Coal mine methane emission estimation method based on concentration gradient characteristics and deep learning

Abstract

The application discloses a coal mine methane emission estimation method based on concentration gradient characteristics and deep learning, and the method steps are as follows: step A, setting a research area according to mine area coordinates; step B, gridding the regional surface XCH4; step C, obtaining the daily hourly emission source methane net flux F; step D, obtaining the daily emission source average methane net flux F; step E, constructing a deep learning optimization framework, taking the physical model inversion result as initial input, introducing ground observation true value and atmospheric auxiliary parameters, and realizing adaptive correction of the emission through feature fusion and nonlinear mapping; and step F, obtaining the CH4 emission of the research area. Compared with the prior art, the application uses the methane column concentration data obtained by satellite remote sensing, combines meteorological wind field information, quantifies the concentration gradient flux, peels off the regional background interference, generates the initial emission flux under the physical constraint, introduces the deep learning model, and realizes high-precision estimation and optimization of the coal mine methane emission.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of mining area carbon emission monitoring, and particularly relates to a deep learning coal mine methane emission estimation method taking gradient change characteristics of methane concentration as physical constraints. BACKGROUND

[0002] Methane (CH4) is one of the key greenhouse gases affecting global climate change, and its unit mass greenhouse effect is much higher than that of carbon dioxide, and it has strong infrared absorption capacity in the atmosphere. Coal mine is one of the important sources of anthropogenic methane emission, especially in the process of mining and exhaust, which will continuously release a large amount of methane into the atmosphere, posing potential risks to regional air quality and global warming. Coal mine methane emission often shows strong burstiness, complex location distribution and large emission intensity variation, and traditional statistical or estimation methods are difficult to fully and dynamically depict its spatial distribution pattern and variation trend.

[0003] At present, the traditional emission estimation method mainly faces three technical bottlenecks: first, the bottom-up estimation based on the IPCC emission factor method is limited by the static parameter assumption, and it is difficult to reflect the dynamic changes of emission caused by factors such as geological conditions, mining technology and ventilation system in the process of coal mining; second, although the ground monitoring technology can provide accurate point data, it is difficult to realize continuous observation at regional scale due to the complex topographic environment and safety control requirements in the mining area; third, the existing satellite inversion method (such as mass balance method and Gaussian diffusion model) has significant uncertainty in the coal mine application scene, which is caused by the mismatch between the 7×7km² spatial resolution of mainstream sensors such as TROPOMI and the local characteristics of coal mine point source emission, and the disturbance effect of mine topography on wind field, making it difficult to apply to fine management.

[0004] In recent years, the concentration gradient flux analysis method as a kind of emission tracing technology based on the dynamic change of gas concentration field has shown unique value in the field of remote sensing inversion of emission source identification due to its intuitive calculation principle and transparent physical mechanism. This method quantifies the change of concentration gradient flux driven by wind field, so as to locate the potential emission hot spot area, and has significant advantages in spatial resolution and positioning accuracy. However, due to the transmission effect of wind field data error, the cascade amplification of signal noise and the simplification of high-order physical process, the single gradient flux analysis model often faces the challenges of large fluctuation of inversion results and high parameter sensitivity.

[0005] Meanwhile, as an efficient data modeling technology developed in recent years, deep learning shows good non-linear fitting capability in the field of remote sensing, and can automatically extract features from massive observation data to realize complex mapping between high-dimensional input and target output. However, pure data-driven models often lack physical constraints and interpretability, and are prone to overfitting and poor generalization, which limits their application in actual emission estimation.

[0006] Based on the above description, there is an urgent need to develop a coal mine methane emission estimation technology based on concentration gradient characteristics and deep learning, to realize mechanism and data fusion estimation of coal mine methane emission, which can not only retain the physical characteristics of methane concentration-based emission, but also effectively improve the learning ability of complex non-linear emission patterns, so as to realize higher resolution, higher precision and stronger adaptability of methane emission inversion and optimization at regional scale, so as to conveniently and accurately obtain the methane emission flux above the mining area. SUMMARY

[0007] The purpose of the present application solves the problems of low spatial resolution, weak physical constraints of the model, insufficient non-linear feature extraction capability and other problems in the existing coal mine methane emission estimation method, and proposes a coal mine methane emission estimation method based on concentration gradient characteristics and deep learning.

[0008] In order to achieve the above purpose, the technical scheme adopted by the present application is: a coal mine methane emission estimation method based on concentration gradient characteristics and deep learning, referring to Figure 1 , the method steps are as follows,

[0009] Step A, setting the research area with the coordinates of the mining area;

[0010] Step B, based on the XCH4 value of the whole layer atmosphere, the surface CH4 concentration is calculated through the relationship between the bottom layer column concentration and the dry air density, the daily surface XCH4 data is obtained, and the daily surface XCH4 data is regionally segmented according to the research area range, and the gridded regional surface XCH4 is obtained;

[0011] Step C, obtaining the hourly wind speed data of the coal mine area, combining the gridded ECMWF wind field data, calculating the gradient change of the regional surface XCH4, and then multiplying by the air density column, taking the wind direction as the weight, to obtain the daily emission source methane net flux F;

[0012] Step D, averaging the methane net flux of the emission source in a period of time to obtain the daily average methane net flux F of the emission source;

[0013] Step E, constructing a deep learning optimization framework, taking the physical model inversion result as the initial input, introducing the ground observation true value and atmospheric auxiliary parameters (temperature, humidity, pressure), realizing adaptive correction of the emission amount through feature fusion and non-linear mapping;

[0014] Step F, the emission amount of the mine site location point is taken as the inversion methane emission amount, which is taken as the input data of subsequent deep learning, and is corrected as a physical constraint to comprehensively calculate the CH4 emission amount of the research area.

[0015] As preferred, in step B, the product of the research area at the corresponding time is obtained from the CH4 product of TROPOMI, which is gridded at a spatial resolution of 7km x 7km, and the CH4 product of the research area at the corresponding time is obtained from the CH4 product of TROPOMI. The standard is used to screen valid pixels to obtain usable images.

[0016] As preferred, in step B, the entire atmospheric column is divided into 12 layers in the TROPOMI CH4 column concentration inversion, and the bottom layer can represent the surface atmospheric column concentration. The CH4 surface column average dry air mixing ratio, i.e. the surface atmosphere XCH4, is obtained by dividing the CH4 molar content of the atmospheric bottom layer in the INPUT_DATA data set by the molar content of the atmospheric bottom layer dry air, and the derivation formula is as follows,

[0017] (1);

[0018] In the formula, and respectively represent the bottom layer data of the TROPOMI CH4 profile data and the dry air density column used by TROPOMI XCH4. In order to facilitate subsequent calculation, the methane emission report of the coal mine area is arranged as daily emission amount, and the monthly emission total amount is calculated for subsequent data verification.

[0019] As described above, the XCH4 value of the atmosphere is given in the form of XCH4 by the TROPOMI satellite CH4 data product, which is calculated by the CH4 vertical sub-column element dry air column element. The TROPOMI CH4 profile data divides the entire atmospheric column into 12 layers, and the bottom layer is the surface atmospheric column concentration with an average height of about 700m. In the past research, the inversion results of CH4 have been verified in detail. By comparing the corrected TROPOMI XCH4 with the XCH4 measured at 13 TCCON stations, the average deviation of all measurement stations is less than 1%, which is -0.2% (-3.4ppb), and the variability between measurement stations is 0.3% (5.6ppb), both of which are lower than the task requirements of TROPOMI XCH4 inversion.

[0020] As preferred, in order to ensure the use of the highest quality data, the pixels classified in step B are filtered to meet the standard of to ensure data quality.

[0021] ​As preferred, in step C, the wind speed data of the coal mine area is obtained, horizontal wind speed and vertical wind speed, with a resolution of 0.25°, and resampled to the same resolution as the gridded regional surface XCH4, so as to calculate subsequently.

[0022] As preferred, in step C, the methane net flux calculation is performed,

[0023] When the atmosphere is in a steady state, the methane net flux (F), emission and sink continuity equation can be expressed as:

[0024] (2) ;

[0025] Since CH4 is a long-lived gas, it can exist in the atmosphere for 12.5 years, so there is atmospheric CH4 emission and background concentration on the surface, and since F is a linear operator, the daily methane net flux (F ) of the surface can be expressed as:

[0026] (3) ;

[0027] In the formula, is the daily methane net flux of the surface background flux, is the daily methane net flux caused by the emission source.

[0028] Since the sink of CH4 can be ignored, it is assumed that the background concentration is completely uniform, the daily methane emission , that is:

[0029] (4) ;

[0030] However, in most cases, the true CH4 background concentration is not uniform due to the difference in elevation between adjacent surfaces or the difference in albedo, so:

[0031] (5) ;

[0032] The methane net flux (F) can be obtained from the divergence of the horizontal flux (J), that is, the methane net flux can be obtained:

[0033] (6) ;

[0034] In the formula, means to take the divergence of the horizontal flux; the divergence of the horizontal flux (J) is calculated by the Laplace operator;

[0035] J can be divided into zonal flux ( ) and the meridional flux (F ) within the grid, which can be obtained by the product of the vertical column (V) and the horizontal wind field (W ), so for the daily methane net flux (F

[0036] (7) ;

[0037] The daily CH4 vertical column concentration (V ) at the surface can be derived as:

[0038] (8) ;

[0039] where, is the dry air density column corresponding to the surface concentration (V ), combining equations (7) and (8), equation (3) can be rewritten as:

[0040] (9) ;

[0041] where, is the background concentration corresponding to the surface concentration (V ).

[0042] Because it is difficult to obtain the accurate background concentration (V ), the regional background (V ) is used to approximate the background concentration (V ) in this study, for each central grid, its daily regional background (V ) can be obtained by the average of the lower 10% of the surrounding ±3 grid values (total 7x7=49 grids), so equation (9) can be changed to:

[0043] (10) ;

[0044] Equation (10) is applied to the daily methane net flux change (F ) to achieve the calculation of the methane concentration gradient, the methane net flux (F ) over a period of time is averaged to obtain the CH4 emission, as shown in equation (11):

[0045] (11) ;

[0046] where, represents the average methane net flux of the regional background.

[0047] However, we found that in some locations and A significant correlation exists between the values, indicating that the estimated emissions still include some background. Strong spatial positive correlations (R0) are typically found in regions with complex topography and less homogeneous backgrounds. The remaining net background methane flux is due to localized wind fields caused by topography.

[0048] In summary, the daily methane emissions from the mining area can be calculated based on the above steps.

[0049] As a preferred option, the deep learning method in step E is as follows.

[0050] GBDT:

[0051] Step a: Preprocess the methane emission prediction data from the above steps, including time series segmentation and feature extraction. The data processing flow is as follows: Figure 3 As shown;

[0052] Step b: Construct an integrated gradient boosting decision tree and deep learning model for predicting methane emissions, and train the model using the training set data; calculate the loss function at each step, calculate the gradient of the output layer neurons, and update the network parameter values ​​of each layer through forward and backward propagation until the cutoff condition is met, then obtain the optimized integrated gradient boosting decision tree and deep learning model and the parameters of each network. The network structure is as follows: Figure 4 As shown.

[0053] Step c: Collect actual meteorological and emission data, process the data using an optimized integrated gradient boosting decision tree and a deep learning model, and obtain predicted methane emissions.

[0054] As a preferred option, the deep learning model constructed in step b is as follows:

[0055] GBDT:

[0056] 1) The gradient boosting decision tree uses 300 trees trained, each with a maximum depth of 5. The prediction data is processed using R2 and MAE to measure model performance as follows:

[0057] The initialization model expression is as follows:

[0058] (12);

[0059] Wherein, the optimal initial constant is the label of all samples. The mean;

[0060] 2) The residual expression is calculated as follows:

[0061] (12);

[0062] in, For the first Samples in the wheel The residual or negative gradient; The loss function; For the model to sample The predicted value; For the previous iteration model of samples The predicted value; The derivative of the loss function with respect to the predicted values;

[0063] Regression Tree The expression for the fitted residual is as follows:

[0064] (13);

[0065] in, For the first The output of the tree for sample x

[0066] The overall model expression is updated as follows:

[0067] (14);

[0068] in, This is the output of the model after the current m-th iteration; This is the output of the model from the previous round; The learning rate; It is a single tree obtained from the current training; the final model is a weighted sum of all the trees.

[0069] The training and learning method of its model is as follows: MLP:

[0070] 1) The following processing is performed in the deep learning model using the MLP module:

[0071] The standardized dataset is input into the activation layer for activation processing. The activation result is then input into the first hidden layer for linear transformation to obtain the first processed result. The first processed result is then input into the Dropout layer to prevent overfitting, thus obtaining the target result.

[0072] (15);

[0073] The target result, after being activated by an activation function, is input into the second hidden layer for compressed feature extraction, yielding the second processed result. This second processed result is then input into a Dropout layer to prevent overfitting, resulting in the final target result.

[0074] (16);

[0075] After the target result is activated by the activation function, it is input into the third hidden layer for dimensionality reduction, reducing the 32 dimensions to 1 dimension, and obtaining the third processed result.

[0076] (17);

[0077] The output layer obtains the methane emission prediction value through linear transformation:

[0078] (18);

[0079] 2) Calculate the loss function (MSE) to perform gradient descent on the MLP:

[0080] (19);

[0081] Where : true value; : methane emission prediction value

[0082] 3) The training process is automatically completed by the Adam optimizer:

[0083] (20);

[0084] Where: is all the weights of the network; is the learning rate; is the gradient of the loss with respect to the parameters.

[0085] As a preferred, in step b, the model optimization method is as follows,

[0086] The integrated prediction method adopts a weighted average method to obtain the predicted methane emission amount, as follows:

[0087] (21);

[0088] Where, is the gradient boosting decision tree prediction of methane emission, is the deep learning prediction of methane emission, is the final methane prediction emission;

[0089] Calculate the determination coefficient (R2) to verify the goodness of fit of the model:

[0090] (22);

[0091] Where, is the sum of squared residuals, i.e., the sum of squared errors between the model prediction value and the actual value; is the total sum of squares, i.e., the sum of squared differences between the actual value and its mean.

[0092] Compared with the prior art, the present application has the advantages of:

[0093] The present application proposes an innovative methane emission inversion framework, which breaks through the limitations of traditional inversion methods and establishes a dual optimization mechanism of "physical driving + data driving". The method uses satellite remote sensing to obtain methane column concentration data, combines meteorological wind field information, quantifies concentration gradient flux, and strips regional background interference to generate initial emission flux under physical constraints. Then, a deep learning model is introduced to learn the nonlinear relationship between methane concentration and emission source terms based on physical structure constraints, achieving high-precision estimation and optimization of coal mine area methane emission. BRIEF DESCRIPTION OF DRAWINGS

[0094] Figure 1 is a method flowchart of the present application;

[0095] Figure 2 is a research area location and a methane concentration diagram in the third quarter of 2024;

[0096] Figure 3 is a deep learning data processing flowchart;

[0097] Figure 4 is an integrated gradient boosting tree and MLP structure diagram;

[0098] Figure 5 is a comparison result of methane predicted emission results and real emission amounts. DETAILED DESCRIPTION

[0099] The present application will be further described below.

[0100] Anhui Huainan Dingji coal mine is selected as the research area to conduct coal mine methane emission inversion experiment, and the specific experimental steps are as follows:

[0101] Step one, research area setting: taking the mine area coordinates as the research area center, and taking the circumscribed rectangle of the vector boundary of Fengtai County and Panji District as the research area, as shown in Figure 2 .

[0102] Step two, TROPMI methane concentration data processing: download the S5P OFFO (L2) CH4 product from February 2024 to November 2024, grid it with a spatial resolution of 7km x 7km, and select valid pixels with the standard , and obtain 132 usable images.

[0103] Step three, ECMWF wind speed data preprocessing: download the hourly wind speed data of the study area, use the 10m meridional component of wind speed and the 10m zonal component of wind speed. The above data is downloaded through the website TheClimateDataStore. The same data filtering is performed on the 10m meridional component of wind speed and the 10m zonal component of wind speed data using the data filtering parameters provided in the level 2 file of TROPOMI, and the wind speed is resampled to 7km x 7km resolution.

[0104] Step three, (1) Surface CH4 concentration calculation: In the TROPOMI CH4 column concentration inversion, the entire atmospheric column is divided into 12 layers, and the bottom layer can represent the surface atmospheric column concentration. We use the CH4 molar content of the bottom layer of the INPUT_DATA dataset divided by the molar content of the dry air of the bottom layer of the atmosphere to obtain the surface column average dry air mixing ratio of CH4 (surface atmospheric XCH4).

[0105] (2) Regional background concentration calculation: The regional background concentration is used to approximate the background concentration. In this study, for each central grid, the daily regional background can be obtained by averaging the average of the surrounding ±3 grid values (total 7x7 grids) below 10%.

[0106] (3) Methane concentration gradient calculation: In gridding, for each grid, its north-south and east-west direction four adjacent grids are first used to obtain the concentration gradient. Only when all four adjacent grids have valid values, the concentration gradient can be calculated; in order to make full use of data, the northeast, southeast, northwest and southwest corners of each grid are also used for calculation, and the weight is determined by the wind direction. The second-order central difference is used to calculate the spatial gradient of methane concentration, multiplied by the dry air density column of the corresponding surface methane concentration, and finally multiplied by the angle weight, which is determined by the wind direction, that is, the 10m meridional component of wind speed and the 10m zonal component of wind speed, as the calculation result of methane net flux.

[0107] (4) Methane net emission calculation: The calculated hourly methane net emission results are averaged over a day to obtain daily emissions, and so on to obtain monthly emissions, and the emissions of the point where Dingji coal mine is located are extracted as the inversion of methane emissions, which becomes the input data of subsequent deep learning as a physical constraint for correction. Figure 2

[0108] ​Step four, deep learning data preprocessing: select the measured data table of Dingji coal mine emission, including daily emission, meteorological elements (temperature, wind speed, wind direction) and climatic periodic factors. First, limit the abnormal extreme value in the emission (5% to 95% quantile cut-off), and then standardize each variable to ensure the stability of model training. At the same time, the date field is extracted and the seasonal and monthly information is expanded to enhance the time series expression ability.

[0109] Step five, data set construction and time series division: the processed data is converted into tensor format, and the time series cross-validation method is used to divide the whole data into training set and validation set. 90% of the data is used as the training set, and the remaining 10% is used as the validation set, keeping the time sequence unchanged.

[0110] Step six, gradient boosting tree model training: the standardized feature variables are used as input, and the standardized methane emission is used as label. The gradient boosting regression model with 300 trees is used for training, and the maximum depth is set to 5. After training, the prediction is carried out on the whole data set, and the R2 score is 0.88, showing strong fitting ability.

[0111] Step seven, building a simplified neural network model: a three-layer fully connected neural network structure is designed, the hidden layers are 64, 64 and 32 nodes respectively, the activation function is ReLU, 0.2 Dropout is added to enhance the generalization ability, and the final output layer is a single prediction value. The mean square error loss function is used, the optimizer is Adam, the learning rate is 0.01, and the training is 1000 rounds. The neural network is optimized on the training set, and the loss change is monitored on the validation set. After the final training is completed, the prediction is carried out on the whole data, and the real prediction value is obtained by inverse standardization, and the independent performance evaluation result is R2 of 0.84.

[0112] Step eight, result visualization and output: the predicted value and the true value are plotted Figure 5 in time sequence, and the high-definition black and white line chart is output to compare the model effect, Figure 5 the integrated model is clearly shown to fit the actual emission trend at each time point.

[0113] The present application estimates the coal mine methane emission of Dingji coal mine in Huainan City, Anhui Province from February to November 2024 based on the concentration feature and deep learning fusion coal mine methane emission estimation technology. In this area from February to November 2024, the algorithm coal mining monthly average methane emission is 35354.069 , and the measured monthly average methane emission is 33579.07843 , so the precision of the present application is as high as 95%.

[0114] The coal mine methane emission estimation method based on the concentration gradient characteristics and deep learning provided by the present application is described in detail above, and the principles and implementation modes of the present application are described by applying specific examples. The above example is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed, and the modification and improvement of the present application will be possible, and will not exceed the concept and scope defined in the appended claims. Therefore, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A coal mine methane emission estimation method based on concentration gradient characteristics and deep learning, characterized in that: The method steps are as follows, Step A: setting the study area in mine coordinates; Step B: based on the XCH4 value of the whole layer atmosphere, the surface CH4 concentration is calculated by the relationship between the bottom layer column concentration and the dry air density, and the daily surface XCH4 data is obtained, and the daily surface XCH4 data is segmented according to the range of the study area to obtain the gridded regional surface XCH4; Step C: obtain the hourly wind speed data of the coal mine area, combine the gridded ECMWF wind field data, calculate the gradient change of the regional surface XCH4, and then multiply by the air density column to obtain the daily methane net flux F of the emission source with the wind direction as the weight; Step D: average the methane net flux of the emission source in a period of time to obtain the average methane net flux F of the emission source per day; Step E: build a deep learning optimization framework, use the physical model inversion result as the initial input, introduce the ground observation true value and atmospheric auxiliary parameters, and realize the adaptive correction of the emission through feature fusion and nonlinear mapping; Step F: comprehensive calculation to obtain the CH4 emission of the study area.

2. The coal mine methane emission estimation method based on concentration gradient characteristics and deep learning according to claim 1, characterized in that: In Step B, the corresponding time product of the study area is obtained from the CH4 product of TROPOMI, gridded at a spatial resolution of 7 km x 7 km, and the standard screening effective pixels are obtained to obtain the available image. In Step B, the corresponding time product of the study area is obtained from the CH4 product of TROPOMI, gridded at a spatial resolution of 7 km x 7 km, and the standard screening effective pixels are obtained to obtain the available image.

3. The coal mine methane emission estimation method based on concentration gradient characteristics and deep learning according to claim 1, characterized in that: In step B, the whole atmospheric column is divided into 12 layers in the TROPOMI CH4 column concentration inversion, and the bottom layer can represent the surface atmospheric column concentration. The surface column average dry air mixing ratio of CH4 is obtained by dividing the CH4 molar content of the atmospheric bottom layer in the INPUT_DATA data set by the molar content of the atmospheric bottom layer dry air, that is, the surface atmospheric XCH4, and the derivation formula is as follows, (1)； wherein: and represent the underlying data of the TROPOMI CH4 profile data and the dry air density column used by TROPOMI XCH4, respectively.

4. The coal mine methane emission estimation method based on concentration gradient characteristics and deep learning according to claim 2, characterized in that: To ensure that the highest quality data is used, the classified pixels are filtered to meet the screening criteria. of the standard.​ 5. The coal mine methane emission estimation method based on concentration gradient characteristics and deep learning according to claim 1, characterized in that: In step C, the wind speed data of the coal mine area is obtained, the horizontal wind speed and the vertical wind speed, and the resolution is 0.25°, which is resampled to the same resolution as the gridded regional surface XCH4.

6. The coal mine methane emission estimation method based on concentration gradient characteristics and deep learning according to claim 1, characterized in that: In step C, the methane net flux calculation is performed, When the atmosphere is in steady state, the continuity equation for the net flux (F) of methane, emissions and sinks can be written as: (2)； Since F is a linear operator, the daily net methane flux at the Earth's surface ( This can be represented as: (3)； wherein is the daily methane net flux of the surface background flux, is the daily methane net flux of the emission source.

7. The coal mine methane emission estimation method based on concentration gradient characteristics and deep learning according to claim 6, characterized in that: Since the sink of CH4is negligible, the background concentration is assumed to be perfectly uniform, is the daily methane emission i.e.: (4)； However, in most cases, the true CH4 background concentration is not uniform due to the difference in elevation between adjacent surfaces or the difference in albedo, so: (5)； The methane net flux (F) can be obtained from the divergence of the horizontal flux (J), that is, the methane net flux: (6)； wherein denotes the divergence of the horizontal flux; the divergence of the horizontal flux (J) is computed by the Laplacian operator; J can be separated into a zonal flux (Jz) ) and a meridional flux (Jm) ), which can be obtained using the product of the vertical column (V) and the horizontal wind field (W) ), so for the daily methane net flux (F): (7)； Daily vertical column concentration of CH4 over the surface Can be derived: (8)； where is the dry air density column corresponding to the surface of the earth, in conjunction with equations (7) (8), equation (3) can be rewritten as: (9)； wherein is corresponding background concentration; Calculations were made using regional background concentration, utilizing regional background ( To approximate simulation Therefore, formula (9) can be changed to: (10)； Equation (10) is applied to the daily methane net flux change of CH4, and the methane concentration gradient is calculated, and the methane net flux in a period of time The average CH4emission is obtained, as shown in equation (11): (11)； In the formula represents the average methane net flux for the regional background.

8. The coal mine methane emission estimation method based on concentration gradient characteristics and deep learning according to claim 1, characterized in that: In step E, the deep learning method is as follows, Step a: preprocessing the above step methane emission prediction data, time series segmentation and feature extraction are performed on the data; Step b: build an integrated gradient boosting decision tree and deep learning model for predicting methane emission, train the model with training set data; calculate the gradient of the output layer neurons each time the loss function is calculated, and update the network parameter values of each layer by forward and backward propagation until the optimized integrated gradient boosting decision tree and deep learning model and each network parameter are obtained when the stopping condition is reached; Step c: collect actual meteorological and emission data, and use the optimized integrated gradient boosting decision tree and deep learning model to process the data to obtain the predicted methane emission.

9. The coal mine methane emission estimation method based on concentration gradient characteristics and deep learning according to claim 8, characterized in that: In step b, the deep learning model is built as follows, 1) The gradient boosting decision tree adopts training of 300 trees, and the maximum depth of each tree is 5. The R2 and MAE methods are used to measure the model performance of the prediction data as follows: The initialization model expression is as follows: (12)； where the optimal initial constant is the mean of all sample labels ; 2) The residual expression is calculated as follows: (12)； wherein, is the first wheel of samples a residual or negative gradient; is a loss function; is a predicted value of the model for the sample ; is a predicted value of the model for the sample of the previous iteration; is a derivative of the loss function with respect to the predicted value; Regression tree The fitted residual expression is as follows: (13)； wherein, is the the output of the tree for the sample x The overall model expression is updated as follows: (14)； wherein, is the model output after the current mth iteration; is the model output of the previous iteration; is the learning rate; is the current tree obtained from training; the final model is the weighted sum of all trees. The training learning method of the model is as follows: 1) The MLP module is used in the deep learning model as follows: The standardized data set is input into the activation layer for activation processing, and the activated result is input into the first hidden layer for linear transformation to obtain the first processing result. The first processing result is input into the Dropout layer for overfitting prevention processing to obtain the target result: (15)； The target result is activated by the activation function and input into the second hidden layer for feature extraction. The second processing result is input into the Dropout layer for overfitting prevention processing to obtain the target result: (16)； The target result is activated by the activation function and input into the third hidden layer for dimension reduction processing, reducing 32 dimensions to 1 dimension to obtain the third processing result: (17)； The output layer is linearly transformed to obtain the methane emission prediction value: (18)； 2) The loss function (MSE) is calculated for gradient descent of the MLP: (19)； wherein : true value; : methane emission predicted value 3) The training process is automatically completed by the Adam optimizer: (20)； where: is the weight of the network; is the learning rate; is the gradient of the loss with respect to the parameters.

10. The coal mine methane emission estimation method based on concentration gradient characteristics and deep learning according to claim 8, characterized in that: In step b, the model optimization method is as follows, The integrated prediction method adopts a weighted average method to obtain the predicted methane emission amount as follows: (21)； wherein, is a gradient boosting decision tree prediction of methane emissions, is a deep learning prediction of methane emissions, is a final methane predicted emission. The determination coefficient (R2) is calculated to verify the goodness of fit of the model: (22)； wherein, is the sum of squares of residuals, i.e., the sum of squares of errors between the model predicted values and the actual values; is the total sum of squares of variation, i.e., the sum of squares of differences between the actual values and their mean.

Images