A Prediction Method for Height of Water Flowing Fractured Zone Based on Sparrow Search Algorithm–Elman Neural Network in Northwest Mining Area

Abstract

The main Jurassic coal seams of the Ordos Basin of northwest mining area have special hosting conditions and complex hydrogeological conditions, and the high-intensity coal mining of the coal seams is likely to cause groundwater loss and negative effects on the surface ecological environment. The research was aimed at predicting the height of the water-flowing fractured zone (WFFZ) in high-intensity coal mining in that area and gave instructions for avoiding water inrush accidents and realizing damage reduction mining during the actual mining procedure of the coal mine. In this study, 18 samples of the measured height of WFFZ in Jurassic coal seams were systematically collected. In the mining method, the ratio of the thickness of the hard rock to the thickness of the soft rock in the bedrock, buried depth, mining height, and working face length was selected as the input vectors, applied the sparrow search algorithm (SSA) to iteratively optimize the weights and thresholds of the Elman neural network (ENN), constructed an SSA-Elman neural network model. The results demonstrate that the improved SSA-Elman neural network model has higher accuracy in predicting the height of the WFFZ compared with traditional prediction algorithms. The results of this study help guide damage-reducing, water-preserving mining of the middle-deep buried Jurassic coal seams in the northwest mining areas.

1. Introduction

Mining-induced fractures are the principal pathways for groundwater seepage. During coal mining, changes in the structure of the overburden strata of the coal seam and the heights and spatial characteristics of the mining-induced fractures are the basic premise for water-preserving and damage-reducing coal mining techniques [1]. An extraordinary deal of practical and theoretical research has been conducted. Qian et al. [2] established the widely accepted three horizontal zones and three vertical zones characteristics of rock formation movement-induced fractures and the mining-induced fracture distribution in the stope. Xu [3] proposed a strategy for predicting the height of the WFFZ based on the location of the primary overburden strata. By studying the mining conditions of the shallow coal seams in Shendong and Northern Shaanxi, Huang [4] established the step rock beam structure model of the overburden strata in shallow coal seams and determined the rules for upward fracture development and downward fracture development. Zhang [5] preliminary established an empirical equation for the height of the WFFZ during Jurassic coal seam mining in the western mining area, which provides guidance for the prevention and control of roof water hazards and water-preserving mining of the western shallow coal seams. Li et al. [6] and Xie et al. [7] considered different influencing factors and used a back propagation (BP) neural network, support vector machine, and an improved model [8,9,10] to predict the height of the water-flowing fractured zone.

The northwest-mining area is an important coal production base in China, where the structure of the overburden bedrock of most of the Jurassic middle-deep buried coal seams in the Inner Mongolia-Shaanxi border area is dominated by weakly cemented Cretaceous and Jurassic strata. The development started relatively late, and the exploration was less extensive. Affected by major factors, such as overburden strata, mining size, and mining rate, the predicted height of the water-flowing fractured zone according to the empirical formula diverges greatly from the actual height.

Based on the geology and technical mining conditions of the middle-deep under-ground Jurassic coal seams in the Ordos Basin and height measurements of the WFFZ and comprehensively considering factors such as the burial depth, the structure of the weakly cemented overburden strata, and the high-intensity mining conditions, we used the sparrow search algorithm (SSA) to optimize the Elman neural network (ENN) and developed an SSA-Elman neural network coupling model to predict the heights of the WFFZ in the study area.

2. Geological Setting

The study area is located in the border area between Inner Mongolia Autonomous Region and Shaanxi Province in Northwest China. It is the main coal-producing area in China, including Hujirt, Nalinhe, Xinjie, and Yushen mining areas (Figure 1). The strata in the study area are Quaternary (Q), Cretaceous Zhidan group (K₁zh), Jurassic Anding group (J₂a), Zhiluo group (J₂z), and Yan’an group (J₂y) from top to bottom. The surface is generally covered by Quaternary aeolian sand and Salawusu Formation aquifer, with abundant water content. The Cretaceous aquifer has a large thickness, medium to strong water richness, and poor cementation. The Yan’an Formation and Zhiluo Formation are distributed with multi-layer confined aquifers, and there is no stable aquifuge between them and coal seams.

The main coal-bearing strata in the research area are the Jurassic Yan’an Formation, and the already put into operation mines such as Muduchaideng, Nalinhe No.2 mine and Menkeqing. The thickness of the main coal seam is 2.8–15 m, and the buried depth is 280–1000 m, the working face is under high-intensity mining, and the maximum water volume of the working face has reached 1000 m³/h. In the process of mining, a series of problems, such as mixed coal and water, drainage difficulties, and roof water inrush, have seriously affected the production efficiency and safety of the mine.

3. Influencing Factors and Data Preprocessing

3.1. Selection of Influencing Factors

The Ordos Basin is currently China’s main coal-producing and energy strategy implementation region [11,12,13]. As the scale and intensity of the Jurassic coal seam mining in the Ordos Basin increase, it has gradually entered deep mining. Due to the complexity of geological structures and aquifers, accurate prediction of the height of water flowing fractured zone has become an important content of water-preserving mining.

Based on the engineering geology and technical mining conditions of the Jurassic coal seams in the middle and deep part of the Inner Mongolia-Shaanxi border area, five influencing factors, i.e., mining method, the ratio of the thickness of hard rock layers and soft rock layers in the bedrock, buried depth, mining height, and working face length, were used as the model’s input vectors. The measured height of water flowing fractured zone was used as the model’s output vectors. The reasons for selecting these factors as predictive features are as follows:

1.

Mining method (I_m);

The mining method has specific requirements for the mining height. Different mining methods selected under different geological conditions have different effects on the damage degree and damage range of the roof.

2.

The ratio of the thickness of hard rock layers and soft rock layers in the bedrock (I_r);

The bedrock in this area is thick, the soft and hard rock layers are distributed alternately, and there are many key strati in bedrock that control rock strata movement and mining-induced fracture development, which brings great difficulty to the accurate prediction of the height of WFFZ. Rock strata with a Platts coefficient (the value is 1/10 of the uniaxial compressive strength limit of rock or soil) less than 3 were defined as soft rock, and rock strata greater than or equal to 3 were defined as hard rock. In the process of working face mining, soft rock is apt to produce plastic strain, closes the overburden fractures, in turn, makes the height of WFFZ decrease. While hard rock strata are prone to cause deformation and failure in a larger range after fracture, resulting in a highly nonlinear mutation of the WFFZ. Therefore, the ratio of the thickness of the hard rock to the soft rock within the range of bedrock is chosen as an index reflecting the structure of the overlying rock.

3.

Buried depth (I_d);

According to the theory of rock mechanics, the deeper the coal seam is buried, the greater the stress on the overlying rock, and the more drastic the degree of deformation and destruction of the corresponding overlying strata after mining.

4.

Mining height (I_h);

In the empirical formula for calculating the height of WFFZ, the mining height is an essential parameter. The larger the mining height, the greater the scope of plastic failure zone of rooves.

5.

Working face length (I_l).

The length of the working face is an important factor affecting the development of the WFFZ. Before critical mining is achieved, the height of the water-flowing fractured zone increases with the increase of the working face length [14].

3.2. Data Preprocessing

Because the magnitudes of the selected predictors are not the same, the min-max normalization method was used to normalize the sample matrix, with values of 0.4 and 0.6 for fully mechanized mining and fully mechanized top caving, respectively [10]. The formula is as follows:

$${\mathrm{y}}_i=\frac{x_i-x_{\mathit{min}}}{x_{\mathit{max}}-x_{\mathit{min}}}$$

(1)

where y_i is the normalized data, I = 1, 2…, n, and x_min and x_max are the minimum and maximum values of the factors, respectively.

According to the selected predictors, the measured heights of WFFZ in the mining of several Jurassic coal seams in the Ordos Basin were collected, The Inner Mongolia-Shaanxi border area mining area was taken as the research object, the mining parameters of 18 working faces were analyzed in detail. The mining methods mainly include fully mechanized mining and fully-mechanized top coal caving. The thickness ratio of hard rock and soft rock in bedrock ranges from 1.45 to 3.56, the mining height ranges from 2.75 to 10.1 m, and the burial depth ranges from 248 to 692 m. The length of the working face is between 90 and 360 m, and the measured height of the WFFZ is between 45.6 and 245.51 m. The five selected influencing factors and the measured height data of the WFFZ were normalized, respectively, and the processed data are between 0 and 1, and the deep learning sample data and normalized values were established (

Table 1. Measured values and normalized values of the heights of the WFFZ.

Panel	Measured Value						Normalized Value
	Im	Ir	Ih (m)	Id (m)	Il (m)	Ia (m)	Im	Ir	Ih	Id	Il	Ia
Hulusu 21102	ZC 1	1.53	2.85	634.8	320	45.6	0	0.04	0.01	0.87	0.85	0.00
Hongqinghe 31101	ZC	2.82	6.5	669	240	106.1	0	0.65	0.51	0.95	0.56	0.30
Nalinhe mine #2	ZC	2.93	5.5	580	247	103.23	0	0.70	0.37	0.75	0.58	0.29
Hanglaiwan 30101	ZC	3.13	4.8	248	300	122.02	0	0.80	0.28	0.00	0.78	0.38
Tingnan mine 107	ZF 2	1.45	9.9	650	116	165.83	1	0.00	0.97	0.91	0.10	0.60
Xiagou mine 2801	ZF	1.45	9.9	332	90	149	1	0.00	0.97	0.19	0.00	0.52
Caojiatan 122106	ZC	3.13	6.0	280	360	136.10	0	0.80	0.44	0.07	1.00	0.45
Huoshizui mine 8712	ZF	1.45	10	628.16	200	220	1	0.00	0.99	0.86	0.41	0.87
Dafosi 40106	ZF	1.45	9.1	450	180	245.51	1	0.00	0.86	0.45	0.33	1.00
Hujiahe 401101	ZF	1.45	10.1	644.52	180	225.43	1	0.00	1.00	0.89	0.33	0.90
Yushuwan 20104	ZC	3.13	5	280	255	138.40	0	0.80	0.31	0.07	0.61	0.46
Jinjitan 12−2, upper 101	ZC	3.13	5.0	260	300	120.25	0	0.80	0.31	0.03	0.78	0.37
Lingxin mine L331	ZC	3.56	2.75	250	280	57.35	0	1.00	0.00	0.00	0.70	0.06
Hongliu mine 1121	ZC	3.56	5.3	330	302	62.5	0	1.00	0.35	0.18	0.79	0.08
Menkeqing 11-3102	ZC	1.53	4.75	692	300	117	0	0.04	0.27	1.00	0.78	0.36
Muduchaideng mine, coal seam 3-1	ZC	1.53	4.8	685	250	106	0	0.04	0.28	0.98	0.59	0.30
Jinfeng mine 011802	ZC	3.56	4.6	500	280	63.12	0	1.00	0.25	0.57	0.70	0.09
Shilawusu mine 221, upper 06A	ZF	1.53	9.5	660	300	240	1	0.04	0.92	0.93	0.78	0.97

¹ ZC denotes fully mechanized mining. ² ZF denotes fully-mechanized top coal caving.

).

3.3. Correlations among Input Features and Output

Pearson correlation coefficient (PCC) was used to analyze the correlation degree of coal seam mining conditions and parameters to the height of WFFZ. The Pearson correlation coefficient is defined as follows:

$${\rho }_{X,Y}=\frac{\mathrm{cov}\left(X,Y\right)}{{\sigma }_X{\sigma }_Y}$$

(2)

where ${\rho }_{X,Y}$ is the Pearson correlation coefficient, cov(X, Y) is the covariance of variables X and Y, ${\sigma }_X$ and ${\sigma }_Y$ are the standard deviation of variables X and Y, respectively. The range of ${\rho }_{X,Y}$ is between −1 and 1, when the value is 1, it denotes that there is a wholly positive correlation between the two random variables; when the value is −1, it indicates a fully negative correlation between the two random variables; when the value is 0, it implies linear independence between the two random variables. The greater the absolute value of the correlation coefficient between the two random variables, the stronger the linear correlation between the variables.

Figure 2 is a distribution diagram of the Pearson correlation coefficient analysis results of input features and output. As shown in Figure 2, mining method, mining height, and buried depth are positively correlated with the height of WFFZ, and the ratio of the thickness of hard rock layers and soft rock layers in the bedrock and working face length are negatively correlated with the height of WFFZ. Among them, mining height (PCC: 0.86) and mining method (PCC: 0.84) had a strong positive correlation with the height of WFFZ, followed by buried depth (PCC: 0.24). the ratio of the thickness of hard rock layers and soft rock layers in the bedrock (PCC: −0.6) and working face length (PCC: −0.46) have a strong negative correlation with the height of WFFZ. Under the mining conditions of middle-deep coal seams, more hard rock layers in the overlying rock structure have a greater impact on the development height of the WFFZ. The greater the thickness of the hard rock layer in the bedrock, the smaller the height of WFFZ.

4. SSA-ELMAN Neural Network Model

4.1. Elman Neural Network

The Elman neural network is a partial feedback neural network with a feed forward connection. It is composed of an input layer, a hidden layer, a context layer, and an output layer. The Elman neural network adds a context layer to the original structure of the BP neural network, and it also propagates the data output from the hidden layer back to the hidden layer, which enhances the dynamic memory ability of the network and the ability to process nonlinear problems [15,16,17,18]. The network’s structure is shown in Figure 3.

The Elman neural network mathematical model is as follows:

$$\left\{\begin{array}{l}X\left(k\right)=f\left\{W_1{X}_c\left(k\right)+W_2\left[U\left(k-1\right)+b_1\right]\right\}\hfill \\ X_c\left(k\right)=X\left(k-1\right)\hfill \\ Y\left(k\right)=g\left[W_{3}X\left(k\right)+b_2\right]\hfill \end{array}\right.$$

(3)

where f and g are the hidden layer and output neurons’ transfer functions, respectively, and W₁, W₂, and W₃ are the hidden layer’s connection weights to the input layer, the context layer’s connection weights to the hidden layer, and the hidden layer’s connection weights to the output layer, respectively. U(k−1) and Y(k) are the input vectors of the input layer and the output vector of the output layer, respectively; b₁ and b₂ are the weights of the input layer and hidden layer respectively; and X(k) and X_c(k) are the outputs of the hidden layer and the context layer, respectively.

4.2. Sparrow Search Algorithm

The SSA is a new swarm intelligence optimization method that draws inspiration from sparrows’ foraging habits. in nature [19]. The sparrow search algorithm has strong stability and quick convergence time when compared to other intelligent optimization algorithms, and it can successfully prevent slipping into a local optimum [20,21,22,23]. The sparrow population is usually composed of producers, scroungers, and guards. We assume that the variable dimension of the problem to be solved is D-dimensional and the number of sparrow population is n. Then the position of ith sparrow in the D-dimensional solution space is X_i = (x₁, x₂..., x_D) and the fitness value is f_i = f (x₁, x₂..., x_D). In the sparrow search algorithm, the reserve energy of an individual sparrow depends on its fitness value, the producers with greater fitness are given priority in the search procedure for food.

1.

Producers;

The producers are usually healthy and fit and are in charge of giving the foraging area and guidance for the scroungers. Their locations are updated as follows:

$$X_{i,j}^t=\left\{\begin{array}{l}{X}_{i,j}^t\cdot \mathit{exp}\left(-\frac{i}{\alpha \cdot ite{r}_{\max }}\right){ \mathrm{if} R}_2< \mathit{ST}\hfill \\ X_{i,j}^t+Q\cdot L{ \mathrm{if} R}_2 \ge \mathit{ST}\hfill \end{array}\right.$$

(4)

where X_ij is the position of the ith sparrow in the jth dimension, j = 1, 2..., d, where d is the solution’s dimension, t is the current iteration, iter_max is the population’s maximum number of iterations, α is a random number between [0, 1], R₂ ∈ [0, 1] and ST ∈ [0.5, 1] indicate the alarm value and the safety threshold, respectively, Q is a random number following a normal distribution, and L is a one-dimensional matrix with all members equal to 1.

2.

Scroungers;

The scroungers always monitor the producers. They will compete for resources and become new producers once they see the producers have moved to an area with greater food. The locations are updated as follows:

$$X_{i,j}^{t+1}=\left\{\begin{array}{l}Q\cdot \mathit{exp}\left(\frac{{\mathrm{X}}_{\mathit{worst}}-X_{i,j}^t}{i^2}\right) \mathit{if} i >\frac{m}{2}\hfill \\ X_p^{t+1}+\left|X_{i,j}^t-X_p^{t+1}\right|\cdot A^{+}\cdot L \mathit{if} i \le \frac{m}{2}\hfill \end{array}\right.$$

(5)

where X_p represents the current global best position sought by the current producer, and X_worst is the current global worst position sought by the current producer, A is a 1 × d matrix with all its elements randomly assigned as 1 or −1, and m is the sparrow population. When I > m/2, the ith scrounger is starving and needs to fly to other places to find food.

3.

Guards;

The number of guards is usually 10% to 20% of the population. When they encounter a predator, the guards will issue an alarm. The producer will direct the gathering to a safe spot if the alert signal exceeds the safe threshold. The guards are randomly generated, and their locations are updated as follows:

$$X_{i,j}^{t+1}=\left\{\begin{array}{l}{X}_{i,j}^k+\beta \left|X_{i,j}^k-X_{b,j}^k\right|{ \mathit{if} f}_i>{ f}_g \hfill \\ X_{i,j}^t+K\cdot \left(\frac{\left|X_{i,j}^k-X_{worst}\right|}{\left(f_i-f_g\right)+\epsilon }\right){ \mathit{if} f}_i={ f}_g\hfill \end{array}\right.$$

(6)

where β represents the step size control parameter, which obeys a normal distribution of N (0, 1), The direction of the sparrow population’s search movement is indicated by k, which is a random number between [0, 1], f_i is the current sparrow’s fitness value, and f_g and f_w are the current global best and worst fitness values, respectively. When f_i > f_g, the sparrows on the periphery of the group are threatened by predators. When f_i = f_g, the sparrows in the population will move closer to each other.

The producers, scroungers, and guards update their locations by constantly comparing the fitness values. When the set termination conditions (the maximum number of iterations) is achieved, the algorithm ends.

4.3. Construction of Prediction Model Based on SSA-Elman Neural Network

The flow chart of the prediction model for the height of the WFFZ based on the SSA-Elman neural network is shown in Figure 4.

Step 1: Model parameter setting

The Elman neural network model parameters were set as follows. The number used for network training was 1000, the minimum training error was 0.0001, the learning rate was 0.01, and the momentum factor was 0.01.

The SSA algorithm model parameters were set as follows. The sparrow population was 30, the maximum number of evolutionary generations was 300, the safety threshold was 0.6, the proportion of producers was 0.7, the target error was 1 × 10⁻⁶, and the proportion of guards was 0.2.

Step 2: Model training

First, the following empirical formula for the hidden layer nodes [24] was used to determine the network structure.

$$num=\sqrt{\left(m+n\right)}+a$$

(7)

where num denotes the quantity of hidden layer nodes, m indicates the amount of input layer nodes, n represents the number of output layer nodes, and a represents an integer between [1,10]. The best number of hidden layer nodes was determined to be seven when the training was completed.

The SSA was then utilized to determine the Elman neural network’s ideal weights and thresholds. The first 13 samples were selected from the normalized data as the training set, when the number of network training reaches 13 times, the prediction error of the model can reach 3.8 × 10⁻⁷, which is lower than the preset target error value of 1 × 10⁻⁶. At this time, the model has met the demand for accuracy and can be used for predicting the height of the WFFZ for the test set. The curve of error convergence during the training procedure is shown in Figure 5.

The mean square error was used as the fitness function, and according to the initial fitness calculation result of the population, the ranking is performed to determine the best and worst-adapted individuals. Based on Equations (3)–(5), the fitness values of the current individuals are calculated and the positions of discoverers, joiners, and guards are updated respectively, if the new position is preferable to the position of the previous iteration, it is updated until the termination condition (the model reaches the maximum training times or meets the demand for accuracy) is satisfied, the global optimal value is found and given to the weights and thresholds of the ELMAN neural network. The fitness function curve during the training process is shown in Figure 6.

The fitness leveled out as the number of iterations reached 36, and the curve leveled off. The obtained optimal weight matrix and threshold matrix were updated onto the neural network, and the last five groups of collected samples were used to evaluate the regression effects of the different models.

Step 3: Model evaluation

The mean absolute error (MAE) and root mean square error (RMSE) were used to evaluate the performances of the prediction models:

$$MAE=\frac{1}{n}{\displaystyle \sum _{i=1}^n\left|{\widehat{y}}_i-y_i\right|}$$

(8)

$$RMSE=\sqrt{\frac{1}{n}{\displaystyle \sum _{i=1}^n{\left({\widehat{y}}_i-y_i\right)}^2}}$$

(9)

where ${\widehat{y}}_i$ represents the predicted value, y_i represents the true value, and n represents the total number of samples.

The trained SSA-Elman neural network prediction model was tested using the test samples to determine the predicted value of the test sample, and it was compared with the performance using the predicted values of the empirical formulas. The results are shown in Figure 7.

The contrast of the prediction performances of the various models is shown in

Table 2. Contrast of prediction performances of various models in the evaluation stage.

Model		MAE	RMSE
SSA-ELMAN		3.93	4.93
ELMAN		55.75	64.54
BP		43.29	53.92
Empirical Equation (1) *	$\begin{array}{l}H=4.82M+60.13\ln \frac{s}{100}\\ +3.43M\ln \frac{b}{100}+16.17\end{array}$	61.70	64.72
Empirical Equation (2)	$H=\frac{100M}{1.6M+3.6}\pm 5.6$	67.86	92.01

* H is the height of WFFZ; M is the buried depth; s is the mining height; b is the working face length.

. The MAE values and RMSE values for the predicted values of the height of the WFFZ obtained using the SSA-Elman neural network model are 3.93 and 4.93, respectively, which are less than the corresponding results calculated using the other methods. The results show that using the SSA to optimize the Elman neural network model achieves a higher accuracy than the other prediction methods.

5. Verification of the Prediction Model for Height of WFFZ

5.1. Engineering Background

Coal Mine A is in the Hujirt mining area, Jurassic coalfield, Ordos Basin (Figure 1). The main coal seam is the Jurassic Yan’an Formation 3⁻¹ coal seam. It is 610.62–626.17 m deep, with an average depth of 620 m, and it has inclination angles of 1~4° and an average thickness of 5.72 m. The long-wall fully mechanized mining scheme is adopted for panel 31110X, with a mining height of 5.3 m, a panel width of 260 m, and a mining length of 2589.3 m. The full caving method is used for roof management.

The positions of the roof aquifers relative to the 3⁻¹ coal seam are shown in Figure 8. The immediate roof of the 3⁻¹ coal seam is mainly sandy mudstone, and its basic roof is mainly sandstone. There are multiple aquifers above the 3⁻¹ coal seam, and the bedrock aquifers from the top to the bottom are as follows: the Lower Cretaceous Zhidan Formation (K₁zh) pore confined groundwater aquifer, the Middle Jurassic (J₂) clastic rock confined groundwater aquifer, and the Middle-Lower Jurassic Yan’an Formation (J_1-2y) clastic rock confined water aquifer. The direct filling groundwater source of the 3⁻¹ coal seam mining is the top aquifer of the Yan’an Formation and the middle aquifer of the Zhiluo Formation. According to the hydrogeological survey, this aquifer group has a strong water abundance, and it is prone to causing abnormal roof water busting during roadway excavation and face mining, which threatens the safety of the mining operations.

Input the relevant mining conditions of panel 31110X in the Coal Mine A into the trained prediction model, and the predicted height of the water-flowing fractured zone at the mine resulted in 124.9 m, with mining fissures lead to the aquifer of Zhiluo Fm. In order to verify the performance of the model, we measured the actual height of the water-flowing fractured zone of panel 31110X.

5.2. Field Measurements of Height of WFFZ

In Coal Mine A, panel 31110X near the stop line was selected for T1 drilling. Using the upward water injection method, the height of the WFFZ was measured on-site. The detection instruments used in this experiment are shown in Figure 9a.

Figure 9b depicts the borehole positions and measuring point layout. The depth of the borehole is 180 m, and the angles with respect to the horizontal and vertical directions are 55° and 12°, respectively. Twenty-one measuring points were set up for the water pressure testing. They are divided into three layout patterns. Two measuring points, 10 m apart, were located at the end of the hole and at a vertical height of 140 m, respectively. Fifteen measuring points, 3 m apart, were located between vertical heights of 140 to 100 m. These measuring points were used to search for the highest point of the water-flowing fractured zone. Three points, 15 m apart, were located at vertical heights of 100 m and below.

Water pressurized into the formation at different measuring points in the borehole under different pressures was shown in Figure 10. At vertical heights of 93, 101, 110, and 124 m, the pressurized water volume reached a peak and then gradually decreased. In addition, water could not be injected at some of the measuring points between 90 and 120 m, indicating there should be no water-conducting fractures. And the total water injected at each measuring point in the borehole under different pressures was shown in Figure 11. After 126 m, the volume of water pressurized into the formation tended to level off, indicating that this position was close to the top of the water-flowing fractured zone. Based on a comprehensive analysis of the changes in the total amount of water injected and the amount of water pressurized into the formation in each test section, the measured height of the WFFZ in panel 31110X was 126 m.

5.3. Model Verification

The factors affecting the height of the WFFZ of panel 31110X in the Coal Mine A were input into the SSA-Elman neural network, Elman neural network, and BP neural network prediction models. Meanwhile, the corresponding parameters were substituted to calculate the results of the empirical formulas.

Table 3. Comparison of the prediction results of the different models.

Model		Height of WFFZ (m)	Proportionality Coefficient	Absolute Error (m)	Relative Error (%)
SSA-ELMAN		124.9	23.51	1.1	0.87
ELMAN		168.9	31.87	42.9	34.05
BP		69.2	10.72	56.8	45.08
Empirical Equation (1)	$\begin{array}{l}H=4.82M+60.13\ln \frac{s}{100}\\ +3.43M\ln \frac{b}{100}+16.17\end{array}$	166.8	31.47	40.8	32.38
Empirical Equation (2)	$H=\frac{100M}{1.6M+3.6}\pm 5.6$	49.47	9.33	76.53	60.74

shows a comparison of the prediction results.

As shown in Table 3, the heights of the water flowing fractured zone are projected to be 49.47–168.9 m by the Elman neural network, BP neural network and empirical formulas. With respect to the in-situ measurement of the height of 126.0 m, the absolute error range is 40.8–76.53 m and the relative error range is 32.38–60.74%. While the SSA-Elman neural network predicted that the height of the WFFZ of coal mine A is 124.9 m, and the relative error range is 0.87%. It is obviously observed that the predicted value of the WFFZ by the SSA-ELMAN model is closer to the measured value, while the values predicted by the empirical formulas, BP neural network, and Elman neural network have a larger error. It is because the empirical formulas that only consider mining height, buried depth, and working face length, the Elman neural network itself has a good learning ability and considers two more factors, the mining method, the ratio of the thickness of hard rock layers and soft rock layers in the bedrock, which makes the prediction more consistent with the in-situ measurement value. Meanwhile, compared with BP neural network, ELMAN neural network has better performance than BP neural network in terms of computational power and stability due to the addition of undertaking layer. Using sparrow search optimization algorithm to optimize ELMAN neural network can effectively improve the convergence speed and stability. In the future, the influence of overlying strata structure on the development of water-flowing fractured zone will be more comprehensively considered. Meanwhile, we consider expanding the collection range of highly correlated data of water flowing fractured zone, further improving the convergence speed and reducing the training time of the model, and improving the accuracy, in order to be applied in a wider range of mining areas.

5.4. Feature Importance

Analysis of feature importance can better interpret the value of existing features and infer which features may be most favorable for predicting. To grasp the influence of input features on output, the importance of input features was analyzed by the algorithm Random Forest [25]. The number of trees in the random forest was set to 100, the deepest depth was 4, and the feature importance calculated by the algorithm Random Forest on the original data set are shown in Figure 12. The horizontal axis is on behalf of the importance score of the input features, and the vertical axis represents the input features. The results manifest that mining height (importance score: 0.416) has the greatest influence on the height of the WFFZ prediction, followed by mining method (importance score: 0.251), working face length (importance score: 0.152), the ratio of the thickness of hard rock layers and soft rock layers in the bedrock (importance score: 0.146), and buried depth (importance score: 0.035).

6. Conclusions

(1)

The Jurassic coal seams in the Inner Mongolia-Shaanxi border area are middle-deep buried underground and have special overburden strata. We considered five factors (mining method, the ratio of the thickness of hard rock layers and soft rock layers in the bedrock, buried depth, mining height, and working face length) that influence the development of the height of the water-flowing fractured zone. In particular, the ratio of the thickness of the hard rock layers to the thickness of the soft rock layers in the bedrock was selected as a parameter to characterize an overburdened structure. The linear correlation between coal seam mining conditions and parameters and the height of WFFZ is analyzed by the Pearson correlation coefficient, which offers a foundation for developing a prediction model.

(2)

An SSA-Elman neural network water-flowing fractured zone height prediction model was established using the sparrow search algorithm to optimize the weights and thresholds of the original Elman neural network. This model can fit the nonlinear relationship between the development of the WFFZ and its influencing factors better in the Inner Mongolia-Shaanxi border area.

(3)

In response to the requirement for accurate water-transmission fissure heights in the middle and deep Jurassic coal seams, we used the MAE and RMSE to quantitatively evaluate the performances of the different models. The results show that the MAE and RMSE of the SSA-Elman neural network model in the evaluation stage are 3.93 and 4.93, respectively, which indicates that this model has the best prediction accuracy compared with traditional calculation methods and prediction models.

(4)

Using the mining conditions of panel 31110X of Coal Mine A as an engineering example, we obtained a prediction value of 124.9 m using the SSA-Elman neural network model. This has the smallest error of the prediction models compared with the measured value of 126 m, demonstrating that the SSA-Elman neural network model developed in the study is more accurate in predicting the height of the water flowing fractured zone and more closely represents the real situation. Thus, it can provide guidance regarding the water inflow from a mining face and the use of advanced water discharge techniques in the Inner Mongolia-Shaanxi border area.