Method for predicting gas emission quantity of coal face

Abstract

The invention discloses a method for predicting the gas emission quantity of a coal face, which comprises the following steps of: performing primary decomposition on gas emission quantity data by adopting a completely adaptive noise set empirical mode decomposition method to obtain a plurality of intrinsic mode components; calculating the complexity of each intrinsic mode component by using fuzzy entropy, and setting a complexity threshold value; performing parameter optimization on the variational mode decomposition method by adopting a fishing optimization algorithm, and performing secondary decomposition on the intrinsic mode component of which the complexity is greater than a complexity threshold by adopting the variational mode decomposition method to obtain the intrinsic mode component; after a flood optimization algorithm is adopted to carry out parameter optimization on the BiGRU prediction model, the BiGRU prediction model is established, and the prediction model is adopted to carry out training and prediction on all intrinsic mode components; and summing the prediction results of the intrinsic mode components to obtain a final prediction result. The method is high in prediction precision and good in prediction effect, and can be effectively applied to coal face gas emission quantity prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of gas disaster prevention and control, and particularly relates to a method for predicting gas emission volume in a coal mining face. Background Technique

[0002] Gas disaster accidents have always threatened the safe production of coal mines. With the continuous mining of mines, the incidence rate of gas overrun accidents has gradually increased, seriously affecting the safe and efficient production of coal mines in China. Therefore, researching and establishing a prediction model for gas emission volume in mines, improving the prediction accuracy of gas emission volume, and reducing the occurrence of disaster accidents are of great significance for the safe production of coal mines and facilitating the construction of a modern industrial system.

[0003] Traditional methods for predicting gas emission volume, such as statistical analysis method and coal seam gas content method, etc., have simple principles, but the prediction results are prone to randomness and it is difficult to make relatively accurate predictions. In recent years, with the progress of technologies such as intelligent information analysis and big data, machine learning methods have been applied by more and more scholars to predict gas emission volume in coal mines. These methods are generally divided into two categories: those based on time series data analysis and those considering multiple factor indicators. Since the conditions of coal mines are different from each other, the importance of each influencing factor is also different when predicting gas emission volume. In addition, gas emission volume is affected by various complex factors, and most coal mines can only provide historical records of gas emission volume and lack detailed data on other influencing factors. These factors together lead to a deviation between the final model and the actual gas emission state. Therefore, many scholars have begun to focus on exploring the prediction of gas emission volume based on time series data analysis.

[0004] Yu Yongqiang et al. proposed to use the grey system theory to predict the gas emission and its change trend in the mine in the paper "Predicting the Gas Emission of the Wu 8 Coal Seam in Pingdingshan No. 1 Mine with Grey System Theory"; Tao Yunqi et al. combined the grey model with the Markov model to construct a prediction model for the gas emission in coal mines in the paper "Predicting the Gas Emission in Coal Mining Faces with an Improved Grey Markov Model"; Shi Shiliang et al. constructed an EMD-PSO-SVM prediction model in the paper "Prediction Method and Application of Coal Mine Gas Emission Based on EMD-PSO-SVM", and used the support vector machine optimized by the particle swarm algorithm for prediction, which is applicable to unstable time series; Dai Wei et al. established a VMD-DE-RVM gas emission interval prediction model in the paper "Interval Prediction Method for Gas Emission in Coal Mining Faces Based on VMD-DE-RVM", and proved that the model has a high prediction accuracy; Zhang Yucai et al. screened input characteristic factors by the Pearson correlation coefficient method in the paper "Research on the Prediction of Gas Emission in Coal Mining Faces Based on WOA-LSTM", constructed a gas emission prediction model based on the whale optimization algorithm (WOA) and long short-term memory network (LSTM), and studied the prediction accuracy of the model under different time steps; Rong Tongrui et al. decomposed the gas emission data twice in the paper "Research on the Prediction Method of Gas Emission in Coal Mining Faces Based on the Combined Model of Quadratic Decomposition and BO-BiLSTM", and introduced the Bayesian algorithm to optimize the bidirectional long short-term memory network, giving full play to the advantages of each algorithm in processing time series, and can accurately predict the gas emission in coal mining faces.

[0005] Compared with the traditional prediction methods, the above models have higher prediction accuracy, but there are still deficiencies: on the one hand, most of the original time series are non-linear and non-stationary, and it is often difficult to capture these complex fluctuations by direct prediction; on the other hand, even if the original data is decomposed first and then predicted, there is still a problem of insufficient model generalization ability. This is because the decomposition algorithm and the machine learning model are sensitive to the setting of parameters, and certain parameter tuning is required to achieve a certain effect. If the setting is improper, the decomposed components will contain components with multiple frequencies, which leads to the need for the model to identify various change information during prediction and cannot focus on learning the respective patterns and trends of each component. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a prediction method for the gas emission in coal mining faces with high prediction accuracy and good prediction effect, which can be effectively applied to the prediction of the gas emission in coal mining faces in view of the above deficiencies in the prior art.

[0007] To solve the above technical problem, the technical solution adopted by the present invention is: a prediction method for the gas emission in coal mining faces, the method includes the following steps:

[0008] Step S1: Use the complete adaptive noise ensemble empirical mode decomposition method to perform a primary decomposition on the gas emission data, and decompose it into multiple intrinsic mode components;

[0009] Step S2: Calculate the complexity of each intrinsic mode component using fuzzy entropy and set a complexity threshold;

[0010] Step S3: After optimizing the parameters of the variational mode decomposition method using the fishing optimization algorithm, use the variational mode decomposition method to perform a secondary decomposition on the intrinsic mode components with a complexity greater than the complexity threshold, and obtain intrinsic mode components with small fluctuations and regularity;

[0011] Step S4: After optimizing the parameters of the BiGRU prediction model using the flood optimization algorithm, establish a BiGRU prediction model, and use the BiGRU prediction model to train and predict all the intrinsic mode components obtained in Step S1 and Step S3;

[0012] Step S5: Sum up the prediction results of each intrinsic mode component to obtain the final gas emission prediction result.

[0013] For the above-mentioned gas emission prediction method for a coal mining face, after Step S5, it further includes Step S6: Evaluate the prediction result to prove the accuracy of the prediction result.

[0014] For the above-mentioned gas emission prediction method for a coal mining face, the specific process of using the complete adaptive noise ensemble empirical mode decomposition method to perform a primary decomposition on the gas emission data in Step S1 is as follows:

[0015] Step S101: Define the original time series signal of the gas emission data as f(t), and add an adaptively decomposable Gaussian white noise sequence a n (t), which is expressed by the formula:

[0016] F n (t) = f(t) + ε0a n (t)(F1)

[0017] Among them, F n (t) is the signal after adding white noise, ε0 is the standard deviation of the originally added white noise, n is the number of decompositions and the value of n is a natural number from 1 to M, and M is the total number of decompositions;

[0018] Step S102: Decompose the signal F n (t) using the empirical mode decomposition method, take the average value of a group of intrinsic mode components obtained by the decomposition, and obtain the first intrinsic mode component IMF1(t), which is expressed by the formula:

[0019]

[0020] Among them, represents the first intrinsic mode component obtained by decomposition when the decomposition times is n;

[0021] Step S103: Calculate the residual r1(t) obtained from the first decomposition, which is expressed by the formula:

[0022] r1(t) = f(t) - IMF1(t)(F3)

[0023] Step S104: Add the Gaussian white noise sequence a n (t) to r1(t), perform M decompositions until it is decomposed into a first-order component, and calculate the second intrinsic mode component IMF2(t), which is expressed by the formula:

[0024]

[0025] Among them, E1 is the first-order component generated by the empirical mode decomposition method, and ε1 is the standard deviation of the white noise added to the first-order component;

[0026] Step S105: Calculate the residual r2(t) obtained from the second decomposition, which is expressed by the formula:

[0027] r2(t) = r1(t) - IMF2(t)(F5)

[0028] Step S106: Continue to perform empirical mode decomposition until the decomposition condition is not satisfied. At this time, the original signal f(t) is decomposed into q intrinsic mode components and a residual component, which is expressed by the formula:

[0029]

[0030] Among them, q is the total number of intrinsic mode components, and IMF i (t) represents the i-th intrinsic mode component, which is expressed by the formula:

[0031]

[0032] E i is the i-th order component generated by the empirical mode decomposition method, and ε i is the standard deviation of the white noise added to the i-th order component; r i (t) is the residual obtained from the i-th decomposition and is expressed by the formula:

[0033] r i (t) = r i-1 (t) - IMF i (t)(F8)

[0034] r i-1 (t) is the residual obtained from the (i - 1)-th decomposition;

[0035] R(t) is the residual component and is expressed by the formula:

[0036] R(t) = r q (t) = r q-1 (t) - IMF q (t)(F9)

[0037] For the above method for predicting the gas emission amount in a coal mining face, in step S3, the use of the fish swarm optimization algorithm to optimize the parameters of the variational mode decomposition method is to use the fish swarm optimization algorithm to optimize the two parameters of the penalty factor α and the decomposition layer number k in the variational mode decomposition method.

[0038] For the above method for predicting the gas emission amount in a coal mining face, the specific process of using the fish swarm optimization algorithm to optimize the two parameters of the penalty factor α and the decomposition layer number k in the variational mode decomposition method in step S3 is as follows:

[0039] Step A1: Initialize the parameters of the fish swarm optimization algorithm, and set the search range of the decomposition layer number and the penalty factor of the variational mode decomposition method;

[0040] Step A2: Set the number of CFOA populations, the maximum number of iterations, and the search dimension;

[0041] Step A3: Select the minimum envelope entropy as the fitness function; calculate the fitness value corresponding to each parameter combination respectively, and compare it with the fitness value of the previous iteration;

[0042] Step A4: If it is less than the fitness value of the previous iteration, update the decomposition layer number and the penalty factor; if it is greater than the fitness value of the previous iteration, do not update the decomposition layer number and the penalty factor;

[0043] Step A5: Continue the iteration until the maximum number of iterations is reached, and output the optimal values of the decomposition layer number and the penalty factor;

[0044] Step A6: Substitute the optimal values into VMD to complete the decomposition.

[0045] For the above method for predicting the gas emission amount in a coal mining face, in step S2, the use of the variational mode decomposition method to perform secondary decomposition on the gas emission amount data is to continuously update the center frequency and bandwidth of each mode by using the alternating direction multiplier method, and finally obtain each mode and the corresponding center frequency, which is expressed by the formula:

[0046]

[0047] Among them, represents the Wiener filtering of the current residue, ω k represents the center frequency of the current mode component, denotes the Fourier transform of f(t), where f(t) represents the original signal processed in step S1; denotes u i (t)'s Fourier transform, where u i (t) represents the mode function of the i-th decomposition in the variational mode decomposition method; denotes the Fourier transform of λ(t), where λ(t) represents the Lagrange multiplier; k is the number of mode functions, n is the number of iterations, α is the penalty factor, ω is the frequency variable, representing any frequency point within the frequency domain range, ω k is the oscillation frequency representing the mode function; denotes u k (t)'s Fourier transform, where u k (t) represents the mode function of the k-th decomposition in the variational mode decomposition method;

[0048] For the above-mentioned method for predicting the gas emission volume in a coal mining face, in step S4, the use of the flood optimization algorithm to optimize the parameters of the BiGRU prediction model is to optimize the initial learning rate and the number of hidden layer nodes of the BiGRU prediction model.

[0049] For the above-mentioned method for predicting the gas emission volume in a coal mining face, the output of the BiGRU prediction model in step S4 is determined by the weighted sum of the results of a forward hidden layer and a backward hidden layer. The calculation formula is as follows:

[0050]

[0051]

[0052] where GRU(·) is the gated recurrent unit, x t is the input data, is the output of the previous moment in the forward direction, is the output of the previous moment in the backward direction; h t is the output, is the output of the forward hidden layer, is the output of the backward hidden layer; α t is the weight of the forward hidden layer, β t is the weight of the backward hidden layer; b t is the bias.

[0053] For the above-mentioned method for predicting the gas emission volume in a coal mining face, when evaluating the prediction result in step S6 to prove the accuracy of the prediction result, the mean absolute error E MAE , mean absolute percentage error E MAPE , root mean square error E RMSE and coefficient of determination R 2, and the specific calculation formulas are as follows:

[0054]

[0055] Among them, α t is the original sequence; is the predicted sequence; t = 1, 2, ···, W is the sequence length.

[0056] The present invention has the following advantages compared with the prior art:

[0057] 1. The present invention proposes a prediction model for gas emission volume in coal mining face based on quadratic decomposition and bidirectional gated recurrent unit. This method first performs a primary decomposition on the gas emission volume data using CEEMADN, then performs a secondary decomposition on the high-complexity components after decomposition using VMD, constructs BiGRU prediction models for each decomposed subsequence respectively, and introduces the fish swarm optimization algorithm (CFOA) and the firefly algorithm (FLA) to optimize the parameters of each model; this combined model fully excavates the non-linear characteristic information of the original data and gives full play to the respective advantages of each algorithm in processing time series.

[0058] 2. The prediction accuracy of the present invention is high, the prediction effect is good, and it can be effectively applied to the prediction of gas emission volume in coal mining face.

[0059] The technical solution of the present invention will be further described in detail below through the drawings and embodiments. Description of the Drawings

[0060] Figure 1 is the flow chart of the method of the present invention;

[0061] Figure 2 is the original gas emission volume data diagram of the present invention;

[0062] Figure 3 is the box plot for outlier discrimination of the present invention;

[0063] Figure 4 is the complexity diagram of each component after the primary decomposition of the gas emission volume data of the present invention;

[0064] Figure 5 is the fitness curve diagram of the optimized VMD of the present invention;

[0065] Figure 6 is the diagram of each component of the secondary decomposition of the IMF1 of the primary decomposition of the gas emission volume data of the present invention;

[0066] Figure 7 is the complexity diagram of each component of the secondary decomposition of the IMF1 of the primary decomposition of the gas emission volume data of the present invention;

[0067] Figure 8It is a comparison graph of the superposition of each component of the present invention and the original gas emission volume;

[0068] Figure 9A It is a convergence curve graph of the first optimization function of the present invention;

[0069] Figure 9B It is a convergence curve graph of the second optimization function of the present invention;

[0070] Figure 10 It is a prediction result graph of each component model of the present invention;

[0071] Figure 11 It is a prediction result graph of the component superposition model of the present invention;

[0072] Figure 12 It is a prediction result graph of M1 and M2 models of the present invention;

[0073] Figure 13 It is a prediction deviation graph of M1 and M2 models of the present invention;

[0074] Figure 14 It is a prediction result graph of M2, M6 and M7 models of the present invention;

[0075] Figure 15 It is a prediction deviation graph of M2, M6 and M7 models of the present invention;

[0076] Figure 16 It is a prediction result graph of M6, M7 and M10 models of the present invention;

[0077] Figure 17 It is a prediction deviation graph of M6, M7 and M10 models of the present invention;

[0078] Figure 18 It is a prediction result graph of M10 and CVB models of the present invention;

[0079] Figure 19 It is a prediction deviation graph of M10 and CVB models of the present invention. Detailed implementation manners

[0080] As Figure 1 shown, the gas emission volume prediction method for the coal mining face of the present invention includes the following steps:

[0081] Step S1: Use the complete adaptive noise ensemble empirical mode decomposition method (CEEMDAN method) to decompose the gas emission volume data once to decompose multiple intrinsic mode components;

[0082] In specific implementation, the construction of the BiGRU prediction model requires ensuring that the original data is complete. To ensure the accuracy of the BiGRU prediction model, before model construction, it is necessary to screen for outliers in the acquired data, that is, to determine whether there are situations such as missing data, errors, or abnormal change trends; before step S1, it is also necessary to detect whether there are outliers in the gas emission data in the origin software. When there are outliers, mean imputation is performed, and then step S1 is carried out; when there are no outliers, step S1 is directly carried out;

[0083] In this embodiment, the specific process of using the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN method) to decompose the gas emission data once in step S1 is as follows:

[0084] Step S101: Define the original time series signal of the gas emission data as f(t), and add an adaptive decomposable Gaussian white noise sequence a n (t), which is expressed by the formula:

[0085] F n (t) = f(t) + ε0a n (t)(F1)

[0086] Among them, F n (t) is the signal after adding white noise, ε0 is the standard deviation of the originally added white noise, n is the number of decomposition times and the value of n is a natural number from 1 to M, and M is the total number of decomposition times;

[0087] Step S102: Decompose the signal F n (t) using the empirical mode decomposition (EMD method), and take the average of a group of intrinsic mode functions obtained by the decomposition to obtain the first intrinsic mode function IMF1(t), which is expressed by the formula:

[0088]

[0089] Among them, IMF1 n (t) represents the first intrinsic mode function obtained by decomposition when the number of decomposition times is n;

[0090] Step S103: Calculate the residual r1(t) obtained by the first decomposition, which is expressed by the formula:

[0091] r1(t) = f(t) - IMF1(t)(F3)

[0092] Step S104: The Gaussian white noise sequence a n(t) is added to r1(t), and decomposed M times until it is decomposed into first-order components, and the second intrinsic mode function IMF2(t) is calculated, which is expressed by the formula:

[0093]

[0094] Among them, E1 is the first-order component generated by the empirical mode decomposition method (EMD method), and ε1 is the standard deviation of the white noise added to the first-order component;

[0095] Step S105, calculate the residual r2(t) obtained from the second decomposition, which is expressed by the formula:

[0096] r2(t) = r1(t) - IMF2(t) (F5)

[0097] Step S106, continue the empirical mode decomposition until the decomposition condition is not satisfied. At this time, the original signal f(t) is decomposed into q intrinsic mode functions and a residual component, which is expressed by the formula:

[0098]

[0099] Among them, q is the total number of intrinsic mode functions, and IMF i (t) represents the i-th intrinsic mode function, which is expressed by the formula:

[0100]

[0101] E i is the i-th order component generated by the empirical mode decomposition method (EMD method), and ε i is the standard deviation of the white noise added to the i-th order component; r i (t) is the residual obtained from the i-th decomposition and is expressed by the formula:

[0102] r i (t) = r i-1 (t) - IMF i (t) (F8)

[0103] r i-1 (t) is the residual obtained from the (i - 1)-th decomposition;

[0104] R(t) is the residual component and is expressed by the formula:

[0105] R(t) = r q (t) = r q-1 (t) - IMF q (t) (F9)

[0106] The gas emission is affected by various factors, including trend factors such as mining depth and coal seam gas content; periodic factors such as daily output and daily advance; and mutation factors such as geological structure and atmospheric pressure. Directly using the original data of gas emission will result in insufficient utilization of the data. Therefore, the CEEMDAN algorithm, which can adaptively decompose various modes in the signal, is used to decompose the time series data of gas emission once.

[0107] The CEEMDAN method adopted in the present invention (Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise) is an improvement based on the EMD method and the EEMD method (Ensemble Empirical Mode Decomposition). The intrinsic mode function (IMF) obtained after EMD decomposition is added to the original sequence, and at the same time, the average value of the IMF obtained by each decomposition is calculated, effectively solving the problems of mode mixing in EMD and residual noise in the reconstruction of EEMD.

[0108] Step S2: Calculate the complexity of each intrinsic mode function using fuzzy entropy and set the complexity threshold.

[0109] In this embodiment, the fishing optimization algorithm (CFOA algorithm) is used to optimize the parameters of the variational mode decomposition method (VMD method) in step S2, that is, the fishing optimization algorithm (CFOA algorithm) is used to optimize the two parameters of the penalty factor α and the decomposition layer number k in the variational mode decomposition method (VMD method).

[0110] In this embodiment, the specific process of using the fishing optimization algorithm (CFOA algorithm) to optimize the two parameters of the penalty factor α and the decomposition layer number k in the variational mode decomposition method (VMD method) in step S2 is as follows:

[0111] Step A1: Initialize the parameters of the fishing optimization algorithm (CFOA algorithm) and set the search range of the decomposition layer number and the penalty factor of the variational mode decomposition method (VMD method).

[0112] Step A2: Set the CFOA population number, the maximum number of iterations and the search dimension.

[0113] Step A3: Select the minimum envelope entropy as the fitness function; calculate the fitness value corresponding to each parameter combination respectively and compare it with the fitness value of the previous iteration.

[0114] Step A4: If it is less than the fitness value of the previous iteration, update the decomposition level and the penalty factor; if it is greater than the fitness value of the previous iteration, do not update the decomposition level and the penalty factor;

[0115] Step A5: Continue the iteration until the maximum number of iterations is reached, and output the optimal values of the decomposition level and the penalty factor;

[0116] Step A6: Substitute the optimal values into VMD to complete the decomposition.

[0117] In this embodiment, the secondary decomposition of the gas emission data by using the variational mode decomposition method (VMD method) in step S2 is to continuously update the center frequencies and bandwidths of each mode by using the alternating direction multiplier method, and finally obtain each mode and the corresponding center frequencies, which is expressed by the formula:

[0118]

[0119] where represents the Wiener filter of the current residual, ω k represents the center frequency of the current mode component, represents the Fourier transform of f(t), and f(t) represents the original signal processed in step S1; represents u i (t)'s Fourier transform, u i (t) represents the mode function of the i-th decomposition in the variational mode decomposition method (VMD method); represents the Fourier transform of λ(t), and λ(t) represents the Lagrange multiplier; k is the number of mode functions, n is the number of iterations, α is the penalty factor, ω is the frequency variable, representing any frequency point within the frequency domain range, ω k is the oscillation frequency representing the mode function; represents u k (t)'s Fourier transform, u k (t) represents the mode function of the k-th decomposition in the variational mode decomposition method (VMD method);

[0120] The Catch Fish Optimization Algorithm (CFOA algorithm) is a novel meta-heuristic optimization algorithm proposed by Heming Jia et al. in 2024 through the fishing strategy of fishermen in a pond, with advantages such as strong optimization ability and rapid convergence.

[0121] The variational mode decomposition method (VMD method) is a non-recursive and quasi-orthogonal signal processing method that can effectively reduce the complexity of time series. However, its effectiveness depends to a large extent on the setting of two parameters, the penalty factor α and the decomposition level k. If the value of k is too small, it may not be able to capture the important components in the signal, resulting in information loss; while if the value of k is too large, redundant information will be introduced, increasing the computational complexity; if the penalty factor α is too small, it may lead to over-flexible modes and overfitting; while if α is too large, it may lead to overly smooth modes and underfitting. Therefore, the present invention uses the fishing optimization algorithm to optimize the two parameters, the penalty factor α and the decomposition level k, in the VMD method, and then uses the variational mode decomposition method (VMD method) to perform secondary decomposition on the gas emission data, which can perform secondary decomposition on the components with higher complexity in the gas emission data after the first decomposition by CEEMDAN, helping to improve the prediction accuracy of the subsequent model.

[0122] Step S3: After optimizing the parameters of the variational mode decomposition method (VMD method) using the fishing optimization algorithm (CFOA algorithm), use the variational mode decomposition method (VMD method) to perform secondary decomposition on the intrinsic mode components with complexity greater than the complexity threshold to obtain intrinsic mode components with small fluctuations and regularity; that is, use the variational mode decomposition method (VMD method) with optimized parameters by the fishing optimization algorithm (CFOA algorithm) to perform secondary decomposition on the high-complexity components of the gas emission data after the first decomposition.

[0123] Step S4: After optimizing the parameters of the BiGRU prediction model using the flood optimization algorithm (Flood algorithm, FLA algorithm), establish a BiGRU prediction model, and use the BiGRU prediction model to train and predict all the intrinsic mode components obtained in Step S1 and Step S3.

[0124] In this embodiment, the step of optimizing the parameters of the BiGRU prediction model using the flood optimization algorithm (Flood algorithm, FLA algorithm) in Step S4 is to optimize the initial learning rate and the number of hidden layer nodes of the BiGRU prediction model.

[0125] In practical applications, appropriate initial learning rate and number of hidden layer nodes are crucial for the prediction results of the BiGRU model. Manual parameter tuning is time-consuming and it is difficult to find the best parameter combination; traditional random search method and grid search method do not fully utilize the previous search results during testing, resulting in low efficiency; the present invention uses the flood optimization algorithm to optimize the initial learning rate and the number of hidden layer nodes of the BiGRU prediction model, which can fully utilize the previous search results, find the best parameter combination, improve the efficiency, and improve the stability of training.

[0126] In this embodiment, the output of the BiGRU prediction model in step S4 is determined by the weighted sum of the results of a forward hidden layer and a backward hidden layer, and the calculation formula is as follows:

[0127]

[0128] where GRU(·) is a gated recurrent unit, x t is the input data, is the output of the previous moment in the forward direction, is the output of the previous moment in the backward direction; h t is the output, is the output of the forward hidden layer, is the output of the backward hidden layer; α t is the weight of the forward hidden layer, β t is the weight of the backward hidden layer; b t is the bias.

[0129] Traditional time series prediction models usually rely on unidirectional analysis methods, which means that they mainly predict future trends based on historical data. However, in the prediction of gas emission volume, in addition to the historical gas emission pattern having an important impact on the prediction of future emission volume, the upcoming mining plan is also crucial. Therefore, the present invention uses a bidirectional gated recurrent unit that can analyze both historical and future information of time series to predict the gas emission volume; the BiGRU prediction model is a model developed on the basis of the gated recurrent unit (GRU) model. By introducing a gating mechanism and a bidirectional structure, the modeling ability of sequence data is enhanced. Its core lies in the bidirectional structure. The BiGRU model contains two independent GRU blocks, which can obtain information from both the forward and backward directions of time series data, enabling the model to make full use of the information of the entire sequence and better capture the relationship between data; the BiGRU network structure is as Figure 2 shown.

[0130] Step S5: Sum the prediction results of each intrinsic mode component to obtain the final gas emission volume prediction result.

[0131] In this embodiment, after step S5, there is also step S6: Evaluate the prediction result to prove the accuracy of the prediction result, that is, prove the superiority of the proposed prediction model.

[0132] In this embodiment, when evaluating the prediction result in step S6 to prove the accuracy of the prediction result, the mean absolute error E MAE , mean absolute percentage error E MAPE , root mean square error E RMSE and coefficient of determination R 2 are used, and the specific calculation formulas are as follows:

[0133]

[0134] Among them, α t is the original sequence; is the predicted sequence; t = 1, 2, ···, W is the sequence length.

[0135] In this embodiment, the predicted sequence is an intrinsic mode component.

[0136] In order to verify the technical effects that the present invention can produce, a certain working face in a certain mine of Shaanxi Coal Group is taken as the research object. The gas ventilation volume and drainage volume data of this working face from May 16, 2020 to April 27, 2021 are monitored in real time, and the absolute gas emission data is obtained as Figure 2 shown;

[0137] When detecting whether there are outliers in the gas emission data before step S1, the data points outside ±1.5IQR (IQR represents the interquartile range) are taken as outliers, and a box plot is drawn Figure 3 , and it can be seen from Figure 3 that there are no outliers in the absolute gas emission data, and they are all within a reasonable range, which can be used for the prediction of the time series model;

[0138] The data after detection is a total of 347 groups from May 16, 2020 to April 27, 2021;

[0139] In step S1, CEEMDAN is used to decompose the original gas emission time series data once. After decomposition, 7 IMF components and 1 RES residual component are obtained. IMF4 - IMF7 have almost no spikes, the changes are relatively gentle, have a certain periodic fluctuation law, and contain most of the key information in the original data. IMF7 reflects the overall change trend of the original gas emission data. Through this component, it can be known that the gas emission of this working face is small in the first 3 months, and then gradually increases and stabilizes at 45 - 60m 3 / min -1 ; The RES residual component is close to a linear distribution and has strong regularity, which is conducive to the training and learning of the final model; IMF1 - IMF3 contain more high-frequency components. Especially IMF1 has more fluctuations and spikes and has obvious non-stationary characteristics, which is not conducive to the prediction of the final model;

[0140] In order to further evaluate the complexity of each component after the first decomposition of the gas emission data, in step S2, fuzzy entropy calculation is used to calculate the complexity of each component of the gas emission data after the first decomposition by CEEMDAN. The larger the value of the fuzzy entropy, the higher the complexity of the corresponding component; The results are as Figure 4 shown;

[0141] It can be seen from Figure 4It can be seen that the complexity of IMF1 far exceeds that of other components, reaching 1.13, which is 98% higher than that of IMF2 with the second highest complexity. This indicates that this component has the strongest volatility and randomness. If the data including it is directly used for gas emission prediction, a large error will occur. Therefore, it is necessary to use VMD to decompose IMF1 twice to further reduce its complexity;

[0142] When performing VMD decomposition, parameters such as the decomposition layer number, penalty factor, noise tolerance, DC part, initialization method, sampling frequency, and convergence tolerance will affect the decomposition result. However, compared with other parameters, the influence of the penalty factor α and the decomposition layer number k is particularly significant. Therefore, in step S3 of the present invention, CFOA is proposed to optimize these two parameters of VMD.

[0143] The initialization parameters in the optimized VMD algorithm are set as follows: the maximum number of iterations is 10, the population size is 30, the optimization dimension is 2, the optimization range of the penalty factor α is [100, 4000], and the optimization range of the decomposition layer number k is [3, 10]. During the optimization process, the minimum envelope entropy is used as the fitness function to calculate the envelope entropy value corresponding to each parameter combination. When this value is small, it indicates that the time-varying characteristics of the signal are less, and the periodicity and regularity are stronger; that is, it means that the complexity of the signal is lower; the fitness curve is shown in Figure 5 .

[0144] It can be seen from Figure 5 that after 6 iterations, the curve tends to be stable, and until the iteration is completed, the fitness value does not continue to decrease, indicating that the optimal solution has been found in the 6th iteration. The values of α and k after optimization by CFOA are [138, 5] respectively. Substitute the obtained parameter combination into the VMD algorithm to complete the second decomposition of IMF1, and the result is as shown in Figure 6 shown, and the complexity of each component is as shown in Figure 7 shown;

[0145] It can be seen from Figure 6 and Figure 7 that IMF1 with higher complexity is decomposed by VMD to obtain 5 IMF components and 1 RES residual component. The complexity of each component after decomposition is lower than 0.6. Compared with 1.1 before decomposition, the complexity has decreased by 50%, and the smoothness has been significantly improved;

[0146] After two decompositions of the original gas emission time series, a total of 11 IMF components and 2 RES residual components are obtained. To evaluate whether the second decomposition will cause the loss of information in the gas emission data, the original gas emission is compared with the superimposed value of the decomposed components, and the result is as shown in Figure 8 shown;

[0147] It can be seen from Figure 8It can be seen that the coincidence degree between the superposition curves of each component and the original time series curve is relatively high, and the absolute error is between 0.00001 - 0.04395m 3 / min, with an average of 0.00646m 3 / min, indicating that there is almost no decomposition loss in the second decomposition, effectively avoiding the prediction deviation caused by information loss due to decomposition;

[0148] When constructing the gas emission prediction model, the setting of parameters directly affects the overall performance of the model. Therefore, in step S4 of the present invention, FLA is used to optimize the parameters of BiGRU;

[0149] Since different optimization algorithms may show different efficiencies in dealing with gas emission datasets and related problems, in order to evaluate the optimization ability of FLA, its performance is tested; through testing, the performance of FLA and other optimization algorithms on different difficulty levels and types of problems can be compared, thus demonstrating the stability and effectiveness of the algorithm in solving complex problems;

[0150] Two representative test functions are used to test the optimization ability of FLA and compare it with several other common optimization algorithms, including genetic algorithm (GA), particle swarm optimization algorithm (PSO), grey wolf optimization algorithm (GWO), and Bayesian optimization algorithm (BO); the expressions of the two test functions are respectively:

[0151]

[0152] where N represents the initial population size, and x i represents the position of the i-th population;

[0153] Here, the parameters of each optimization algorithm are uniformly set. The population size is set to 30, and the maximum number of iterations is set to 300; the convergence curves of the two optimization functions are as Figure 9A and 9B shown;

[0154] From Figure 9A and 9B it can be seen that as the number of iterations increases, the convergence speed of FLA is faster, and it always maintains the highest search efficiency and accuracy, indicating that FLA performs better than other algorithms on both test functions, verifying its significant advantage in optimizing the gas emission prediction parameters based on the BiGRU model finally.

[0155] BiGRU prediction models are respectively constructed for all subsequences obtained after the first decomposition by CEEMDAN and the second decomposition by VMD;

[0156] The data obtained by decomposing the original gas emission volume are divided into a training set and a prediction set according to a ratio of 8:2 (277 groups in the training set and 70 groups in the prediction set); to improve the prediction performance of the model and overcome the limitations of manual parameter tuning, the initial learning rate and the number of hidden layer nodes of each component BiGRU model are optimized using the training set data through the FLA optimization algorithm, and the optimal values are shown in Table 1; thus, the construction of the gas emission volume prediction model based on BiGRU is completed.

[0157] Table 1 Optimization values of each component of the BiGRU model obtained by FLA

[0158]

[0159] In Table 1, BestA and BestB are respectively the best initial learning rate and the best number of hidden layer nodes for optimizing BiGRU by FLA;

[0160] So far, the CEEMDAN-VMD-BiGRU gas emission volume prediction model proposed by the present invention has been built. To verify the prediction effect of the model, the gas emission volume for a total of 70 days from February 17, 2021 to April 27, 2021 at the working face, which is the prediction set data, is predicted. The prediction results of each component CEEMDAN-VMD-BiGRU model are as Figure 10 shown.

[0161] As Figure 10 can be seen, each component prediction model shows good prediction performance, the overlap between the predicted values and the true values is relatively high, and the distribution of each fitting curve and data further shows that the error between the predicted values and the true values is small and has a high degree of consistency.

[0162] Finally, the prediction results of each component are superimposed to obtain the final gas emission volume prediction result as Figure 11 shown. As Figure 11 can be seen, the final predicted value curve is very close to the true value curve, and the absolute error is between 0.0004 - 0.7928 m 3 / min, with an average of 0.1630 m 3 / min, and the error is small. It can better predict the gas emission volume trend of the working face in the next 70 days, verifying that the proposed model has high feasibility.

[0163] To further verify the prediction accuracy of the BiGRU prediction model of the present invention, 11 different comparison models are also selected for gas emission volume prediction evaluation. The division ratio of the training set and the prediction set of each model is the same as that of the model proposed by the present invention. These models cover different combinations of the decomposition and prediction algorithms used in the present invention. For the convenience of subsequent analysis and discussion, the 12 models are named and numbered according to a certain format, and the naming format of each model is shown in Table 2:

[0164] Table 2 Model naming method

[0165]

[0166] The comparison of each model evaluation index is shown in Table 3 as follows:

[0167] Table 3 Comparison of evaluation indexes of each prediction model

[0168]

[0169] In order to further verify the rationality of the gas emission prediction model combination proposed by the present invention with different decomposition algorithms and prediction algorithms, the following discussions are carried out respectively:

[0170] 1) Direct prediction comparison analysis based on the BiGRU model (M1 and M2); among them, M2 incorporates FLA to optimize the parameters of BiGRU;

[0171] The gas emission prediction models are respectively established by directly using BiGRU and optimized BiGRU for the original data, and the prediction results are as Figure 12 shown. The prediction deviations of each model are as Figure 13 shown.

[0172] From Figure 12 , Figure 13 and combined with Table 3, it can be seen that M1 and M2 have the same change trend, M2 is closer to the true gas emission value, and its various evaluation indexes are also improved compared with M1. This also verifies the improvement of the FLA optimization algorithm on the prediction performance. However, generally speaking, the prediction effects of the two models in different scenarios are not very ideal. This is because the method of directly predicting the original gas emission data fails to capture the non-linear characteristics in the data signal well, and thus it is impossible to accurately judge the change of the gas emission data at the next moment. This phenomenon is more obvious when the gas emission data fluctuates.

[0173] 2) Primary decomposition prediction comparison analysis based on the optimized BiGRU (M2, M6 and M7); among them, M6 and M7 use different primary decomposition methods to decompose the gas emission data first, and then use the optimized BiGRU for prediction.

[0174] As can be seen from the previous section, without data decomposition, directly applying the machine learning model to the gas emission time series data with obvious non-stationary characteristics for prediction cannot achieve ideal results. Therefore, the primary decomposition models of CEEMDAN and VMD are respectively established on the basis of M2. The results are as Figure 14 shown, and the goodness of fit between the predicted values and the true values of each model is as Figure 15 shown.

[0175] The prediction results show that the combined model with the signal decomposition algorithm has better gas emission prediction effect than the single BiGRU and optimized BiGRU. 2 It reached 0.96, and the correlation between the predicted value and the true value was also higher, indicating that the subsequences of the original gas emission time series after signal decomposition processing showed more obvious regularity, which is conducive to the training of the prediction model.

[0176] 3) Comparative analysis of secondary decomposition prediction models based on optimized BiGRU (M6, M7 and M10); among them, M10 uses CEEMDAN-VMD to perform secondary decomposition on gas emission data, and then uses optimized BiGRU for prediction.

[0177] In order to verify that the secondary decomposition can further reduce the nonlinearity and non-stationarity of each subsequence after the primary decomposition and to further explore the data characteristics, a secondary decomposition model of CEEMDAN-VMD based on M6 and M7 is established. The comparison results are shown in Figure 16 As shown in the figure, the goodness of fit between the predicted values ​​of each model and the true value is as follows Figure 17 shown.

[0178] From the comparison results and combined with Table 3, we can see that compared with the single CEEMDAN or VMD decomposition prediction model, the prediction accuracy of the model after CEEMDAN-VMD decomposition is significantly improved. MAE 、E MAPE and E RMSE The average reductions were 0.46, 0.79 and 0.41, achieving a relatively ideal prediction. This further illustrates that for time series with large fluctuations, using only one decomposition is not enough. The method of reasonably using the combined model constructed by the decomposition algorithm to predict gas emission is reasonable and effective.

[0179] 4) Comparative analysis of the quadratic optimization decomposition prediction models based on optimized BiGRU (M10 and CVB); among them, CVB uses CFOA to optimize the key parameters of VMD based on quadratic decomposition.

[0180] The gas emission prediction model proposed in this paper uses CEEMDAN-VMD secondary decomposition and also uses CFOA to optimize two important parameters required for VMD decomposition. In order to verify that optimizing the decomposition algorithm can further improve the prediction accuracy of the model, a CVB final prediction model based on M10 is established. The comparison results are shown in Figure 2. Figure 18 As shown in the figure, the goodness of fit between the predicted values ​​of each model and the true value is as follows Figure 19 shown.

[0181] Depend on Figure 18 , Figure 19It can be seen that the prediction results of both models are very close to the true values. In the local range, the fitting degree of M10 and the true curve even exceeds that of CVB. Generally speaking, however, the prediction performance of CVB is better. Especially in the stage with relatively large fluctuations, its ability to capture non-linear features is significantly better than that of M10, which also proves that the optimized decomposition algorithm can indeed improve the prediction accuracy of the model.

[0182] Based on the above comparison results, it can be concluded that the CVB model proposed in the present invention has the best prediction effect, and the prediction result is the closest to the true value, which also verifies the feasibility of introducing quadratic optimization decomposition and FLA optimization on the basis of the direct prediction model.

[0183] To sum up, the present invention proposes a prediction model for gas emission in coal mining faces based on quadratic decomposition and bidirectional gated recurrent units. The method first decomposes the gas emission data once using CEEMADN, then performs quadratic decomposition on the high-complexity components after decomposition using VMD, constructs BiGRU prediction models for each subsequence obtained by decomposition, and introduces the fish swarm optimization algorithm (CFOA) and flood optimization algorithm (FLA) to optimize the parameters of each model; this combined model fully exploits the non-linear feature information of the original data and gives full play to the respective advantages of each algorithm in processing time series. The present invention compares the prediction effects of different combined models of 12 decomposition and prediction algorithms used in the present invention. The results show that the CEEMADN-VMD-BiGRU model has the best prediction effect, and its MAE E MAPE E RMSE and 2 R

[0184] The above are only the preferred embodiments of the present invention, and do not impose any limitations on the present invention. Any simple modifications, changes, and equivalent structural changes made to the above embodiments according to the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for predicting the gas emission amount in a coal mining face, characterized in that, The method includes the following steps: Step S1: Use the complete adaptive noise ensemble empirical mode decomposition method to perform a primary decomposition on the gas emission data, and decompose it into multiple intrinsic mode components; Step S2: Calculate the complexity of each intrinsic mode component using fuzzy entropy, and set a complexity threshold; Step S3: After optimizing the parameters of the variational mode decomposition method using the fishing optimization algorithm, use the variational mode decomposition method to perform a secondary decomposition on the intrinsic mode components with complexity greater than the complexity threshold to obtain intrinsic mode components with small fluctuations and regularity; Step S4: After optimizing the parameters of the BiGRU prediction model using the flood optimization algorithm, establish a BiGRU prediction model, and use the BiGRU prediction model to train and predict all the intrinsic mode components obtained in Step S1 and Step S3; Step S5: Sum up the prediction results of each intrinsic mode component to obtain the final gas emission prediction result.

2. A method for predicting the gas emission amount in a coal mining face according to claim 1, characterized in that: After Step S5, there is also Step S6: Evaluate the prediction result to prove the accuracy of the prediction result.

3. A method for predicting the gas emission amount in a coal mining face according to claim 1, characterized in that: The specific process of performing a primary decomposition on the gas emission data using the complete adaptive noise ensemble empirical mode decomposition method in Step S1 is as follows: Step S101: Define the original time series signal of gas emission volume data as f(t), and add an adaptively decomposable Gaussian white noise sequence a n (t), which is expressed by the formula as: F n y(t) = f(t) + ε0a n y(t) (F1) Among them, F n (t) is the signal after adding white noise, ε0 is the standard deviation of the originally added white noise, n is the number of decompositions and the value of n is a natural number from 1 to M, and M is the total number of decompositions; Step S102: Decompose the signal F n (t) using the empirical mode decomposition method, take the average value of a group of intrinsic mode components obtained by the decomposition to obtain the first intrinsic mode component IMF1(t), which is expressed by the formula as follows: Among them, IMF1 n (t) represents the first intrinsic mode function component obtained by decomposition when the decomposition times is n; Step S103: Calculate the residual r1(t) obtained from the first decomposition, which is expressed by the formula: r1(t) = f(t) - IMF1(t) (F3) Step S104: Add the Gaussian white noise sequence a n (t) to r1(t), perform M decompositions until it is decomposed into first-order components, and calculate the second intrinsic mode function IMF2(t), which is expressed by the formula: where E1 is the first-order component generated by the empirical mode decomposition method, and ε1 is the standard deviation of the white noise added to the first-order component; Step S105: Calculate the residual r2(t) obtained from the second decomposition, which is expressed by the formula: r2(t) = r1(t) - IMF2(t) (F5) Step S106: Continue to perform empirical mode decomposition until the decomposition condition is not satisfied. At this time, the original signal f(t) is decomposed into q intrinsic mode components and a residual component, which is expressed by the formula: where q is the total number of intrinsic mode components, and IMF i (t) represents the i-th intrinsic mode component, which is expressed by the formula: E i is the i-th order component generated by the empirical mode decomposition method, and ε i is the standard deviation of the white noise added to the i-th order component; r i (t) is the residual obtained from the i-th decomposition and is expressed by the formula: r i r(t) = i-1 r(t) - IMF i r(t)(F8) r i-1 (t) is the residual obtained from the (i - 1)-th decomposition; R(t) is the residual component and is expressed by the formula: R(t) = r q (t) = r q-1 (t) - IMF q (t)(F9).

4. A method for predicting the gas emission amount in a coal mining face according to claim 1, characterized in that: In Step S3, using the fishing optimization algorithm to optimize the parameters of the variational mode decomposition method means using the fishing optimization algorithm to optimize the two parameters of the penalty factor α and the decomposition layer number k in the variational mode decomposition method.

5. A method for predicting the gas emission amount in a coal mining face according to claim 1, characterized in that: The specific process of using the fishing optimization algorithm to optimize the two parameters of the penalty factor α and the decomposition layer number k in the variational mode decomposition method in Step S3 is as follows: Step A1: Initialize the parameters of the fishing optimization algorithm, and set the search range of the decomposition layer number and the penalty factor of the variational mode decomposition method; Step A2: Set the CFOA population size, the maximum number of iterations, and the search dimension; Step A3: Select the minimum envelope entropy as the fitness function; calculate the fitness value corresponding to each parameter combination respectively, and compare it with the fitness value of the previous iteration; Step A4: If it is less than the fitness value of the previous iteration, update the decomposition layer number and the penalty factor; if it is greater than the fitness value of the previous iteration, do not update the decomposition layer number and the penalty factor; Step A5: Continue to iterate until the maximum number of iterations is reached, and output the optimal values of the decomposition layer number and the penalty factor; Step A6: Substitute the optimal values into the VMD to complete the decomposition.

6. A method for predicting the gas emission amount in a coal mining face according to claim 1 or 4, characterized in that: In step S2, the variational mode decomposition method is used to perform secondary decomposition on the gas emission data. The alternating direction multiplier method is used to continuously update the center frequency and bandwidth of each mode. Finally, each mode and the corresponding center frequency are obtained, which is expressed by the formula: Among them, Wiener filtering representing the current margin, ω k represents the center frequency of the current modal component, represents the Fourier transform of f(t), and f(t) represents the original signal processed through step S1; represents u i (t)’s Fourier transform, u i (t) represents the modal function of the i-th decomposition in the variational mode decomposition method; represents the Fourier transform of λ(t), and λ(t) represents the Lagrange multiplier; k is the number of modal functions, n is the number of iterations, α is the penalty factor, ω is the frequency variable, representing any frequency point within the frequency domain range, ω k is the oscillation frequency representing the modal function; represents u k (t)’s Fourier transform, u k (t) represents the modal function of the k-th decomposition in the variational mode decomposition method.

7. A method for predicting the gas emission amount in a coal mining face according to claim 1, characterized in that: In step S4, the flood optimization algorithm is used to optimize the parameters of the BiGRU prediction model, which optimizes the initial learning rate and the number of hidden layer nodes of the BiGRU prediction model.

8. A method for predicting the gas emission amount in a coal mining face according to claim 1 or 7, characterized in that: In step S4, the output of the BiGRU prediction model is determined by the weighted sum of the results of a forward hidden layer and a backward hidden layer. The calculation formula is as follows: Among them, GRU(·) is a gated recurrent unit, and x t is the input data, is the output of the previous moment in the forward direction, is the output at the previous moment from back to front; h t is the output, is the output of the forward hidden layer, is the output of the backward hidden layer; α t is the weight of the forward hidden layer, β t is the weight of the backward hidden layer; b t is the bias.

9. A method for predicting the gas emission amount in a coal mining face according to claim 2, characterized in that: When evaluating the prediction results described in step S6 to prove the accuracy of the prediction results, the mean absolute error E MAE , mean absolute percentage error E MAPE , root mean square error E RMSE and coefficient of determination R 2 are used. The specific calculation formulas are as follows: Among them, α t is the original sequence; is the predicted sequence; t = 1, 2, ···, W is the sequence length.

Images