Academic Journal of Science and Technology ISSN: 2771-3032 | Vol. 11, No. 1, 2024 197 A new IPCA‐CPSO‐BP Model for Predicting Gas Emission in Underground Coal Mines Guangsheng Chen School of Energy Science and Engineering, Henan Polytechnic University, Jiaozuo, Henan, 454003 China Abstract: Gas has become the primary factor restricting the safe and efficient production of coal mines. Gas emission prediction model is very important for mine gas emission prediction and gas disaster prevention. The improved principal component analysis (IPCA) was used to reduce the dimension of 13 influencing factors of gas emission, and the CPSO-BP neural network prediction model was built with MATLAB software to predict the absolute gas emission of Zhongtai Mining.The results show that the cumulative contribution rate of principal components is not less than 95%, and the number of principal components after improvement is reduced from the original 6 to 3, which effectively solves the problem of excessive principal components and data redundancy caused by the correlation difference between various influencing factors and gas emission amount. At the same time, the optimized particle swarm optimization algorithm alleviates the trouble that the particle swarm optimization algorithm is prone to fall into the local optimal solution. By coupling the improved particle swarm optimization algorithm with BP neural network, the problem of BP neural network's over-dependence on the initial value of weights and thresholds is solved, and the optimal initial weights and thresholds are provided for BP neural network, improving the accuracy of model prediction results. The average relative error of the original BP neural network is 1.6638%, the normalized mean square error is 0.3155, and the regression correlation index is 0.8157. The IPCA-CPSO-BP prediction model decreased to 0.5749%, 0.1143 and 0.9758, respectively. IPCA improves the data dimensionality reduction ability of principal component analysis, IPCA-CPSO-BP enhances the stability of particle swarm optimization algorithm, and significantly improves the prediction accuracy of BP model. The prediction trend is highly consistent with the actual value, which verifies the reliability of the model and provides strong data support for mine gas prevention and control. Keywords: Keywords: Prediction of gas emission; Spearman correlation coefficient; Principal component analysis; Particle swarm optimization; BP neural network. 1. Introduction Mine gas gushing is one of the main causes of gas accidents. According to the requirements of Coal mine Safety regulations, it is necessary to predict the gas emission quantity of coal mining face before mining, so the prediction research of gas emission quantity is very important [1-3].At present, many scholars have carried out a lot of work on the prediction of gas emission and obtained a lot of results. Xie Minglu et al.[4] used the method of multiple linear regression and SPSS software to process the data, carried out multiple regression analysis on the influencing factors of gas emission, and worked out the regression equation for prediction; Lu Guobin et al. [6] used the PCA-BP model to predict the amount of gas emission in the mining face; Zhou Xihua et al.[7] established GA-BP neural network prediction model based on principal factor analysis, and introduced momentum term in the process of weight reverse updating; Xie Xingjun et al.[8] used kernel principal component analysis (KPCA) to reduce the dimensionality of the influencing factors. To solve the problems of slow convergence and easy to fall into local optimal solution in BP neural network (BPNN), they used CMGA to optimize BPNN. The coupled algorithm of CMGA and BPNN (CMGANN) was constructed to predict the gas emission; Lin Haifei et al.[9] based on the monitoring data of absolute gas emission at a coal face in Huangling, Shaanxi Province, applied periodic trend decomposition based on local weighted regression and integrated empirical mode decomposition to construct a SEGS model by optimizing parameters of support vector regression machine through genetic algorithm, and obtained the final prediction result of gas emission. Feng Shengcheng et al.[10] established a prediction model of gas emission in mining face based on PCA and PSO-LS-SVM by taking advantage of the characteristics of fast training speed, global optimal solution and good generalization of vector machine, and successfully applied it in practice.; Xiao Peng et al [11] proposed to combine the wavelet packet decomposition method with the extreme learning machine to build a time-varying prediction model of the gas emission quantity, so as to improve the prediction accuracy and reliability of the gas emission quantity.; Liu Feng [12] used simulated annealing algorithm to optimize particle swarm optimization and established PCA-SAPSO- BP model to predict the amount of gas emission, and the prediction results showed that the model had fast convergence speed and high accuracy.; Fu Hua et al.[13] proposed an algorithm combining kernel principal component analysis and autoregressive integral moving average model, expressed the gas emission amount as a time series, and established a prediction model of the gas emission amount with autoregressive integral moving average, which has high prediction accuracy. Zhang Yucai et al.[14] used WOA, a whale optimization algorithm, to optimize LSTM to predict the amount of gas emission. By comparing the experimental results of different time steps, they found that the model had strong robustness and high accuracy; Zhou Xu et al.[15] proposed a combination of non-negative matrix decomposition NMF and random forest RF, and used the Hunger Games search algorithm HGS to optimize RF parameters, and established a coal seam gas emission prediction model with stable operating results and high precision. 198 The above gas emission prediction methods have achieved results to a certain extent, but the factors affecting the gas emission amount are complex and diverse, and many influencing factors carry a large amount of redundant information to affect the prediction accuracy and running speed of the model. How to optimize the data pre-processing algorithm and improve the accuracy of the prediction results has become an urgent task for the research of prediction models. In order to solve the shortcomings of the above prediction models and improve the prediction accuracy, the IPCA- CPSO-BP model of gas emission prediction is proposed in this paper. By introducing Spearman correlation coefficient into the traditional principal component analysis method, the number of principal components decreases and the information carried by the first principal component increases after data dimensionality reduction. Adaptive inertia weight and contraction factor are added to the PSO algorithm to ensure that the PSO algorithm has a good search effect before and after. It is used to optimize the defects of BP neural network, such as excessive dependence on initial weights and thresholds, and slow convergence, so as to improve the prediction accuracy. 2. Principal Component Analysis (PCA) Method Improved Principal component analysis (PCA) is a common data dimensionality reduction method, which has obvious dimensionality reduction effect in data processing, small error, simple operation, easy implementation and strong practicability, and is widely used in the field of gas emission prediction. 2.1. Influencing factors of gas emission quantity The approved production capacity of Zhongtai Mining mine is 1.5 million tons/year. It adopts underground mining (longwall caving and falling mining method), multi-level down-hill development of vertical shaft and multi-level down-hill mining of Permian Shanxi Formation 21 coal seam. There are 5 vertical shafts and shafts, and the ventilation method of mine is mixed type and extraction type. The ventilation system of the mine is regionalized, and the gas grade evaluation of the mine from 2017 to 2022 is all outburst mine. Based on the field investigation of gas emission in Ⅱ1 coal seam of Zhongtai Mining, the influencing factors of gas emission are selected from geological factors and mining factors [16, 17]：X1 promotion degree, X2 recovery rate, X3 coal thickness, X4 burial depth, X5 mining height, X6 temperature, X7 underground pressure, X8 adjacent coal seam spacing, X9 air volume, X10 daily production, X11 coal seam inclination, X12 adjacent layer gas content and X13 mining coal seam gas content, the prediction work of gas emission Y of Zhongtai mining industry is studied. The original data are shown in Table 1: Table 1. Data of influencing factors of gas emission Group X1/ （m/d） X2/ % X3/ m X4/ m X5/ m X6/ ℃ X7/ KPa X8/ m X9/ （m3/min） X10/ （t/d） X11/ ° X12/ （m3/t） X13/ （m3/t） 1 4 84.5 5.85 625.17 3 19 87.26 3 1000 3591 8 5.2 4.93 2 4.3 85.7 8.87 640.01 3 20 87.14 4 1008 3857 9 5.63 4.98 30 3 84.7 5.04 636.17 3 21 86.95 4.3 967 2856 9 5.41 4.72 2.2. Spearman correlation coefficient improved principal component analysis (PCA) As principal component analysis (PCA) treats each influencing factor equally in the process of data processing, the impact of the influencing factor that is not strongly correlated with the target variable becomes greater after dimensionality reduction. In order to reduce this impact, Spearman correlation coefficient is used to assign weight to each influencing factor. If the influencing factor is strongly correlated with the target variable, the weight is significant. Small weight, this method has no specific requirements on the distribution of variables, can be applied to small samples (sample size does not exceed 30) and large sample range. 2.2.1. Spearman correlation coefficient Spearman's correlation coefficient calculation method first arranges n samples according to the value size of the impact factor from small to large, and then arranges the target variables in the same way, which can be evaluated by substituting them into formula (1): (1) Where: represents the correlation between the feature variable and the target variable, and the larger the correlation, the stronger the correlation; n represents the number of samples; Represents the ranking difference between the feature variable and the target variable. 2.2.2. Improvement step The main steps of Spearman correlation coefficient to improve PCA are as follows: Step 1: Select the original sample matrix ，Where n represents the number of samples and m represents the number of impact factor variables. Step 2: Spearman correlation coefficient is calculated between each impact factor variable and target variable, and the diagonal matrix formed by the calculation result is denoted as , as follows: (2) Where: is the Spearman correlation coefficient between 199 the impact factor and the target variable . Step 3: The original sample matrix is normalized (de-dimensional) according to formula (3) to obtain the matrix. . (3) Step 4: Multiply the normalized matrix with the diagonal matrix of Spearman correlation coefficient to get the improved PCA initial feature variable matrix . Step 5: Calculate the covariance matrix of the eigenvariable matrix . Step 6: Calculate the eigenvalues and eigenvectors of the covariance matrix.. Step 7: Calculate the principal component contribution rate of each characteristic variable, and obtain the final principal component as the prediction index according to the cumulative contribution rate greater than or equal to 95%. 2.3. Comparison of data dimensionality reduction before and after improvement 13 principal components were obtained by principal component analysis (PCA) and improved principal component analysis (IPCA) after dimensionality reduction. The results are shown in Table 2 and Figure 1 Table 2. Changes in the contribution rate of principal component analysis before and after improvement Principal component PCA IPCA Contribution contribution (%) Cumulative contribution rate (%) Contribution contribution (%) Cumulative contribution rate (%) Z1 61.2484 61.2484 89.6354 89.6354 Z2 9.7231 70.9714 3.1107 92.7461 Z3 8.6822 79.6536 2.4600 95.2060 Z4 8.4483 88.1019 1.6003 96.8063 Z5 4.5867 92.6886 0.8557 97.6620 Z6 4.3779 97.0665 0.7171 98.3791 Z7 1.5388 98.6053 0.5624 98.9415 Z8 0.8199 99.4252 0.4102 99.3517 Z9 0.2569 99.6821 0.2513 99.6030 Z10 0.1154 99.7975 0.1499 99.7529 Z11 0.0939 99.8914 0.1316 99.8845 Z12 0.0779 99.9694 0.0683 99.9528 Z13 0.0306 100.0000 0.0472 100.0000 (a) (b) Figure 1. Effect diagram of principal component analysis before and after improvement Compared with the effect before and after the improvement of principal component analysis, principal components were selected according to the cumulative principal component was greater than or equal to 95%. Principal component analysis extracted 6 principal components with a cumulative contribution rate of 97.0665%, while the improved principal component analysis only needed to extract 3 principal components with a cumulative contribution rate of 95.2060%. Although the cumulative contribution rate of PCA is greater than that of IPCA, the first principal component of IPCA carries more information. While the cumulative contribution rate is greater than or equal to 95%, the number of principal components is reduced, the calculation accuracy of the model is improved, and the calculation amount of the model is effectively reduced and the calculation speed of the model is increased. The improved principal component analysis method has better effect on dimensionality reduction. IPCA method is superior to PCA method, and can be used as the 1 2 3 4 5 6 7 8 9 10 11 12 13 0 20 40 60 80 100 C on tr ib ut io n ra te o f im pa ct f ac to r (% ) Principal component PCA Cumulative contribution rate 1 2 3 4 5 6 7 8 9 10 11 12 13 0 20 40 60 80 100 C on tr ib ut io n ra te o f im pa ct f ac to r (% ) Principal component IPCA Cumulative contribution rate 200 final method to reduce the dimension of gas emission prediction index. 3. BP Neural Network Model Based on CP-SO Optimization 3.1. PSO algorithm improvement Particle Swarm Optimization (PSO) is a stochastic global optimization algorithm first proposed by Kennedy and Eberhart in 1995[18]. 3.1.1. Basic PSO algorithm The basic PSO algorithm uses formulas (4) (5) to operate on particles: (4) (5) Where: called inertia factor, is non-negative; and are called learning factor and is usually 2; and are random number between 0 and 1;𝑥 is the position of the i-th particle at step d. The velocity evolution equation expressed in formula (4) and (5) consists of three parts. The first part is the inertia of the particle itself, and the velocity in a certain direction when the particle carries out the next search; The second part is the learning of the particle itself, which comes from the "experience" of the particle's previous flight, namely the local optimal solution. The third part is the information sharing among particles, that is, the global optimal solution. In addition, when the particle is constantly adjusting its position according to the speed, it is also limited by the maximum speed 𝑣 . The termination condition of iteration is generally selected as the maximum number of iterations or the optimal location of the current location of the population meets the predetermined minimum fitness threshold. 3.1.2. Adaptive inertial weight particle swarm optimization A large inertial weight value is conducive to global search, while a small weight value is more conducive to local search.[19, 20] In order to make PSO more stable, the inertia weight of PSO is changed adaptively. Compared with the original PSO, the inertia weight and the number of iterations are related to the fitness of each particle. For the minimum value problem, the change rule of inertia weight is shown in equation (6). For the maximum value problem, the change rule of inertia weight is shown in equation (7). (6) (7) Where: 𝑤 and 𝑤 are the minimum and maximum inertia coefficients given in advance, generally 0.4 and 0.9. The average fitness of all particles at the DTH iteration is shown in equation (8): (8) Minimum fitness of all particles at the DTH iteration, as shown in equation (9) : (9) 3.1.3. Compression factor particle swarm optimization Individual learning factors and social learning factors determine the influence of the experiential information of particles and other particles on the trajectory of particles, which reflects the information exchange between particle groups [21-23].Setting a large 𝑐 value will make the particles search too much in their own range and converge slowly, and a large 𝑐 value will cause the particles to converge prematurely to the local optimal value. Setting a compression factor k, such as equations (10) and (11) can delay the arrival of the local optimal solution. (10) (11) Where: , . 3.1.4. Combinatorial particle swarm optimization (CPSO) n the early stage, the fitness of particle swarm is large, and global search is more necessary. Corresponding to each particle, it requires rapid search within its own scope, and requires large inertia weight , individual learning factor and small social learning factor . In the later stage of particle swarm search, fitness decreases gradually, so smaller inertia weights, individual learning factors and larger social learning factors are needed. However, in the later stage, particle swarm will hover around itself too much and easily fall into the local optimal solution. CPSO (Combine adaptive inertia weights and learning factors PSO) algorithm combines "adaptive inertia weights" and "compression factor algorithm". It can balance the local search and global search capabilities of PSO algorithm to avoid the shortcomings of premature convergence of PSO algorithm and falling into local optimal solutions, and the speed update formula is shown in (12). (12) 3.2. BP neural network BP neural network is a kind of multi-layer feedforward neural network which corrects error by error backpropagation algorithm. The signal propagates forward while the error 201 propagates back. In the forward propagation process, the input signal is processed layer by layer through the input layer and hidden layer. When reaching the output layer, if the result or the number of iterations does not meet the requirements, the backpropagation process is entered, the error signal is returned in the original way, and the weight of each layer is modified. The number of hidden layer neurons can be calculated by empirical formula (13) [24]: (13) Where: a is the number of hidden layer nodes; h is the number of nodes in the output layer; c is the number of nodes in the input layer; d is the adjustment constant, generally 1,2,... 10. 3.3. Calculation steps of CPSO-BP model Step 1: Data normalization, establish BP neural network structure; Step 2: Set the particle swarm hyperparameters; Step 3: Conduct training according to mean square error as fitness function of CP-SO objective function (MSE, as shown in equation (14)), iteratively optimize the weights and thresholds of BP algorithm; (14) Where: is the predicted value of the training group, is the actual value of the training group. Step 4: Calculate the fitness value of the particle swarm, compare and update the sum after each calculation, and bring the speed and position of updating the particle swarm into equation (9) and equation (2) until the required accuracy or number of iterations are reached, and the final , namely the optimal weight value and threshold value are obtained; Step 5: Bring the test group data into the updated BP neural network for calculation until the number of iterations ends or the error condition is reached, and the predicted results are output. 4. IPCA-CPSO-BP Model Application 4.1. IPCA principal component analysis Through MATLAB software programming, the IPCA algorithm was used to reduce the dimensionality of the original data, and three principal component feature variables were obtained. The eigenvector data of the principal component about the impact factor are shown in Table 3: Table 3. Improved principal component factor component data Influence factor F1 F2 F3 X1 0.0015 0.0008 -0.0011 X2 0.0024 0.0017 -0.0030 X3 0.0088 0.0047 -0.0073 X4 0.0038 0.0064 -0.0088 X5 -0.9266 0.3524 0.0779 X6 0.0343 -0.1115 0.9913 X7 0.0066 0.0077 -0.0013 X8 0.0008 0.0009 -0.0001 X9 0.0046 -0.0068 -0.0079 X10 0.1036 -0.0358 0.0520 X11 -0.3597 -0.9284 -0.0916 X12 0.0035 0.0084 -0.0071 X13 0.0024 0.0015 0.0004 The three principal components obtained after dimensionality reduction by improved principal component analysis become new prediction indexes for gas emission prediction. As can be seen in Table 3.3 of the data obtained by MATLAB analysis, the data information in the first principal component mainly includes the degree of advancement of mining coal seam, coal thickness, recovery rate, mining height, recovery amount and gas content of mining coal seam in the original prediction index. The second principal component includes the underground pressure, the spacing of adjacent coal seams, the inclination of coal seams and the gas content of adjacent layers. The third principal component includes burial depth, temperature and air volume. 4.2. Parameter setting The five algorithms BP, PSO-BP, PCA-PSO-BP, IPCA- PSO-BP and IPCA-CPSO-BP were programmed and calculated with MATLAB software. BP neural network was constructed with Newff function, with the maximum training times of 1000 and the learning rate of 0.01. Training requires an accuracy of 0.000001. Set groups 1-25 as the training group, 26-30 as the test group, the input, hidden and output layers of the first two groups as 13, 11 and 1, the third group as 8, 9 and 1, and the last two groups as 4, 6 and 1. 4.3. Prediction result analysis The predicted values and relative errors (hereinafter referred to as errors) of the model are shown in Table 4: 202 Table 4. Test data results group Actual value BP PSO-BP PCA-PSO-BP IPCA-PSO-BP IPCA-CPSO-BP forecast error forecast error forecast error forecast error forecast error 26 16.5 17.0189 0.5189 16.7819 0.2819 16.6531 0.1531 16.4495 0.0505 16.6748 0.1748 27 15.0 14.8042 0.1958 15.4738 0.4738 15.2683 0.2683 15.0076 0.0076 14.9129 0.0871 28 15.0 15.0634 0.0634 15.0536 0.0536 15.3371 0.3371 14.7403 0.2597 14.9802 0.0198 29 16.5 16.6096 0.1096 16.7949 0.2949 16.3260 0.1740 16.1537 0.3463 16.3371 0.1629 30 15.0 14.5828 0.4172 14.9649 0.0351 15.2247 0.2247 15.2073 0.2073 14.9827 0.0173 Average relative error MRE, normalized root-mean-square error Nmse and regression correlation index R2 were introduced to evaluate the prediction results, as shown in Table 5: Table 5. Evaluation indeZes of prediction results of each model Evaluation index MRE/% Nmse R2 BP 1.6638 0.3155 0.8157 PSO-BP 1.4491 0.2811 0.8537 PCA-PSO-BP 1.5034 0.2408 0.8926 IPCA-PSO-BP 1.1138 0.2159 0.9137 IPCA-CPSO-BP 0.5749 0.1143 0.9758 The evaluation indicators of different models were visualized, as shown in FIG. 2 and FIG. 3 are the comparison graphs between the prediction results of different models and the real values. With the introduction of different algorithms and improved algorithms in a certain order, the average relative error MRE increases first and then decreases, and the normalized root- mean-square error Nmse and the regression correlation index R2 change monotonically. The smaller the mean relative error MRE and the normalized root-mean-square error Nmse, the more accurate the prediction result. The optimization effect of different algorithms on BP neural network is shown in Figure 2. The IPCA-CPSO-BP model is the most accurate in prediction, with MRE value of 0.5749 and Nmse value of 0.1143. Compared with BP prediction model, the average relative error of IPCA-CPSO-BP model is reduced by 1.0889%, and the average relative error of BP prediction model is about 3 times that of IPCA-CPSO-BP model. The normalized root-mean-square error is reduced by 0.2012, and the normalized root-mean-square error of BP prediction model is about 2.8 times that of IPCA-CPSO-BP model, which significantly improves the prediction accuracy of IPCA-CPSO-BP model. The closer the regression correlation index R2 is to 1, the better the fitting effect of the model is. As shown in Figure 2 and 3, IPCA-CPSO-BP model has the best fitting effect, and the R2 value is 0.9758. After particle swarm optimization and improved particle swarm optimization are added to the model, the fitting effect is improved and the dispersion of data distribution is low. The lines are more consistent with the true value. Principal component analysis (PCA) can more effectively remove redundant data information and increase the accuracy of prediction results in the whole model. The improved PCA strengthens this feature. Particle swarm optimization (PSO) mainly controls the stability of the model, and the improved PSO also strengthens this property. Figure 2. Evaluation indicators of different models BP PSO-BP PCA-PSO-BP IPCA-PSO-BP IPCA-CPSO-BP 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 1.66377 1.4491 1.50343 1.1138 0.5749 0.81565 0.85371 0.89263 0.91372 0.97579 0.31551 0.28106 0.24079 0.21585 0.11433 E va lu at io n in de x va lu e Different prediction models MRE（%） R2 N mse 203 Figure 3. Comparison between the predicted value and the real value of different models Figure 4 shows the curve of PSO iterations of each model. The model iteration curve is in the shape of a ladder. The model with the fastest convergence speed is IPCA-PSO-BP model, which tends to converge at the 25th step after iteration. The convergence speed of PSO-BP model and IPCA-CPSO- BP model is relatively slow, and the hopping amplitude of fitness value convergence is relatively large. Because PSO- BP model has not undergone data dimensionality reduction, the convergence process is random, and the number of iterations increases correspondingly. IPCA-CPSO-BP model adds adaptive inertia weight and shrinkage factor. The model can have global search particles before and after the search, so it is not easy to fall into the local optimal solution, the convergence rate decreases, and the prediction accuracy increases. Figure 4. Iteration curves of different models BP model relies too much on initial parameter values in the process of use, PSO convergence speed in the process of parameter optimization leads to easy to fall into the local optimal solution, PCA is not strong enough to reduce the dimension of data, resulting in the model has randomness in the prediction process. IPCA-CPSO-BP model improves the pre-processing ability of data, slows down the convergence speed and improves the prediction accuracy. 5. Conclusion (1) The Spearman correlation coefficient (absolute value) was added to the principal component analysis as the weight of each influencing factor, which improved the principal component analysis. After dimensionality reduction, the traditional principal component analysis produced 6 principal components, with a cumulative contribution rate of 97.0665%, while the improved principal component analysis had only 3 principal components, with a cumulative contribution rate of 95.2060%. The improved principal component has better dimensionality reduction under the premise of preserving 95% information. (2) The traditional particle swarm optimization (PSO) algorithm is improved by adding adaptive inertia weights and factors of compressed search space.The optimized algorithm not only guarantees the search accuracy, but also shows the faster convergence speed, thus achieving double improvement in the optimization efficiency and the accuracy of the predicted results. (3) The IPCA-CPSO-BP neural network prediction model realized the prediction of gas emission at 3309 working face of Zhongtai Mining, and the average absolute error (MAE) was 0.0924. The mean relative error (MAPE) was 0.5749%. The regression correlation index R2 was 0.9758. (4) Four models, namely BP model, PSO-BP model, PCA- PSO-BP model and IPCA-PSO-BP model, were established respectively, and the amount of gas emission was predicted, and the predicted results were compared with IPCA-CPSO- BP algorithm. Through the result analysis, it is found that the principal component analysis method controls the stability of the model operation, the particle swarm optimization algorithm controls the prediction accuracy of the model, and the calculation accuracy of IPCA-CPSO-BP model is about 3 times that of BP neural network model. The data confirm the feasibility of the IPCA-CPSO-BP neural network prediction model studied in this paper in the prediction of gas emission quantity, and also has a certain reliability in the prediction effect. References [1] QIN Y J, SU W W, JIANG W Z, et al. Research Progress and Development Direction of Mine Gas Emission Forecast Technology in China[J]. Safety in Coal Mines. 2020, 51(10): 52-59. 26 27 28 29 30 14.5 15.0 15.5 16.0 16.5 17.0 G as e m is si on ( m 3 /m in ) group Actual value IPCA-CPSO-BP BP PSO-BP PCA-PSO-BP IPCA-PSO-BP 0 50 100 150 200 0.02 0.04 0.06 0.08 0.10 F itn es s va lu e Number of iterations PSO-BP PCA-PSO-BP IPCA-PSO-BP IPCA-CPSO-BP 204 [2] LI R Q, SHI S L, NIAN Q F, et al. Research on Coalmine Gas Accident Rules in China in Recent Decade[J]. China Safety Science Journal. 2011, 21(09): 143-151. [3] WU S Y, ZHANG J L, HAN T T, et al. Revision suggestions on “Prediction Methods of Mine Gas Emission Rate” industry standard[J]. China Coal. 2012, 38(09): 117-120. [4] XIE M J, MA S Q. Prediction of coal mine gas emission based on multiple linear regression theory[J]. Shaanxi Coal. 2021, 40(01): 26-29. [5] SUN X Y. Research and application of neural network gas emission prediction system[J]. Coal Mine Modernization. 2023, 32(05): 65-68. [6] LU G B, KANG J K, BAI G, et al. Application of PCA-BP to gas emission prediction of mining working face[J]. Journal of Liaoning Technical University (Natural Science). 2015, 34(12): 1329-1334. [7] ZHOU X H, SUN J Z. Prediction of Gas Emission based on Principal Factor Analysis and Improved BP Neural Network[J]. Mining Safety & Environmental Protection. 2018, 45(06): 43- 47. [8] XIAO P, XIE X J, SHUANG H Q, et al. Prediction of gas emission quantity based on KPCA-CMGANN algorithm[J]. China Safety Science Journal. 2020, 30(05): 39-47. [9] LIN H F, LIU S H, ZHOU J, et al. Prediction method and application of gas emission from mining workface based on STL-EEMD-GA-SVR[J]. Coal Geology & Exploration. 2022, 50(12): 131-141. [10] FENG S C, SHAO L S, LU W J, et al. Application of PCA- PSO-LSSVM model in gas emission prediction[J]. Journal of Liaoning Technical University (Natural Science). 2019, 38(02): 124-129. [11] XIAO P, XIE X J, SHUANG H Q, et al. Application of wavelet-extreme learning machine in time-varying series prediction of gas emission quantity[J]. JOURNAL OF XI' AN UNIVERSITY OF SCIENCE ANDTECHNOLOGY. 2020, 40(05): 839-845. [12] LIU F. Prediction of gas emission based on PCA-SAPSO-BP neural network[J]. Safety in Coal Mines. 2023, 54(04): 60-68. [13] FU H, FU Y, ZHAO J C, et al. Prediction of gas emission based on KPCA-ARIMA algorithm[J]. Journal of Liaoning Technical University (Natural Science). 2022, 41(05): 406-412. [14] ZHANG Y C, WANG Y, GUO K Y. Research on prediction of gas emission in working face based on WOA-LSTM[J]. Mining Safety & Environmental Protection. 2023: 1-6. [15] ZHOU X, YANG J P, ZU Y W, et al. Gas emission prediction based on NMF-HGS-RF[J]. Mining Safety & Environmental Protection. 2023, 50(03): 117-123. [16] LIN Y B, QIN Y, WANG X, et al. Geology and emission of mine gas in Bintang mining area with low rank coal and high mine gas[J]. Journal of China Coal Society. 2019, 44(07): 2151-2158. [17] GUO C, XIA Y C, SUN X Y, et al. Method and practice of gas geological grading evaluation on coal mining face of high gas m[J]. Journal of China Coal Society. 2019, 44(08): 2409-2418. [18] EBERHART, KENNEDY. A new optimizer using particle swarm theory[C]. Nagoya, Japan: IEEE, 1995. [19] KANG Y S, ZANG S L. Improved Particle Swarm Optimization Algorithm Based on Multiple Strategies[J]. Journal of Northeastern University (Natural Science). 2023, 44(08): 1089-1097. [20] ZHANG H, WANG X L. Adaptive inertia weight particle swarm optimization algorithm[J]. Intelligent Computer and Applications. 2023, 13(09): 5-8. [21] EBERHART, Y. Shi. Comparing Inertia Weights and Constriction Factors in Particle Swarm Optimization[C]. La Jolla, CA, USA: CEC00, 2000. [22] VANDENBERGH, ENGELBRECHT. A Cooperative Approach to Particle Swarm Optimization[J]. IEEE Transactions on Evolutionary Computation. 2004, 8(3): 225- 239. [23] LIU Y D, ZHOU M R, XIE J Y, et al. RESEARCH ON SIMULATION MODEL FOR DATA ASSIMILATION IN SOLAR RADIATION PREDICTION BASED ON PARTICLE SWARM OPTIMIZER WITH TIME VARYING CONSTRICT FACTOR[J]. Acta Energeia Solaris Sonica. 2021, 42(04): 181-185. [24] WANG R B, XU H Y, LI B, et al. Research on Method of Determining Hidden Layer Nodes in BP Neural Network[J]. Computer Technology and Development. 2018, 28(04): 31-35.