A novel intelligent risk prediction model for the effectiveness of CO2/N2–ECGD technology

To solve the problems of high sampling requirements and low predictive accuracy resulting from the complexity, uniqueness, and randomness of predicting the risk of the CO2 and N2 injection to enhance coal seam gas drainage (CO2/N2–ECGD) technology. The principal component analysis (PCA) method to reduce the dimensionality of the factor data that contribute to the effect risk of the technology was adopted. And the particle swarm optimization (PSO) method was implemented to search for optimal hyperparameters in support vector machine (SVM) by particle search, as a solution to the traditional SVM hyperparameters optimization problem. A novel risk prediction model using machine learning algorithms for gas injection displacement technology was constructed. The prediction results were tested and compared with those of backpropagation (BP), Random Forest (RF), and Decision Tree (DT) models using data from 29 gas injection displacement field projects in China. The results demonstrated that the SVM model had greater accuracy in prediction than the other three models. Additionally, after PSO optimization and dimensionality reduction, the PCA–PSO–SVM model reached 100% prediction accuracy, while requiring less modeling and operation time. The study provided a reliable and reasonable model for predicting technical effects, along with a theoretical basis for risk management and prevention. First, the technology's influencing indicators were analyzed by examining its mechanisms. Second, we utilized the PCA method to reduce the dimensionality of the factor data that contribute to the risk of the technology's effects. Third, we implemented the PSO method to search for optimal hyperparameters in the SVM through particle search, as a solution to the traditional SVM hyperparameters optimization problem. Finally, the prediction results were tested and compared with those of BP, RF, and DT models using data from 29 gas injection displacement field projects in China. The SVM model was found to have greater accuracy in prediction than the other three models. After PSO optimization and dimensionality reduction, the PCA–PSO–SVM model achieved 100% prediction accuracy while requiring less modeling and operation time. The study presents a valid and reasonable model for predicting technical effects and a theoretical basis for risk management and prevention.


| INTRODUCTION
CO 2 and N 2 injection to enhance coal seam gas drainage (ECGD) is a technology that can effectively improve coalbed methane recovery and has worldwide recognition, as reported in past decades. 1,2Some scholars have recently explored utilizing this technology for the prevention and control of mine gas disasters, achieving notable results. 3,4The injected gases used are typically N 2 , CO 2 , or a combination of both.Due to the higher adsorption capacity of coal for CO 2 in comparison to CH 4 , CO 2 is more effectively adsorbed into the coal matrix, occupying the adsorption site of CH 4 and promoting CH 4 desorption. 5,6Injecting N 2 replaces CH 4 in coal fractures, reducing the partial pressure of CH 4 and promoting CH 4 desorption.In recent years, the advancement of big data and artificial intelligence has led to the use of machine learning (ML) as a nonlinear prediction technique based on self-learning machines.This technique bypasses the complex physical process of gas injection displacement, possesses robust computing and learning capabilities, and finds applications in intricate geological settings in coal mines, including situations, such as coal and gas outbursts, coal and rock strength, and the prediction of spontaneous combustion in goafs. 7However, there is a noticeable lack of research on the prediction of displacement effects based on deep learning algorithms in the context of gas injection displacement technology as applied to coal seams.In the field of safety management research, several techniques have been employed for risk assessment, including empirical, 8,9 analytical, 10,11 numerical, [12][13][14] experimental, 15,16 intelligent, [17][18][19] and expert system 20 approaches.Every method has its benefits and drawbacks.Nonetheless, the traditional prediction method is mainly reliant on manual labor and makes a meticulous evaluation of the likelihood of an event occurring when compared with the intelligent method.In the past, studies have thoroughly discussed all these methods. 21ecently, the use of intelligent ML for predicting the risk of coal mine gas has gained popularity due to its ability to handle complex and nonlinear input-output relationships without explicit programming, thus replacing the need for human intervention.ML surpasses traditional technology by tackling high-complexity problems without manual operation.Previous research, for example, Zhao et al., 22 aimed to investigate the use of ML for coal mine outburst prediction.To predict the risk of coal and gas eruption, Wang et al. 7 established the improved grey wolf optimizer-SVM model.ML was used to study the gas occurrence pattern by Jiao et al. 23 For the prediction of coal spontaneous combustion risk, Wang et al. 24 proposed the SSA-CNN model.However, it failed to address significant concepts such as the constraints of conventional assessment techniques in long-term and short-term forecasting methods, and the feasibility of ML in overcoming limitations.Similarly, the model description pertaining to supervised and unsupervised ML methods, which are widely used in the study of coal mine risk assessment, has not been provided.The terms related to data-driven modeling methods are vague and unspecified, and there is no discussion of parameter descriptions or comparative research on data size used for performance measurement and interpretation.As a result, there are still gaps in the systematic study of ML application in risk assessment of gas injection displacement technology in coal mines, which could benefit researchers in related fields.
Therefore, our work aims to provide a comprehensive and resourceful risk assessment study of the CO 2 / N 2 -ECGD technology in coal mines, covering all relevant aspects.The work is divided into five sections.Section 1 introduces an exposition of the research background and purposes.Section 2 outlines the principle and influencing factors of the CO 2 /N 2 -ECGD technology.Section 3 provides a detailed introduction of working mechanisms to the support vector machine (SVM) algorithm, particle swarm optimization (PSO) algorithm, and principal component analysis (PCA) methods.Next, Section 4 constructs the risk assessment model for the effectiveness of CO 2 /N 2 -ECGD technology.Section 5 discusses the outcomes of utilizing an ML algorithm for predictions.
Finally, the conclusion of the research paper is summarized along with future prospects.

ECGD technology
Coal is the primary energy source in China, and gas provides an effective and environmentally friendly resource in conjunction with coal.Efficient gas extraction is essential to prevent accidents and gas disasters while also promoting clean energy usage.The structure of China's coal seams is intricate, and the coal permeability is generally below standard.With the gradual depletion of shallow coal resources, coal development has shifted to greater depths and has become commonplace.The low permeability features of coal seams have become increasingly prominent, resulting in a rise in gas drainage difficulty and pressure on gas disaster prevention and control efforts.The conventional negative pressure drainage technique is insufficient in resolving the issue of coal seam low permeability, while predrainage requires an extended period and yields sluggish outcomes, falling short for mining replacement, and subterranean safety production purposes.Thanks to the unceasing progress of science and technology, gas injection displacement technology has now achieved a significant breakthrough in addressing the low permeability in coal seams.The success of the Enhanced Coal Bed Methane (ECBM) test offers a fresh approach for enhancing the efficacy of subterranean gas extraction via gas injection displacement.Yang 4,5 referred to the technology of underground coal seam gas displacement as ECGD based on the concept of the ECBM recovery promotion in surface wells.This technology not only increases the internal gas pressure of the coal body and improves the seepage velocity of mixed gases but also reduces the effective partial pressure of gas, promotes the desorption of adsorbed gas, effectively addresses the issue of reservoir pressure drop in the later stage of drainage, and provides adequate power and a dependable migration channel for the reservoir flow field.This technology has garnered wide interest due to its safety, costeffectiveness, ecofriendliness, and notable enhancement of gas recovery capabilities, with promising prospects for its implementation.
In most areas of China, the coal strength of coal reservoirs is considerably low, preventing the formation of significant coal seam fractures and deep burial of the coal seam, leading to high ground stress and consequently resulting in poor permeability of coal reservoirs.In many instances, the primary hindrance to gas drainage is low permeability.Therefore, enhancing permeability has become a significant point of focus within gas-related research fields in China.It is important to note that the relationship between the pressure of coal seam gas and its content adheres to a typical Langmuir curve.The parameters a and b of this curve indicate the adsorption pressure at saturated adsorption content and 1/2 adsorption capacity, respectively.Taking the Langmuir adsorption characteristics of common coal as an example (as demonstrated in Figure 1A), this study shows that while gas pressure in coal seams can be significantly decreased through drainage (e.g., 2.0-0.74MPa), the residual gas content in the coal seam remains relatively high at around 5.8 m 3 /t.The gas content in coal seams reduced by drainage, on the other hand, is only approximately 3 m 3 /t.
On the basis of the above analysis, it is evident that a significant reduction in coal seam gas content, elimination of the danger of coal and gas outburst, and decrease in gas emission during tunneling and mining can be achieved only by drastically lowering the coal seam gas pressure.However, it should be noted that as gas pressure gradually decreases in the coal seam, the gradient between the coal mining area and drainage borehole is greatly reduced.This leads to a lack of "gas flow driving force," which explains the significant decrease in pure gas flow after the gas drainage enters the depletion period.However, the gas content in the coal seam still remains high, as depicted in Figure 1B.In addition, the displacement process can be viewed as a whole.Where there is seepage of the injection gas and CH 4 within the coal fractures, with the injection gas diffusing from the fractures into the matrix coal body, and the desorbed CH 4 diffusing from the matrix coal body to the fractures, the injection gas and CH 4 adsorbs and desorbs within the pore fractures of the coal matrix, as illustrated in Figure 1C.
Gas drainage is a process of gas flow that results from a combination of "power" (gas pressure) and "resistance" (coal permeability).To rapidly reduce the gas content in coal seams by gas drainage, it is no longer adequate to simply increase the permeability of the coal body and reduce "resistance."On the basis of the two-stage mechanism of gas injection displacement established by Lin et al., 6 the process of gas injection displacement in coal seams can be divided into two stages: stages I and II, which are illustrated in Figure 2A,B.
In conventional gas extraction, the pressure of the gas (P 0 ) in the gas-producing borehole is approximately equal to atmospheric pressure, while the gas pressure in the coal deposit is higher than that in the borehole (P 1 > P 0 ).This causes a gas exchange process primarily governed by Darcy flow.The pressure gradient ( P  ) drives gas from the coal deposit into the gas-producing borehole through fissures.With the discharge time progressing, the pressure gradient ( P  ) decreases gradually, and the gas outlet flow from the gas-producing borehole follows a second-order attenuation exponential function curve.Before the injection of displacement gas into the coal seam, the gas pressure between the matrix (P m ) and the fracture (P f ) is in a state of dynamic equilibrium, wherein P m is equal to P f .Under the condition of injecting gas at a constant pressure, the total gas pressure (P = P m + P f ) within the coal seam steadily rises and eventually stabilizes.This, in turn, results in an increase in the fracture pressure differential P  between the coal body and the gas-producing borehole.
As per Darcy's law, the velocity at which gas permeates through the coal body is directly proportional to the fracture pressure difference P  between the coal body and the gas-producing borehole.As the pressure differential increases, the velocity at which gas seeps through the fracture in the coal body also increases.In stage I of gas injection, the primary product is free-state CH 4 in its original form.During the second stage of gas injection, CH 4 adsorbed on the coal matrix's pore surface desorb and migrate towards the gas-producing borehole for production.The continuous injection of foreign gas and the constant outflow of free-state CH 4 from the fracture create a gas concentration gradient between the coal matrix and the fracture.This pressure gradient drives gas exchange between the two areas, promoting the desorption and diffusion of CH 4 into the fracture.FICK's diffusion law is used to describe this gas exchange: where Q i represents the amount of gas that is exchanged between the fracture and matrix, g; C f m is the concentration gradient of different gases between matrix and fracture, g/m 3 ; D i is the diffusion coefficient of different gases; M i is the molar mass of different gases, g/mol; i represents different gases.As gas injection progresses, the concentration gradient between the matrix and fracture steadily decreases along with the gas exchange rate.At this stage, the primary output is CH 4 that has been desorbed from its adsorbed state.

| Analysis of influencing factors of the CO 2 /N 2 -ECGD technology
There are numerous potential factors that impact the effectiveness of gas injection for displacing coal gas.Several complex nonlinear relationships exist among these factors and the associated safety risks.Consequently, studying them through quantitative methods is a difficult undertaking.When utilizing gas injection displacement and flow enhancement to extract gas, it is crucial to assess the suitability of the tested coal seam.On the basis of a statistical analysis of numerous safety risk examples, and taking into account the risk factors affecting coal mine gas safety, as well as relevant literature on the causes of coal mine safety risks, 7,[23][24][25][26][27][28][29][30][31][32][33] the potential influencing factors have been identified as follows: thickness of coal (a), Protodyakonov's coefficient (b), water content (c), ash content (d), porosity (e), permeability (f), gas content (g), gas pressure (h), gas saturation (i), Langmuir pressure constant (j), and Langmuir volume constant (k).
Among these, Protodyakonov's coefficient represents the coal's firmness, reflecting its ability to resist external deformation and damage.Gas injection displacement leads to an increase in air pressure within the pores of the coal.If the tensile stress on the pore wall is enough to overcome in situ stress and tensile strength of the coal, the pore structure of the coal will be destroyed.Protodyakonov's coefficient affects the degree of change in porosity and permeability of coal throughout gas injection displacement.This, in turn, significantly impacts the effectiveness of the CO 2 /N 2 -ECGD technology.Second, the gas pressure indicates the equilibrium pressure of gas in the initial coal seam, which reflects the force of coal gas and determines the gas injection pressure required for the gas injection displacement process, which significantly affects the gas injection displacement effect.However, the full extent of potential influencing factors on the displacement effect of gas injection remains unclear.Therefore, it is essential to establish clear methods for determining the existence and degree of the impact.
The success or failure of gas injection displacement in existing field tests in China is primarily judged based on four aspects: increasing mixed gas flow, increasing pure CH 4 flow, gas utilization, and meeting the extraction standard.Hence, the criteria for determining the efficacy of gas injection displacement include raising the flow rate of mixed gas (α 1 ), increasing pure CH 4 flow (β 1 ), reducing gas volume fraction (γ 1 ), and achieving the gas drainage standard (δ 1 ).The calculation formulas are shown in Equations ( 2)-( 5).Gas drainage is typically considered to meet standard requirements when α 0 1  , β 0 1  , and γ 50.46% 1  (assuming the positive value average of γ 1 ).As established by the Detailed Rules for Prevention and Control of Coal and Gas Outburst and Coal Mine Safety Regulations of Chinese outburst mines and gas, mines must adhere to the following gas drainage conditions: gas content <8.00 m 3 /t, gas pressure <0.74 MPa, and predrainage rate δ 1 > 30%.Cases meeting the aforementioned four criteria are considered successful examples, and vice versa.
where Q m1 and Q m2 represent the mixed gas volume before and after injecting CO 2 /N 2 to ECGD, respectively, m 3 /min; Q c1 and Q c2 represent the gas volume fraction before and after injecting CO 2 /N 2 to ECGD, respectively, %; m c1 and m c2 represent the gas content before and after injecting CO 2 /N 2 to ECGD, respectively, cm 3 /g.

| Support vector machine
The concept of SVM emerged from the notion of the optimal hyperplane, that is, the linearly separable optimal classification plane. 34As illustrated in Figure 3A, multiple planes can partition a data set into two types of samples.To construct the optimal classification plane, we seek a plane that maximizes the distance between the two types of sample points.Consequently, we refer to this plane as the optimal hyperplane or the optimal classification plane.
Assuming that there is a training sample set, it can be expressed as , where s is the number of training samples and n is the dimension of the input sample.An optimal plane for effective sample classification is required.Assume this plane to be where ω is the weight of the vector and b is the threshold value.
For normalized description of hyperplane, the following formula is used: It can be obtained from formula (7): Suppose is the optimal hyperplane, the margin is the interval between its two hyperplanes, which is recorded as And our purpose is to find the biggest interval of 2 .According to formulas ( 8) and ( 9), we can introduce Lagrange function 35 and get the following formula: where . It is evident that this problem is an instance of convex quadratic programming, which can be reformulated as a constrained optimization problem using the duality principle to solve the corresponding functional.

    y y x x δ δ δ
WANG ET AL.
By solving Equation ( 11), the pertinent parameters of the optimal classification plane ω In the case of linear inseparability, the training samples do not fully satisfy the condition of formula (11).
To address this, we can permit a few sample points to appear between the classification lines.This can be achieved by introducing relaxation variables and a nonnegative penalty coefficient c (c > 0) to constrain the relaxation variables.The corresponding functional can be obtained: Similarly, in the case of linear inseparability, dual problems are addressed for a solution.
Find the linear decision function The SVM was initially created to solve linear classification problems.Further research led to optimization and improvement by incorporating relaxation variables and penalty coefficients, resulting in a set of calculation methods for linear unclassified problems.Through extensive research and development by numerous researchers, SVMs have expanded beyond classification problems to tackle regression problems.This broadens the scope of solutions for real-life regression problems.
When addressing the regression issue, the primary concept behind SVM involves a transformation that separates the high-dimensional sample space from the low-dimensional sample space.The assumption is that a linear function can split the transformed samples, but if it is impossible, then they are linearly inseparable, as depicted in Figure 3C.When the sample undergoes a transformation from a low-dimensional space to a highdimensional space, an increase in the number of dimensions occurs, leading to an extremely complicated model.However, the proposal of using a kernel function, which transforms the inner product operation in highdimensional space to calculate the kernel function in low-dimensional space, 36 effectively solves this issue.Numerous studies have established that a function must be symmetric and satisfy Mercer theorem to qualify as a kernel function.The structural diagram of the SVM, with K x x ( , ) m serving as the kernel function, is illustrated in Figure 4.
It is clear that the kernel function plays a central role in predicting the regression of the entire SVM and is critical to solving the problem presented.The following three types of kernel functions are generally used.
(1) Polynomial kernel function: where u is the degree of polynomial, which has a great relationship with the vapnik-chervonenkis dimension of the whole system.(2) Sigmoid kernel function: To satisfy the Mercer condition, it is necessary to appropriately select the values of v and a in the formula.
(3) Radial basis function (RBF): where γ is the scale parameter of the function, which not only affects the sample distance, but also directly determines the fitting result of the final function.The versatility of the RBF kernel function makes it a widely used tool in various research fields.By transforming dimensions for uncertain samples, it can still provide the best response, resulting in a more accurate fit.The RBF kernel function is chosen in this work.

| Particle swarm optimization
PSO is a swarm intelligence optimization algorithm proposed by American researchers James Kennedy and Russell Eberhart in 1995.They developed this algorithm after studying the behavior of birds during the predation process.The objective of PSO is to find the optimal solution to a problem by iteratively adjusting a population of potential solutions, or particles, within a search space.PSO has been widely applied in various fields, including engineering, finance, and image processing.The process of finding the best target using PSO involves the following steps.Initially, birds search for food independently or in groups of three to five, and there are numerous paths to find food.With an increasing number of birds seeking food, there must be an optimal path from the nest to the feeding location in the group, despite some birds not discovering a better feeding path and being unable to do anything.Finding food quickly, they will determine their future direction based on their current position and observe the state of other group members, constantly revising and updating their flight direction and speed.This data will be disseminated until an optimal feeding route is discovered.When we apply this pattern to our scientific queries, a nonlinear regression difficulty resembles that of birds seeking sustenance.The act of birds in search of nutrition mirrors that of humans resolving problems.Birds can achieve their goals efficiently by constantly seeking the optimal foraging path, which also serves as the global optimal solution we seek.The fundamental process of PSO algorithm is illustrated in Figure 5.
As can be seen from Figure 5, the basic principle of PSO is that if the number of particles in the population is m and the search space of the population is D dimension, the population can be expressed as X X X X = ( , , …, ) , the flying velocity of i i X ( 1) . The flying speed and position of particles are constantly updated by formulas (19)  and (20): x t x t v t ( + 1) = ( ) + ( + 1), id id id (20)   where v t ( ) is the number of searching dimension; is the number of particles; c 1 and c 2 are the acceleration constant; rand () 1 and rand () 2 are the random numbers evenly distributed in the interval [0, 1]; p t ( ) id is the historical optimal position of the particle at the t iteration; p t ( ) gd is the historical optimal position of the t iteration of the population; x t ( ) id is the position of the i particle in the d-dimensional space.
To enhance the optimization performance of the PSO algorithm, a new parameter, inertia weight w, is incorporated into the formula, resulting in From the formula (21), it is easy to break the equation down into three segments.The initial part on the right side of the equal sign signifies the postiteration velocity of the particles.Following this, we have the second segment that represents the particle's ability to fly towards its optimal position.Postiteration, the particle makes adjustments to its position to locate the best possible position.The third principle states that particles modify their state conditions across generations via the optimal positioning of the entirety.

| Principal component analysis
PCA is an unsupervised feature extraction method that was proposed by Hotelling in 1933.It is used to find patterns and relationships in data by reducing the dimensionality of the data while preserving the important information.PCA is a widely used technique in ML and data analytics due to its ability to handle high-dimensional data and simplify complex datasets.Among them, the principal component refers to linear independent variables, and an orthogonal transformation is used to convert linear related data into a few linearly independent variables.This method is commonly used in ML to handle multidimensional data variables.It is critical to note that technical terminology will be explained when first used, and the overall content will remain objective and concise, with clear causal connections between statements.After being processed by the PCA algorithm, the B-dimensional vector (where A > B) can be obtained, featuring important characteristics.Consistent formatting features, style guides, and citation will be followed while maintaining grammatical accuracy.
This process involves combining various factors into several vital principal component indices.Subsequently, the weights of these principal components are assigned, which are then utilized in creating a comprehensive evaluation index.This index is eventually employed for model prediction.The proposed approach eliminates noise and redundant data in time series information, ensures index independence to prevent information overlap, significantly minimizes computational requirements, enhances prediction accuracy, mitigates the risk of a "dimension disaster," and fully captures data information.
The impact of the CO 2 /N 2 -ECGD technology is dependent on various qualitative and quantitative data.Utilizing the technology for prediction results in using 11 dimensions due to its high data dimension.Predicting with the model will exponentially increase calculations, leading to a possible "dimension disaster."Also, high-dimensional features may hinder learning algorithms from generalizing well, consequently greatly affecting the accuracy of the final gas injection displacement technology effect prediction.This cannot be improved further as it already adheres to the principles or lacks context.The related algorithms of PCA are as follows: Assuming a data set comprising of n samples and a indicators, the a indicators can be considered a random variable, which can be recorded as X X X , , …, a . Through the training and calculation of the PCA algorithm, an index is transformed into multiple new indexes.These data create an independent information index.
where i a = 1,2, …, .Due to the varying definitions and attributes of each dimension of data, standardizing the data can alleviate the prediction error in PCA results due to the magnitude gap.Following standardization, the data appear as follows: Obtain the covariance matrix S of the sample, as follows: The characteristic equation is solved by the covariance matrix, where I is the identity matrix, and the characteristic roots of a indicators are obtained: . The principal component expression is obtained as follows:

⋯
Finally, we obtain the principal components and sort the corresponding eigenvalues by size.The top m principal component data with high contribution rates are selected based on the cumulative contribution rate.The selection principles are outlined as follows: 4 | RISK ASSESSMENT MODEL FOR THE EFFECTIVENESS OF CO 2 /N 2 -ECGD TECHNOLOGY

| Model establishment
According to the principles outlined in the classification prediction algorithm, selecting the correct kernel function is crucial for effectively solving nonlinear regression.The kernel function serves as the centerpiece for the entire model, and the number of its associated parameters determines the complexity of our approach.On the basis of our prioritization of simplicity in the model, we have opted to use a radial basis kernel function in this research.
Compared with a commonly used polynomial kernel function, the radial basis kernel function has fewer parameters, thereby reducing the model's complexity and improving its operational speed.Additionally, this kernel function exhibits strong anti-interference ability towards certain noise data during the data processing stage.On the other hand, the radial basis kernel function efficiently processes prior distribution data and can effectively adapt to statistical samples.However, the Sigmoid kernel function yields similar results to RBF only under specific parameter settings, and it may not be valid with unsuitable parameter configurations.
Radial basis kernel function has two main parameters in SVM, penalty factor C and radial basis kernel function parameter g.The penalty factor C is mainly to control and coordinate the empirical risk and the confidence range.Coordinating these two risks to a reasonable value will minimize the final actual risk and improve the generalization ability of SVMs.If the value of C is small, it will make SVMs underlearn, resulting in a large empirical error, which is not conducive to the generalization of SVMs.On the contrary, if the value of C is large, it will lead to overlearning of SVMs, which only makes the empirical error larger.
Another parameter, g, mainly reflects the complexity of the sample mapping from low-dimensional space to high-dimensional space.A small value of g will make the training error smaller and the test error larger, while a large value of g will increase both the training error and the test error.Therefore, it is particularly important to choose an appropriate value of these two parameters, so that the samples transformed by the kernel function can have a suitable distribution in high-dimensional space, which is of great help to the solution of the whole model and the accuracy of the obtained results.The PSO algorithm to optimize the selection of these two parameter combinations of SVM is used; their optimal combinations in a certain range are determined; their optimal solutions are attained; and then the obtained solutions are used to make regression prediction, which is shown in Figure 6.

| Data collection
According to the analysis in Section 2.2, there are many potential factors affecting the gas displacement effect of the coal body by gas injection, so it is necessary to judge the applicability of the test coal seam when applying the technology of increasing gas drainage by underground gas injection.The data of 29 cases 7,23-33 of the CO 2 / N 2 -ECGD technology in China (ignoring the difference The flowchart of the PSO algorithm optimizing parameters of the SVM.PSO, particle swarm optimization; SVM, support vector machine. of gas injection displacement technology, the default process parameters are the best) are obtained from the public report to explore the influence law of potential impact indicators on gas injection displacement effect (Table 1).
respective mean value is subtracted from each feature value.This enables the data center to be relocated to the origin (zero point) and thus eliminate data deviations at different scales.Next, each sample data's standard deviation is determined and divided by each value.The purpose of this step is to ensure that each feature makes an equal contribution to the model by standardizing the distribution to have a unit variance and avoid imbalances caused by differing feature scales.The normalized value calculation formula for each feature is presented in formula (27).Partial data after pretreatment are shown in Table 2.
Where μ is the average of sample data, and σ represents the standard deviation of features.

| Index selection and model construction
To eliminate information, overlap between indicators, the principal component features of indicator data using the PCA method were extracted.Following the previously established principles and specific implementation steps of PCA, an algorithm in Python software to calculate the cumulative contribution rate of principal component indicators was conducted, the results are displayed in Figure 7.
The number of principal component indicators to be extracted is determined to be 6, since six kernel principal component indicators have been extracted, and the cumulated contribution rate is 87%, which satisfies the dimension reduction condition.
PCA is utilized to analyze data on influential factors.The contribution rate of the principal component is obtained, and feature extraction is subsequently performed based on this rate of contribution.To demonstrate the correlation between the extracted kernel principal component index and the original index, the kernel principal component matrix undergoes rotation through the maximum variance method in factor analysis.This leads to the polarization of the kernel principal component load, enabling convenient analysis and explanation of the kernel principal component index.According to the principles of discriminant validity, the rotated kernel principal component matrix is simplified, and the principal component load table can be achieved, with the results retaining five decimal places.The obtained data are displayed in Table 3.
The principal component loading table explicates the correlation between the original index information and the derived principal component.Table 3 displays that the principal component P 1 is comprised mostly of moisture, gas pressure, gas content, and other factors, which can be classified as "characteristics of gas."Furthermore, it can be observed that the principal component P 2 contains moisture content and ash content.Principal component P 2 heavily influences the industrial analysis parameters of coal, thus deserving the label "characteristics of industrial analysis."Principal component P 3 includes the thickness of coal, Protodyakonov's coefficient, Langmuir volume constant, and others.Given its significant relation to coal thickness attributes, it can be referred to as "characteristics of coal."Principal component P 4 contains Protodyakonov's coefficient, porosity, and Langmuir volume constant.As it heavily influences both permeability and porosity, it can be referred to as the "characteristics of permeability."Principal component P 5 contains the Langmuir pressure constant and Langmuir volume constant and can be named as the "characteristics of gas adsorption."Principal component P 6 includes Protodyakonov's coefficient and gas saturation, which are collectively termed as "characteristics of gas saturation."On the basis of the variance contribution rate, six kernel principal component indexes have been determined, and Python software was used to perform PCA.On the basis of the PCA, the dimension reduction results of the gas emission influencing factors are presented in Table 4.
The calculation results indicate that the risk influence factor index of gas injection displacement technology has been reduced from 11 to 6 dimensions.The reduced dimension data can now serve as input data for the prediction model.
On the basis of the PCA results, the risk prediction index system for gas injection displacement technology can be expressed in Figure 8.The figure shows that PCA extracted influencing factor indicators and represented them as new indicators: characteristics of gas (P 1 ), characteristics of industrial analysis (P 2 ), characteristics of coal (P 3 ), characteristics of permeability (P 4 ), characteristics of gas adsorption (P 5 ), and characteristics of gas saturation (P 6 ).The cumulative contribution rate of principal components reaches 87%.By retaining 87% of the original information, PCA reduces data redundancy, which in turn reduces complexity in the calculation process of neural networks.Additionally, PCA lays a foundation for The results of the model's prediction are displayed in Figure 10.From the figure, it is evident that the PCA-PSO-SVM model prediction results match the actual coal mine results, achieving a 100% success rate.This indicates a low prediction error and a good prediction outcome for the model.The variability of the objective function value throughout the optimization process based on the number of iterations is illustrated in Figure 11.Initially, the value of the objective function is low, but it gradually stabilizes as the number of iterations increases.The 25th iteration results in the value nearing the algorithm's optimum value, providing empirical evidence for PSO rapid convergence and strong global optimization capabilities.From Figure 12, the prediction accuracy of the DT and RF models on the training and test set was 100% and 66.67%.It is evident that while the training set of both the DT and RF models attains 100% accuracy, their prediction effect on the test set is significantly low, revealing inadequate model generalization ability.The prediction accuracy of the BP models on the training set was 90% and 44.45%, while the prediction accuracy of the BP models on the test set was 90% and 77.78%.Consequently, the prediction effect is inferior compared with that of SVM in a small sample data set.Additionally, it has been established by this study that while SVM exhibits good prediction accuracy on small sample data, the prediction accuracy of the BP neural network is not superior to SVM.However, further improvements of the SVM model are still indispensable.

| Comparative analysis of model prediction performance
(2) Comparison of prediction accuracy of the PCA-PSO-SVM, PSO-SVM, and SVM models To determine the superiority of PSO, we compared the SVM model to the PCA-SVM, and PCA-PSO-SVM models based on optimization speed and accuracy.The findings are displayed in Table 5, indicating that the prediction accuracy of SVM, PSO-SVM, and PCA-SVM on the test set was 66.67%, 77.78%, and 88.89%, respectively.
Notably, the rebuilt PCA-PSO-SVM achieved 100% accuracy, which corresponds with actual results.Upon comparison, it is evident that PCA-PSO-SVM yields the highest prediction accuracy, smallest standard error, fast convergence speed, and good robustness, thereby affirming the model's efficacy.

| CONCLUSIONS
The work examines the correlation between the influencing factors of the CO 2 /N 2 -ECGD technology principle with the aid of theoretical analysis.Given the nonlinear and overlapping nature of the influencing factors of technical risk, we propose utilizing the PCA technique to reduce the dimensionality of factor data for the risk assessment of gas injection displacement technology.Additionally, we optimize the SVM using PSO and integrate it with a coupling prediction algorithm.On the basis of Python software and 29 sets of field data, the model is applied to confirm the accuracy of the prediction algorithm.The following specific conclusions are drawn: The correlation between the factors that influence the CO 2 /N 2 -ECGD technology was analyzed through theoretical analysis.As technical risk factors are nonlinear and overlap, PCA is conducted.In addition, the SVM was optimized by PSO and integrated with a coupling prediction algorithm.The following specific conclusions were drawn: (1) Predicting the risk of the impact of gas injection displacement technology is a complex, nonlinear prediction problem due to the overlapping and nonlinear characteristics of the influencing factors.
PCA is proposed to extract the features of the initial index system with 11 influencing factors, resulting in six principal components being obtained from the data of this small sample size.PCA can effectively reduce data redundancy by retaining 87% of the original information.This technique plays a crucial role in reducing complexity in the calculation process of neural networks, which in turn lays a foundation to improve the efficiency and accuracy of risk prediction for the CO 2 /N 2 -ECGD technology.(2) Aiming at the problem that the C and g hyperparameters in the SVM are randomly generated initially, the PSO-SVM prediction algorithm is obtained by coupling the PSO with the SVM.The experimental results of 29 sets of data sets show that the number of iterations is 60, and the prediction accuracy is 88.89%.The prediction algorithm is feasible and reliable.(3) The experimental results indicate that PCA-PSO-SVM is more accurate and efficient than PSO-SVM when compared with the original data set without dimensionality reduction by PCA.Moreover, it implies that in complex nonlinear systems, to enhance prediction accuracy and efficiency, extracting the features of the influencing factors and reducing dimensions is crucial.
In the context of information and intelligence, the presented intelligent risk assessment model can be integrated into the coal mine safety risk information management system to achieve real-time risk prediction and early warning, which is faster and more accurate than conventional manual approaches.However, there are some areas in need of improvement in our research.Due to statistical limitations, we cannot guarantee the inclusion of all influencing factors related to the CO 2 /N 2 -ECGD technology.For instance, factors such as gas injection pressure, flow rate, radius, and time parameters, are not included in our work.These exclusions may affect risk prediction results.It is hoped that this model performance will be further applied and improved in the future.| 2107

F I G U R E 1
Mechanism diagram of the CO 2 /N 2 -ECGD technology.(A) Curve of Langmuir adsorption equilibrium equation, (B) The seepage and diffusion model of gas in coal seam, and (C) Migration model of mixed gas (N 2 /CO 2 ) in coal seam.ECGD, enhance coal seam gas drainage.

F I G U R E 2
Schematic diagram of the two-stage principle of the CO 2 /N 2 -ECGD technology.(A) Stage I (CH 4 in the original free state is the main component of drainage gas) and (B) Stage II (CH 4 in the absorbed state is the main component of drainage gas).ECGD, enhance coal seam gas drainage.

F I G U R E 3
Schematic diagram of support vector machine algorithm model.(A) Linearly separable samples in two-dimensional space, (B) margin maximizing hyperplane and support vectors, and (C) translation of lower dimension data into higher dimensional data to classify nonlinear data.

F I G U R E 4
Schematic diagram of kernel function structure of support vector machine.

F I G U R E 5
The flow diagram of particle swarm optimization algorithm.

( 1 )
Comparison of prediction accuracy of backpropagation (BP), Decision Tree (DT), Random Forest (RF), and SVM To assess the effectiveness of the SVM model, this work compares it to the BP, DT, and RF models.The BP neural network is configured with three layers, 300 training instances, a learning rate of 0.005, and an error rate of 0.001.The data set is separated into training and testing samples following a 7:3 distribution, with the first 20 data groups used for training and the remaining nine groups for testing.The prediction outcomes of each model are shown in Figure 12.

F I G U R E 9
Risk prediction model prediction flowchart of the CO 2 /N 2 -ECGD technology.ECGD, enhance coal seam gas drainage; PCA, principal component analysis; PSO, particle swarm optimization; SVM, support vector machine.

F
I G U R E 10 Prediction results of the PCA-PSO-SVM model.PCA, principal component analysis; PSO, particle swarm optimization; SVM, support vector machine.F I G U R E 11 Variation of model prediction accuracy with iteration times.PCA, principal component analysis; PSO, particle swarm optimization; SVM, support vector machine.F G U R E 12Comparison of prediction results of the BP, Decision Tree, and Random Forest models.BP, backpropagation; SVM, support vector machine.
Relevant parameters of coal seams with gas injection displacement in China. Note: Table of principal component load.
T A B L E 5 Comparison of results of multiple models.