Research and Application of a Coal and Gas Outburst Prediction Model Based on IPSO-SVM

Pan Jingtao¹,², Zhang Bohan¹, Zhao Dan², Shu Chang³
¹ School of Mining, Liaoning Technical University, Fuxin 123000, Liaoning
² College of Safety Science and Engineering, Liaoning Technical University, Fuxin 123000, Liaoning

Abstract: Coal and gas outbursts represent major safety hazards in China's mining production, posing severe threats to underground personnel and constraining mining efficiency. To address this issue, this paper comprehensively investigates the occurrence mechanism and key influencing factors of coal and gas outbursts through theoretical analysis, model construction, and practical application. Using grey relational analysis combined with real-time mine monitoring data, we quantitatively calculated the correlation degree between ten preselected influencing factors and outburst intensity target values. The calculations demonstrated the effect weight of each factor on outburst intensity, providing data support for index optimization. Building upon this foundation, we innovatively propose an Improved Particle Swarm Optimization (IPSO) algorithm that enhances global optimization capability by dynamically adjusting inertia weight and learning factors, thereby providing optimized support for parameter configuration of the Support Vector Machine (SVM) model and forming an IPSO-SVM coupled prediction model. This model addresses the insufficient accuracy of traditional prediction methods to a certain extent and improves the dynamic identification and early warning capabilities for outburst risks. The research findings provide theoretical basis and technical means for scientific prevention and control of coal and gas outbursts. Additionally, the results can reduce accident rates through risk prediction, promote intelligent upgrading of coal mine safety management systems, and achieve dual improvements in safety benefits and resource exploitation efficiency.

Keywords: Coal and gas outburst; Support Vector Machine; Improved Particle Swarm Optimization Algorithm; Grey correlation degree; Prediction model

1 Introduction

China's energy structure is characterized by "abundant coal, scarce oil, and limited gas." Although renewable energy sources such as wind and solar power are developing rapidly, coal's pivotal position in energy supply cannot be replaced in the short term. Particularly in the process of advancing carbon emission control goals, coal serves not only as a guarantee for energy security but also as a key support for stable supply. However, coal mine safety issues have become a pain point for the industry. Over the past six years, coal mine accidents have occurred frequently with persistently high casualties, among which coal and gas outbursts are extremely destructive, causing both personnel casualties and mine production shutdowns. These accidents are triggered by complex factors, including sudden changes in geological conditions, peak gas pressure, or coal structure fragility, all of which can lead to sudden disasters. To reduce accident occurrence, relying solely on empirical prevention and control is insufficient; it is necessary to understand the underlying patterns from a scientific perspective and establish more accurate early warning measures.

To eliminate potential bias introduced by the original data order, we selected 28 sets of typical coal and gas outburst monitoring data from publicly available National Bureau of Statistics data to construct a sample set, which was randomly shuffled using the randperm function in MATLAB R2018b. The data were divided into training and test sets at a 7:3 ratio, with 20 sets used for model training to learn feature patterns and adjust parameters, and the remaining 8 sets used to validate model prediction accuracy and generalization capability in practical scenarios, ensuring reliability under complex working conditions. The outburst intensity is classified into four risk levels: Level 1 indicates no outburst risk and no outburst intensity; Level 2 represents general outburst risk with relatively small intensity; Level 3 indicates moderate outburst risk with medium intensity; and Level 4 represents severe outburst risk with relatively large intensity. This paper employs the normalization function built into MATLAB R2018 software to standardize the collected data, transforming raw data into dimensionless pure numerical values within the [0,1] interval to eliminate the influence of data units. The normalized results are shown in [TABLE:1].2.

Domestic and international scholars have conducted extensive research on coal and gas outburst prediction. Former Soviet scholar Frid analyzed the coal rock fracture process and discovered a strong correlation between electromagnetic radiation signal anomalies and outburst disasters, providing a theoretical foundation for electromagnetic monitoring and early warning. Chinese scholar Wang Jian achieved precise prediction of outburst areas under small sample conditions using a Particle Swarm Optimization-Support Vector Machine (PSO-SVM) model. Wang Pengfei improved algorithm convergence performance through an improved compression factor optimization strategy (FPSO). Shi Y et al. proposed a globally optimized particle swarm algorithm that further enhanced parameter optimization efficiency. The successful application of Support Vector Machine (SVM) and its improved models in fault diagnosis fields (such as diesel engine fault detection) also provides technical references for coal mine disaster prediction. However, existing methods still have significant limitations: SVM relies on manual experience for parameter selection, which can easily lead to insufficient model generalization ability; traditional Particle Swarm Optimization (PSO) is prone to falling into local optima in complex problems and has slow parameter optimization speed. These issues constrain the accuracy and practicality of prediction models, necessitating breakthroughs through algorithmic innovation.

To address these problems, this paper proposes a coal and gas outburst prediction model based on an Improved Particle Swarm Optimization (IPSO) algorithm optimizing Support Vector Machine (SVM), designated as IPSO-SVM. This study employs grey relational analysis to quantify the correlation between ten primary influencing factors and outburst intensity, using them as model inputs while excluding redundant data interference from core indicators such as gas content, ground stress, and coal hardness coefficient. By dynamically adjusting inertia weight and learning factors, the global search capability of the particle swarm optimization algorithm is enhanced to avoid premature convergence. The penalty coefficient and kernel function parameters of SVM are optimized to improve the model's fitting efficiency for nonlinear relationships. To validate model performance, this study comprehensively compares training results of three models—SVM, PSO-SVM, and IPSO-SVM—using both publicly available National Bureau of Statistics data and monitoring data from multiple excellent coal mines in Shanxi Province. Experimental results demonstrate that compared with traditional PSO-SVM, IPSO-SVM improves learning speed by approximately 18% and achieves a prediction accuracy of 93.7%. The model exhibits stronger adaptability to data from different working faces and demonstrates stable early warning capabilities, particularly under complex geological conditions. The successful application of IPSO-SVM not only provides a dynamic, high-precision prediction tool for coal and gas outburst risks but also optimizes outburst prevention measures through real-time monitoring and early warning, substantially reducing accident rates. This achievement will provide technical support for the intelligent transformation of coal mine safety management and help the coal industry achieve both safety and sustainable development goals under the "dual carbon" targets.

1.2 Parameter Selection and Prediction Results Analysis

In constructing the nonlinear Support Vector Machine (SVM) classification model, we employed the Radial Basis Function (RBF) as the kernel function and conducted parameter optimization based on a grid search method combined with cross-validation strategy. Cross-validation (K=5) was set up to evaluate model generalization ability, with the search space for penalty parameter C and kernel function parameter g defined as [2⁻⁴, 2⁴] and search step size set to 1, achieving global optimization by traversing parameter combinations within this logarithmic space. To systematically analyze the influence of parameters on model classification performance, we simultaneously generated three-dimensional surface plots and contour maps of parameter optimization, visually characterizing the variation规律 of model accuracy under the synergistic effects of C and g, revealing parameter-sensitive regions and optimal solution distribution characteristics to provide quantitative basis for subsequent algorithm improvement. As shown in [FIGURE:1].1 and [FIGURE:1].2, the first-round parameter optimization results demonstrate that the model cross-validation accuracy reaches 89.4737% based on grid search and cross-validation, with optimal solutions for penalty parameter C and kernel function parameter g being 1.0443 and 2.0885, respectively, confirming that this parameter combination can significantly improve the discriminant performance of the coal and gas outburst classification model.

Building upon preliminary search results, we conducted a second round of refined parameter tuning to further improve cross-validation accuracy. In this phase, we maintained K=5 while narrowing the optimization range for penalty parameter C to [2⁻², 2²] and kernel function parameter g to [2⁻², 2²], with step size adjusted to 0.5 for both parameters. The contour map and 3D view of the second SVC parameter selection are shown in [FIGURE:1].3 and [FIGURE:1].4. After two rounds of refined parameter adjustment and cross-validation, the optimal parameter combination was determined as C=1.1892 and g=1.6818, which were applied to the Support Vector Machine (SVM) classification prediction model. As shown in [FIGURE:1].5, although we found the optimal parameter combination through refined search using cross-validation, the prediction accuracy of traditional Support Vector Machine (SVM) for coal and gas outburst prediction still only reached 77.7778%, with test results falling short of expectations.

2.1 Inertia Weight Strategy Improvement

The inertia weight w is set to decrease nonlinearly with iteration times to optimize particle swarm algorithm performance. In the early algorithm stage, particles are given relatively large inertia weight to facilitate rapid global search and improve search efficiency. The inertia weight variation formula is:

$$w = w_o + (w_i - w_o) (1 + e$$

where $w_i$ is the initial inertia weight setting, $w_o$ is the final determined inertia weight value, $t$ is the current iteration number, and $T$ is the maximum iteration number. [FIGURE:2].1 shows the simulation results. As iteration times increase, the attenuation speed of inertia weight gradually accelerates, which helps particles search within the global scope and maintain population diversity. In the later stage, inertia weight gradually decreases with accelerated attenuation speed, enhancing particle local search ability and accelerating population convergence. The adaptive nonlinear decreasing strategy for inertia weight can effectively balance global and local search capabilities and optimize algorithm performance.

2.2 Population Initialization Strategy Improvement

To ensure population diversity while enhancing algorithm search capability, we employ Tent mapping in chaotic motion to generate chaotic sequences. The generated chaotic sequence $Q_\tau$ replaces random numbers for population initialization and maps it to the search space to enhance population diversity and global search capability. The mathematical formulas can be expressed as:

$$Q_{\tau+1} = \begin{cases}
2Q_\tau, & 0 \leq Q_\tau < 0.5 \
2(1 - Q_\tau), & 0.5 \leq Q_\tau \leq 1
\end{cases}$$

where $Q_{\tau+1} \in [0,1]$ is the generated chaotic sequence.

$$x_i^z = u_p + Q_{\tau+1}(u_p + l_p)$$

where $u_p$ and $l_p$ are the upper and lower boundaries of the search space, which can be used to initialize population positions.

2.3 IPSO Optimization Process

The optimization steps of the Improved Particle Swarm Optimization (IPSO) algorithm are as follows:

Step 1: Initialize population positions $x_i^z$ using Tent mapping and initialize particle swarm parameters, including population size $M$, maximum iteration times $T$, nonlinear inertia weight $w$, initial particle velocity and position, individual learning factor $c_1$ and social learning factor $c_2$, and set the search range for penalty parameter $C$ and kernel function parameter $g$.

Step 2: Update the adaptive nonlinear inertia weight $w$.

Step 3: Calculate the current individual fitness value of each particle and the global optimal fitness value of the particle population.

Step 4: Update each particle's velocity and position based on the target function judgment of individual optimal position and global optimal position.

Step 5: Check whether the algorithm termination conditions are satisfied. If satisfied, end iteration and output optimal parameters; if not satisfied, return to Step 3 to continue iteration.

Step 6: Successfully output the optimal parameters $C$ and $g$ found by the IPSO algorithm. The flowchart is shown in [FIGURE:2].2.

Optimization testing was conducted in the MATLAB R2018b environment. Particle swarm size was set to 40, and maximum iteration times to 1000. To comprehensively evaluate the performance of the Improved Particle Swarm Optimization (IPSO) algorithm, we compared it with Standard Particle Swarm Optimization (PSO) and Adaptive Particle Swarm Optimization (TACPSO). Testing was performed on functions $F_1$, $F_2$, and $F_3$, with specific parameter settings shown in [TABLE:2].3 and [TABLE:2].4. To ensure result stability and reliability, each algorithm was independently run 30 times on each test function. The three-dimensional diagrams and convergence curves of test results are shown in [FIGURE:2].3.

Performance Testing of Improved Particle Swarm Algorithm

To verify the performance of the Improved Particle Swarm Optimization (IPSO) algorithm and evaluate its effectiveness in parameter optimization, we used benchmark test functions as evaluation tools. Five unimodal test functions and three multimodal test functions were selected, as shown in [TABLE:2].1 and [TABLE:2].2, aiming to comprehensively compare and evaluate the optimization capability of the improved particle swarm algorithm.

In experiments on unimodal test functions $F_1$ to $F_3$, we observed that although Standard Particle Swarm Optimization (PSO) had relatively fast initial convergence speed, its final solution accuracy was insufficient and it easily fell into local optima. Particularly on function $F_2$, three improved algorithms—Modified Particle Swarm Optimization (MPSO), Improved Particle Swarm Optimization (IPSO), and Adaptive Particle Swarm Optimization (TACPSO)—showed similar performance with comparable convergence speeds and ultimately achieved the same optimization accuracy. Notably, across these test functions, IPSO consistently achieved better solution accuracy than other algorithms, highlighting its significant advantages in local fine search capability.

Coal and Gas Outburst Prediction Model Based on IPSO-SVM

Utilizing the three improved particle swarm optimization algorithms (MPSO, IPSO, and TACPSO) combined with Support Vector Machine (SVM), we established different coal and gas outburst prediction models. Parameters for each particle swarm were initialized as follows: individual learning factor $c_1$ set to 1.5, social learning factor $c_2$ set to 1.7, model evolution iteration limit set to 100, and maximum population size set to 20, aiming to balance search breadth and efficiency. Additionally, the search range for SVM penalty parameter $C$ was set to [0.1, 100], and kernel function parameter $g$ to [0.01, 1000], as shown in [FIGURE:4].11, [FIGURE:4].12, and [FIGURE:4].13.

The TACPSO-SVM prediction model misclassified test sample No. 9, resulting in prediction accuracy of 88.8889%, as shown in [FIGURE:4].11, exposing its limitation of easily falling into local optimization. In contrast, both MPSO-SVM and IPSO-SVM prediction models demonstrated excellent performance with perfect prediction accuracy of 100%, as shown in [FIGURE:2].4 and [FIGURE:2].5. To further investigate the prediction effects of IPSO-SVM and MPSO-SVM models, we conducted detailed analysis of their optimization comparison diagrams. As shown in [FIGURE:4].14, we observed that with increasing iteration times, the Improved Particle Swarm Optimization (IPSO) reached the optimal fitness value of the swarm first after 15 iterations, and its average fitness was higher than that of the MPSO algorithm, directly revealing performance differences among different improved particle swarm algorithms when optimizing prediction models. Comprehensive analysis shows that the IPSO-SVM prediction model takes 0.7049 seconds for prediction with excellent prediction performance.

3.1 Coal Mine Basic Overview

The Shanxi Provincial Coal Industry Bureau identified coal seam No. 81 as an outburst-prone seam in 2009; the China Coal Technology & Engineering Group Chongqing Research Institute identified coal seam No. 15 as an outburst-prone seam in 2012; and the 2020 mine gas grade identification report showed that the mine's absolute gas emission rate reached 106.75 m³/min with a relative emission rate of 78.78 m³/t, among which the maximum emission rate at the mining face reached 49.63 m³/min and at the excavation face reached 2.29 m³/min, all indicating substantial gas occurrence in the coal seams.

Due to gas occurrence conditions and outburst risk in the mine seams, we decided to adopt the drilling cuttings index method for outburst risk prediction at the coal roadway excavation face. The specific implementation scheme includes: arranging no fewer than 3 prediction boreholes; borehole diameter of 4 mm; and setting the projection depth along the roadway axis direction to 10 meters. Coal and gas outburst monitoring data were collected, with 4-5 groups shown in [TABLE:3].1. Outburst intensity is divided into two levels: Level 1 represents no outburst risk; Level 2 represents existing outburst risk.

We used Support Vector Machine (SVM), Particle Swarm Optimization-Support Vector Machine (PSO-SVM), and Improved Particle Swarm Optimization-Support Vector Machine (IPSO-SVM) prediction models for result prediction. The prediction accuracies of the three models for the test set of outburst monitoring data at this coal mine's excavation face were 84.6154%, 92.3077%, and 100%, respectively. The IPSO-SVM prediction model demonstrated accuracy and reliability in predicting coal and gas outbursts at the excavation face of Pingshu Mine in Shanxi, serving as an effective prediction tool for coal mine safety production.

Coal and gas outbursts are frequent and extremely destructive disasters in mine mining, making rapid and accurate prediction crucial. Since practical sample data are often limited while prediction accuracy requirements are extremely high, the Support Vector Machine (SVM) algorithm demonstrates unique advantages in handling such problems. Therefore, we propose a novel prediction model: the IPSO-based SVM model (IPSO-SVM). The core idea is to utilize the IPSO algorithm to optimize key parameters of SVM, aiming to improve prediction speed and accuracy. We conducted detailed research on this new model and performed sufficient training and validation using actual field monitoring data from coal mines. Validation shows that the IPSO-SVM model exhibits excellent performance in coal and gas outburst prediction tasks, with high prediction accuracy and stable, reliable results. This model provides a new and effective prediction tool and technical approach for coal mine safety production, helping to improve mine disaster early warning capabilities.

References

[1] Pang J, Liang Y, Sun K, et al. Analysis of decoupling between China's economic growth and coal consumption and its influencing factors[J]. China Environmental Science, 2024, 44(2).

[2] Frid V I, Shabarov A N, Proskuryakov V M, et al. Formation of electromagnetic radiation in coal stratum[J]. Journal of Mining Science, 1992, 28(2): 139-145.

[3] Frid V, Vozoff K. Electromagnetic radiation induced by mining rock failure[J]. International Journal of Coal Geology, 2005, 64(1): 57-65.

[4] Wang J. Coal and gas outburst prediction model based on SVM and its application[J]. Shaanxi Coal, 2020, 39(2): 109-113.

[5] Wang P, Ren L, Gao Y. Particle swarm optimization algorithm based on improved contraction factor[J]. Electronic Science and Technology, 2022, 35(5): 14-18+46.

[6] Yang C, Guan Z, Liu Y, et al. A new method for coal and gas outburst prediction based on IAEFA-LSSVM[J]. Mechanical & Electrical Engineering Technology, 2023, 52(2).

[7] Hao Q, Shang H. Cause selection and risk level prediction of coal mine gas explosion accidents[J]. Coal Mine Safety, 2023, 54(10): 243-249.

[8] Cheng X, Chen X. Application of grey relational analysis in coal and gas outburst simulation tests[J]. Coal Technology, 2018, 37(10): 179-181.

[9] Wan Y, Qi J. Application of particle swarm algorithm in coal and gas outburst danger prediction[J]. Shanxi Coal, 2019, 39(4): 75-78.

[10] Li Y, Nan X, Lin W. Coal and gas outburst danger prediction[J]. Industry and Mine Automation, 2022, 48(3): 99-106.

[11] Wang H, Wang X, Zhao X, et al. Analysis and engineering application of the influence of low-level roadway excavation process on gas drainage of underlying coal seams[J]. Journal of Xi'an University of Science and Technology, 2022, 42(5): 909-917.

[12] Alkayem N F, Shen L, Mayya A, et al. Prediction of concrete and FRC properties at high temperature using machine and deep learning: A review of recent advances and future perspectives[J]. Journal of Building Engineering, 2024, 83: 108369.

[13] Fang Y. Research on fault diagnosis and prediction of belt conveyor based on support vector machine[D]. Xuzhou: China University of Mining and Technology, 2019.

[14] Su P, Zhang W. Tabu synchronized random learning factor particle swarm algorithm based on chaotic mapping[J]. Small Microcomputer Systems, 2022, 43(8): 1675-1680.

[15] Shan D. Investigation on sensitive indicators and critical values for coal seam outburst prediction with gas as the dominant factor[J]. Coal, 2021, 30(12): 20-22+93.

[16] Cheng X, Chen X. Application of grey relational analysis in coal and gas outburst simulation tests[J]. Coal Technology, 2018, 37(10): 179-181.

[17] Ziyu T, Dingxue Z. A Modified Particle Swarm Optimization with an Adaptive Acceleration Coefficients[C]. Asia-Pacific Conference on Information Processing. 2009: 330-333.