Development process and height of the mining-induced water fractured zone over the longwall goaf

Numerous works have been done on the monitoring and evaluation of the height of the WFFZ. In terms of monitoring the WFFZ in a coalfield, Palchik (2010) found that the ratio between the maximum height of the WFFZ and the thickness of the extracted coal seam increases with the increasing number of rock layer interfaces and decreases with the increasing stiffness of the immediate roof. Miao et al. (2011) believed that borehole photography and borehole TV camera techniques are the most effective methods to detect fractures that have developed inside the overlying strata. On the basis of the quantity and energy distributions of the microseismic events recorded using the microseismic monitoring technique, a zoning method was established by Cheng et al. (2017) for the roof strata movement in the vertical and horizontal directions. Liu et al. (2018) and Du et al. (2021) used the Brillouin optical time-domain reflectometry to detect the mining-induced fractures of rock and soil layers, respectively. What is more, an in-site monitoring method based on formation micro-scanner image technology for the WFFZ penetrating into N₂ laterite was carried out first by Liu et al. (2020) in western China. Using borehole video observation equipment, Wu et al. (2021) successfully observed the maximum height of the fissure zone and the location of bed separations at a depth of more than 600 m at a China coal mine. However, although field data are valuable, it is an expensive, time-consuming, and challenging task to measure on site the deformation characteristics of all overlying strata within the WFFZ. In addition, before coal seam mining, the in situ test cannot evaluate the height of the WFFZ due to the lag of the testing process.

From the perspective of evaluating the height of the WFFZ, the common methods include empirical formula, analytical calculation of mechanical model, physical model, and numerical simulation. For example, according to regression analysis of measured results of numerous coalfields in north China, an empirical formula was proposed by Liu (1981), which has been widely used in China. Key stratum theory is a mechanical model which takes into account both the layered characteristics of overlying strata and the difference in deformation characteristics of different rock formations (Qian et al. 1996). The key stratum theory believed that the entire or partial overburden strata movement after mining is controlled by strata structures of the ‘key stratum’, and the formation process of the WFFZ is closely related to the structural characteristics of the key stratum. Additionally, for the accurate estimation of the WFFZ height, the theoretical prediction method was introduced by Huang et al. (2018) based on the overlying composite strata structure and an empirical equation. Based on the failure characteristics of different rocks, Wang et al. (2012) established a mechanics model to predict the height of the WFFZ using strain and stress as the indices of water flow in soft and hard rock strata. Based on the plate and shell theory, a model to calculate the height of the WFFZ was established by Zhu et al. (2020), and the development of the WFFZ in the bedrock and loess layer was discussed in detail. Several important discoveries were obtained in the previous research and helped to strengthen the prediction of the WFFZ height. However, the development of the WFFZ is a very complicated nonlinear issue that is influenced by many factors, mainly geological factors and mining factors (Fan et al. 2022b). Those above have several defects in practical application, such as complex calculation and large prediction errors.

Reasonable consideration of the contribution discrepancy of various factors in the WFFZ height is the premise of accurate prediction. In this paper, taking the No. 839 coalface in the Qingdong mining area, Anhui, China, as a case study, an RBF model is established based on 39 collected groups of measured data. In this RBF model, besides the mining thickness, mining depth, and inclined length of the coalface (LC), the proportion coefficient of hard rock is introduced to quantify the overburden structure. Furthermore, parallel electrical monitoring (PEM) technology and numerical simulation method are also used to study the entire process of WFFZ changes, including the evolution of height and shape. Finally, the accuracy of the RBF model was verified by measured data. The results are significant for mining subsidence control, methane drainage, and preventing water inrush from the roof.

The Huaibei coal mining area is a concealed coalfield covered by loose layers (Figure 2(b)). The Subei fault divides the Huaibei coalfield into north and south parts. The thickness of the loose layer in the north of Huaibei Coalfield is relatively thin, with an average of 50–150 m. The thickness of the loose layer in the south is relatively thick, with an average of 150–350 m, and the thickest part is 667 m. According to the pumping test data over the years, the maximum water inflow at the bottom of the loose layers is 2.389 L×(m×s)⁻¹, which is the main source of roof water inrush in the working face near the loose layers.

In recent years, with the development of electrical exploration techniques and monitoring equipment, PEM technology, one of the direct current methods, had been widely applied and popularized in the evaluation of surrounding rock deformation and failure, groundwater seepage monitoring, grouting detection, and effect evaluation, water-rich abnormal area identification, etc. Compared with the other direct current methods, PEM has huge advantages of boasting faster data collection with parallel, efficient, and massive data collection and processing (Sun et al. 2021; Bharti et al. 2022).

The basic principle of PEM is that the electrodes are placed inside the borehole at intervals of 1 m and the initial apparent resistivity of overlying strata is determined before coal mining. During coal mining, the resistivity change of the overlying strata around the borehole due to deformation and fracture can be reflected by electrodes located on the drilling wall. In this process of coal seam mining, on the one hand, many empty cavities form inside overlying strata after fracture and collapse of the roof due to redistribution of stress, increasing the apparent resistivity of the surrounding rock strata. On the other hand, for the water-rich area, these cavities provide a channel for water, causing the water to flow freely and the corresponding apparent resistivity to reduce because water has a strong conductive effect. So, it is feasible through monitoring the change in rock mass apparent resistivity to measure the range that is affected by the coal mining.

Compared with other prediction methods, the artificial neural network has the characteristics of nonlinear mapping and self-learning, and can recognize patterns in complex data by simulating the structure and operational mechanism of the human neural system. The artificial neural network has been successfully applied to wind-speed forecasting (Liu et al. 2012), predicting unconfined compressive strength of rocks (Momeni et al. 2015), stability of rubble mound breakwaters (Gedik 2018), predicting floor failure depth under deep seam mining (Hu et al. 2019) and data gap filling (Silva et al. 2018). The back propagation neural network and radial basis function (RBF) neural network are probably the most widely used models. In the identification of neural network systems, the main difference between the back propagation and RBF neural network is how the excitation function is used. The excitation function of the back propagation neural network is global, and the excitation function of the RBF neural network is local. For the global excitation function, all the weights will be constantly corrected in each calculation process, resulting in a large amount of calculation and slow convergence speed, and it easily falls into a local minimum due to its own defects. In contrast, the convergence rate of the RBF neural network is significantly better than that of the back propagation neural network, and the problem of falling into a local minimum is avoided.

In this equation, x_i, x_min, x_max are the ith input parameter of the network, minimum input, and maximum input, respectively.

The regularized values of the four WFFZ height impact factors were used as the input layer, and the quantified result values were the output layer. Based on the proportion 1:6, the network model was trained and verified by normalizing 33 groups of data and then tested with six groups of data. The validation samples were regularized and treated as input values for the neural network; the corresponding output values were then obtained through the trained network. The actual output values were then compared with the values of the goal output. If the errors of the computations are within the permissible range, the constructed network is then acceptable. Otherwise, the network has to be retrained until this requirement has been satisfied.

In the training phase, the values of NS, MAE, and MAPE are 0.971, 1.961, and 3.326 m (Figure 8), respectively, which indicated that the RBF model was highly precise. This treatment is also true in the testing phase, with NS, MAE and MAPE at 0.908, 3.536, and 5.237 m (Figure 8). Therefore, it is practicable to predict the WFFZ height in deep coal excavation through this RBF model.

Another discovery is about the shape of the WFFZ. Limited by traditional monitoring methods, previous studies have focused on the final shape of the WFFZ rather than the evolution process. According to previous research results, the final shape of the WFFZ presented several patterns, mainly consisting of the ‘rectangle shape’, ‘arched shape’ and ‘saddle shape’ (Huang et al. 2018). In this study, the evolution process of WFFZ shape has been proved. The lateral boundary of the WFFZ had undergone many changes in the process of mining. The initial shape of the WFFZ lateral boundary was near saddle-shaped vertically when DB was 60 m (Figure 10(c)). Then, the neck part of the boundary gradually disappeared with the development of the WFFZ (Figure 10(d)). Next, the shape of the WFFZ lateral boundary was near half-saddle shape when DB was 33.5 m (Figure 10(e)). Finally, the lateral boundary of the WFFZ tended to be U-shaped (Figure 10(f)). The main reasons for the shape changes are the inhomogeneous structure of strata and that the rock destruction apparently lags behind the advanced distance of the working face, especially for the hard stratum.

The height of the WFFZ in this study was predicted using the RBF neural network model. The mining thickness is 6.4 m, the inclined LC is 160 m, the mining depth is 402.96 m, and the proportional coefficient of hard rock is 0.6; then, the RBF neural network model gives a predicted height of the WFFZ of No. 839 coalface of 67.32 m.

When the excavation reached 70–85 m, the height of the WFFZ increased in the advanced direction but changed less in the vertical direction (Figure 11(e)), and after the excavation was 85 m, the height of the WFFZ tended to stabilize, as shown in Figure 11. The height of the WFFZ was 65.36 m when the excavation was 100 m (Figure 11(f)).

The comparison between the predictions from the four models is shown in Table 3. The table shows that the predicted values obtained by the RBF neural network model and by numerical simulation coincided with an observed result by PEM, with a relative error of 0.27% and 3.17%, respectively. In comparison, the results from the RBF neural network model are seen to be closer to the measured values than the results from the calculation based on numerical simulation or empirical formula, with less error and higher precision, which meets practical needs.

The table also demonstrates a significant difference between the predicted result by empirical formula and the observed result by PEM, with a relative error of 25.3%–41.6%. It proves that the predicted results of the existing empirical formula cannot effectively predict the height of the WFFZ. Also, the inapplicability of the empirical formula is attributed to the limited mining technology, outdated measured data, different geological conditions, and coal seam characteristics (Dai et al. 2020).

In this study, the development process and height of the mining-induced water fractured zone were researched by PEM technology, neural network model, numerical simulation model and the empirical formula methods. The major results are:

• The results by PEM and numerical simulation have shown that the WFFZ did not develop upwards simply layer-by-layer from the bottom up. Rather, the fractured areas appear first in soft layers above the goaf. With the advance of footage, the destruction gradually extends from soft layers to hard layers, and the WFFZ is formed when the fractured areas are connected with each other. Also, the change process of the WFFZ shape is closely related to rock mass heterogeneity.

• Providing accurate influencing factors is a key factor for building the RBF neural network model. Besides the mining thickness, mining depth, and inclined LC, the proportion coefficient of hard rock was introduced to quantify the overburden structure. Finally, an effective RBF model was established through learning and training based on the 39 collected groups of measured data.

• Based on the detection results by PEM, the rationality of the RBF model was proved. Also, the RBF model is a more effective method to predict the height of the WFFZ in comparison with numerical simulation and the traditional empirical formula method.

Financial support for this work is provided by the State Key Program of the National Natural Science Foundation of China under Grant No. 41931284 and the Natural Science Foundation of Jiangsu Province under Grant BK20190646.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.