection and elimination, data integration, and data normalization.

Specifically, the Pandas and NumPy libraries in Python can be used for data analysis and processing. Missing values can be detected by the isnull() function and filled by the fillna() function. The "3σ" rule is that for data with normal distribution N(μ,), about 99.73% of the data falls within the range of μ plus or minus 3σ. Data outside the 3σ range are basically missing values and are deleted. Data cleaning also includes deleting blanks, erroneous and abnormal characters, checking the data row and column structure, whether the row and column headers are correct, and whether there are repeated arrays. The calculation formula for data normalization is as follows:

In the formula, is the normalized value of the jth data of the ith feature, x _ij is the value of the jth parameter of the ith feature; _Xi is the set of all parameters of the ith feature, x _ij ∈X _i .

Step 2: Based on the correlation between each characteristic parameter in the preprocessed data and the leakage flow, the characteristic parameters with strong correlation are selected, and the characteristic parameters with strong correlation in the preprocessed data are used as the initial parameters, and LOESS denoising is performed on the initial parameters.

There are more than 30 relevant feature parameters in the preprocessed data. As is known, each feature parameter has a different effect on the loss. Therefore, it is necessary to select the feature parameters with the strongest correlation as the feature parameters in the later model. This embodiment specifically uses the following method to select the feature parameters with strong correlation:

S1. Use the Spearman correlation coefficient, mutual information method and LightGBM models to process the preprocessed data and determine the correlation between each characteristic parameter and the loss flow in each model;

S2. Determine the importance of each characteristic parameter in different models according to the correlation between each characteristic parameter and the loss flow. In this embodiment, the importance of the characteristic parameters is characterized by a scoring method. For example, the characteristic parameters of the same model are sorted in the order of increasing correlation. The score of each characteristic parameter is equal to the value of its rank, that is, the first-rank characteristic parameter is 1 point, and the m-th-rank characteristic parameter is m points;

S3. Determine the comprehensive importance of the feature parameter by comprehensively considering the importance of the same feature parameter in different models, and select feature parameters with strong comprehensive correlation. For example, in this implementation, calculate the total score of each feature parameter, and finally determine the comprehensive importance of each feature parameter according to the total score ranking, and then normalize the score (that is, the sum of the total scores of all feature parameters is 1), select the feature parameters with normalized values greater than 0.01 as the target feature parameters, and finally obtain 16 feature factors, as shown in Table 1.

The calculation formula for the total score of each characteristic parameter is as follows:

In the formula, _Dj is the total score of the j-th characteristic parameter, n is the total number of models characterizing the correlation, which is 3 in this embodiment, _fi is the weight of the score of the i-th correlation model in the total score, which is 1/3 in this embodiment; _Dij is the score of the j-th characteristic parameter in the i-th correlation model.

The calculation formula for normalizing the characteristic parameter values is as follows:

In the formula, is the normalized score of the jth feature parameter, and m is the total number of feature parameters.

Table 1 Final results of three models

S4: The characteristic parameters with strong correlation in the preprocessed data are used as the initial parameters, and the LOESS algorithm is used to reduce the noise of the initial parameters. A local weighted regression is used to fit the data near each data point to obtain a smoother curve or surface without being affected by the global fitting. For each data point x _i , LOESS uses a weight function w(x _i ,x) to adjust the influence of nearby data points.

Step 3: Classify the degree of leakage according to the leakage flow. The specific number of levels can be selected as needed. For example, this implementation determines it as 4 levels. The specific classification standards are shown in Table 2. It can be seen from the table that the data of each level is highly unbalanced. Most of the data is concentrated in Class 0 and Class 1, and Class 2 and Class 3 account for less than 2% in total. Without taking other measures, it is difficult to accurately and comprehensively identify Class 2 and Class 3 with the most serious leakage using only machine learning or deep learning models. Therefore, the proportion of data of various missing levels in the initial parameters after LOESS processing is statistically calculated, and the missing level with the highest proportion of missing data is taken as the benchmark. For the remaining missing levels, if the ratio of its proportion to the proportion of the missing level with the highest proportion is lower than the set classification threshold, then the missing level belongs to the minority class, otherwise it belongs to the majority class. The classification threshold here can be considered as a regulation, but in machine learning or deep learning, generally speaking, when the minority class is less than 20% of the majority class, it can be considered that there is a class imbalance problem. The class imbalance problem will cause the model to not learn enough about the minority class data, resulting in low prediction accuracy and recall rate of the minority class. Therefore, here, the classification threshold is set to 20%, and its classification is shown in Table 2.

Table 2 Classification table of leakage level in a well

Step 4: Input the data processed in step 3 into the WGAN-GP model. WGAN-GP is a generative adversarial network with a gradient penalty term Wasserstein distance. The random noise corresponding to the categories of small dropouts, medium dropouts, and severe dropouts is input into the generator. The output of the generator generates data of the corresponding category, and the data error is discriminated at the output of the discriminator to generate minority class data.

This embodiment constructs the WGAN-GP network as follows: its generator has 4 hidden layers, the number of neurons in the hidden layer is 256, 128, 64, 64, the generator input layer accepts input variables, and the output layer outputs variables; the discriminator has 5 hidden layers, the number of neurons in the hidden layer is 256, 128, 64, 64, 32, the discriminator accepts variables as input, and the discriminator output layer is 1 neuron for judging true or false. In addition, in order to avoid overfitting, a Dropout layer is set after the generator and discriminator hidden layers, with a dropout rate of 0.25. The discriminator output layer has a node for judging the authenticity of the input sample, and Adam is used as the optimization function.

The method of generating data using WGAN-GP is as follows:

(2a) When the generator inputs random noise, a category label is added for category guidance. At the output of the discriminator, the category to which the data belongs (majority class or minority class) is determined, thereby generating data by category.

(2b) The real samples and the generated samples are fed into the discriminator together, so that it can learn to capture the characteristics of the data distribution, so as to make the generated samples closer to the real samples. As the data distribution gradually approaches, hierarchical tasks are introduced to further train the discriminator to distinguish between real samples and generated samples, and ensure that the generated samples also have good performance in the classification task.

(3c) Using the Adam method, the loss value of the discriminator and the loss value of the generator in the WGAN-GP network at the current iteration are used to update the parameters of the discriminator and the generator in the WGAN-GP network in turn (Figure 3).

As can be seen from Figure 3, in 6000 rounds of training, the errors between the generator and the discriminator reach the best around 5500 rounds, so the training can be stopped early and the next step can be taken:

_LG ＝1-D(G(z))

Where L _G is the loss function of the generator; G(·) is the generating function; z is the input noise; D(·) is the discriminator function; L(G,D) is the discriminator loss function; G is the generator; D is the discriminator; E(·) is the expected function; x is the normalized input data; P _t (·) is the true data distribution; λ is the penalty term coefficient; · _p is the p-norm; is the gradient operator; /> is a random interpolation between the real data and the generated data, /> ζ follows a uniform distribution in the range [0,1];/> To sample uniformly between the sampling points from the true data distribution and the generated data distribution.

Step 5: Use the initial parameters after LOESS denoising in step 2 and the minority class data generated in step 4 to train the well leakage warning model for 1000 rounds, and then evaluate the performance of the model on the test set; if the evaluation is qualified, determine the model parameters to obtain the TabNet model, otherwise, continue to train the model.

The generated minority data and the initial parameters after LOESS processing are combined as feature-enhanced data, and the TabNet well leakage warning model is trained with the feature-enhanced data samples. The data set is randomly divided into training set and test set in an 8:2 ratio, and the data is input into the TabNet model. Since the TabNet model hyperparameters have an important impact on performance, the TPE algorithm is used to tune the model parameters and establish a well leakage warning model.

The TabNet model of this embodiment is as follows:

(4.1) TabNet is composed of multiple decision steps stacked together, each of which consists of Feature transformer and Attentive transformer, Mask layer, Split layer and ReLU. If the input sample features have discrete features, TabNet first uses the training embedding method to map the discretized features into continuous numerical features, and then ensures that the data input form of each decision step is a B×D matrix, where B represents the batch size and D represents the dimension of the well leakage parameter. The features of each decision step are output by the Attentive transformer of the previous decision step, and finally the output results of the decision step are integrated into the overall decision, as shown in Figure 3.

(4.2) The function of Feature transformer is to realize the feature calculation of the decision step. Feature transformer consists of BN layer, gated linear unit (GLU) layer and fully connected layer. The purpose of GLU is to add a gate unit on the basis of the original FC layer. The calculation formula is as follows:

Where h(X) is the output of the feature transformer; X is the input feature; W, b are the weight and bias of the fully connected layer respectively; * represents matrix multiplication (matrix dot product); represents the element-level XOR operation; σ is the sigmoid activation function; V and c are the weight and bias of the GLU layer respectively. The feature transformation layer consists of two parts. The first half of this layer belongs to the shared decision step, and the parameters of the shared decision step of the feature transformer of each decision step are shared, while the second half is an independent decision step, and the parameters need to be trained separately at each decision step.

(4.3) The function of the Split layer is to cut the vector output by the Feature transformer. The calculation is as follows:

[d[i]，a[i]]＝ _fi (M[i].f)

In the above formula, d[i] represents the final output of the calculation model, a[i] represents the Mask layer of the next decision step; _fi is a function used to process the i-th element of the vector output by the Feature Transformer; M[i] is the i-th element of the vector output by the Feature Transformer model; and f is the Feature Transformer model.

(4.4) The Attentive transformer obtains the Mask layer matrix of the current decision step based on the output result of the previous decision step, and makes the Mask matrix sparse and non-repetitive.

In addition, in order to overcome the problem of data class imbalance, this embodiment further explores the influence of the amount of data generated at each level in the minority class on the accuracy of the model. Specifically, the minority class sample data is trained using the WGAN-GP algorithm, the 16 well leakage characteristic factors in A are grouped according to Table 3, 8 groups of data with different proportions are designed, TabNet is used as the classification model, and four evaluation indicators of precision, recall, F1 value and G-mean are used on the test set to find the optimal amount of minority class data generated. The results are shown in Figure 5 (in the figure, the horizontal axis number represents the experimental number in Table 3). When the ratio of the number of data of each level in the minority class data to the number of benchmark data (0 leakage level) in the feature-enhanced data (the generated minority class data and the initial parameters after LOESS processing are merged) is 5:1, the effect is best, and the class imbalance is effectively overcome. Therefore, when setting the generation of minority class data in the WGAN-GP-TabNet model, the ratio of the amount of data of each level in the minority class data to the benchmark level in the final feature-enhanced data is 5:1.

Table 3. Dataset after adding WGAN-GP generated data

Step 6: Perform the same data processing as step 1 and extract 16 well-screened well leakage characteristic factors, input them into the trained TabNet model, and the leakage degree can be predicted.

The present invention uses the model to test sample data at 15 different depths, with interval sampling, starting from 500m when the data starts to be recorded, 750m, 1000m, 1250m, 1500m, 1750m, 2000m, 2250m, 2500m, 2750m, 3000m, 3250m, and 3500m. The prediction results are shown in the following table.

Table 4. Leakage prediction results

Depth (*10m) 50 75 100 125 150 175 200 225 250 275 300 325 350 actual value 0 0 0 0 0 1 2 0 3 1 1 0 0 Predictive value 0 0 0 0 0 1 2 0 2 1 1 0 0

The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the embodiments of the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (8)

1. A lost circulation early warning method based on WGAN-GP-TabNet algorithm is characterized by comprising the following steps:

Step 1, collecting drilling engineering data from the site, establishing a leakage prevention and blocking large database, and preprocessing the data;

step 2, screening the characteristic parameters with strong correlation based on the correlation between each characteristic parameter in the preprocessed data and the leakage flow, taking the characteristic parameters with strong correlation in the preprocessed data as initial parameters, and carrying out LOESS noise reduction on the initial parameters;

step 3, classifying the initial parameters after LOESS noise reduction according to the leakage flow, classifying each class according to the number of the initial parameters in each class after classification, and classifying the classes into a majority class and a minority class;

Step 4, inputting the data processed in the step 3 into WGAN-GP model to generate minority class data;

The WGAN-GP model is as follows: the generator has 4 hidden layers with a number of hidden layer neurons of 256,128,64, 64; the arbiter has 5 hidden layers, the number of neurons in the hidden layers is 256,128,64,64,32, and the output layer of the arbiter is 1 neuron; a Dropout layer is arranged behind the generator and the discriminator hiding layer, the discarding rate is 0.25, the discriminator output layer is provided with a node for judging the authenticity of an input sample, and Adam is adopted as an optimization function;

step 5, training and evaluating the lost circulation early warning model TabNet by using the initial parameters after the LOESS noise reduction in the step 2 and the minority class data generated in the step 4;

The TabNet model is formed by stacking a plurality of decision steps, and each decision step consists of Feature transformer and ATTENTIVE TRANSFORMER, a Mask layer, a Split layer and a ReLU; the input sample features are discrete features, tabNet firstly map the discrete features into continuous numerical features by using a training embedding mode, and then ensure that the data input form of each decision step is a B X D matrix, wherein B represents the size of batch size, and D represents the dimension of lost circulation parameters; the characteristics of each decision step are output by ATTENTIVE TRANSFORMER of the last decision step, and finally, the output result of the decision step is integrated into the overall decision;

and 6, acquiring field data and predicting the leakage degree by using a TabNet model which is trained, wherein the field data comprises the characteristic parameters with strong correlation.

2. The lost circulation warning method based on WGAN-GP-TabNet algorithm according to claim 1, wherein the initial parameters after LOESS noise reduction are classified according to the leakage flow, and specific classification standards are as follows:

if the leak rate is 0m ³/h, the leak grade is no leak, and the leak grade is 0;

If the leak rate is greater than 0m ³/h and less than or equal to 5m ³/h, the leak rating is a small amount of leak, the leak rating is 1;

If the leak rate is greater than 5m ³/h and less than or equal to 15m ³/h, the leak rating is medium leak and the leak rating is 2;

if the leak rate is greater than 15m ³/h, the leak rating is severe and the leak rating is 3.

3. The lost circulation warning method based on WGAN-GP-TabNet algorithm according to claim 1, wherein classifying each level according to the number of initial parameters in each level after classification specifically comprises: and counting the proportion of the data of various leakage grades in the initial parameters after the LOESS processing, taking the leakage grade with the highest leakage data proportion as a reference, and if the ratio of the proportion of the data of all the leakage grades to the proportion of the leakage grade with the highest proportion is lower than a set classifying threshold value, the leakage grade belongs to a minority class, and otherwise, the leakage grade belongs to a majority class.

4. The lost circulation warning method based on WGAN-GP-TabNet algorithm according to claim 3, wherein the set classification threshold is 20%.

5. The lost circulation warning method based on WGAN-GP-TabNet algorithm according to claim 1, wherein the step 2 of screening the characteristic parameters with strong correlation comprises the following steps:

S1, respectively processing the preprocessed data by using three models of a Szechwan correlation coefficient, a mutual information method and LightGBM, and determining the correlation between each characteristic parameter in each model and the leakage flow;

S2, determining the importance degree of each characteristic parameter in different models according to the correlation between each characteristic parameter and the leakage flow;

S3, determining the comprehensive importance degree of the same feature parameter in different models by integrating the importance degree of the feature parameter, and screening the feature parameter with strong comprehensive relevance.

6. The lost circulation warning method based on WGAN-GP-TabNet algorithm according to claim 4, wherein step S2 includes ranking the feature parameters of the same model in the order of increasing correlation, the score of each feature parameter being equal to the value of its rank, step S3 includes calculating the total score of each feature parameter, and determining the comprehensive importance of each feature parameter according to the total score ranking;

Wherein, the calculation formula of each characteristic parameter total score is as follows:

Wherein D _j is the total score of the jth feature parameter, n is the total number of models characterizing the correlation, and f _i is the weight of the score of the ith correlation model in the total score; d _ij is the score for the j-th feature parameter in the i-th correlation model.

7. The lost circulation warning method based on WGAN-GP-TabNet algorithm according to claim 1, wherein,

Step 4 comprises the following steps:

Adding a class label to conduct class guidance when random noise is input into the generator, judging the class to which the data belongs at the output end of the discriminator, and generating the data according to the class;

inputting the real sample and the generated sample into a discriminator together, so that the characteristics of the data distribution are learned and captured, and the generated sample is more similar to the real sample; introducing hierarchical tasks while gradually approaching data distribution, further training a discriminator to distinguish real samples from generated samples, and ensuring that the generated samples have good performance on classification tasks; in the iteration process, parameters of the discriminator and the generator in the WGAN-GP network are sequentially updated by using the Adam method according to the loss value of the discriminator in the WGAN-GP network and the loss value of the generator in the current iteration until errors of the generator and the discriminator are reached.

8. The lost circulation warning method based on WGAN-GP-TabNet algorithm according to claim 1, wherein the number of minority data generated in step 4 satisfies the following conditions: after merging the initial parameters after the LOESS noise reduction and the minority class data generated in the step 4, the ratio of the data quantity of the class with the largest data quantity to the data quantity of each class in the minority class is 5:1.