2.1. General Approach

The core of the dual-level prediction method for roof pressure in fully mechanized mining combines statistical prediction and deep learning prediction, as illustrated in Figure 1.

The statistical prediction method involves analyzing recent historical roof pressure data to determine the average and standard deviation of key metrics, including pressure intervals, timing, and intensity. This method allows for early predictions, typically one cycle in advance, but the accuracy is highly affected by geological conditions. For simpler geological conditions and more consistent pressure patterns, the predictions are fairly accurate. In contrast, for complex geological conditions and variable pressure patterns, the prediction range is generally broader. Therefore, statistical prediction can be used for early, rough forecasting of the next pressure event.

The deep learning prediction method involves analyzing historical roof pressure data to build and train a roof pressure prediction model using deep learning algorithms. This model can predict future pressure trends, determine the timing and location of the next roof pressure event, and trigger a warning if the prediction meets the set pressure alert criteria. This method offers accurate predictions but with limited lead time; the closer it is to the actual pressure event, the more accurate the prediction. Therefore, deep learning prediction is suitable for real-time, precise forecasting of upcoming pressure events. Combining these two methods provides complementary advantages: early lead time from statistical prediction and timeliness and accuracy from deep learning.

2.2. Statistical Prediction Method

2.2.2. Steps in the Method

Step 1: Calculate the sum of the mean and standard deviation of the cycle end resistance as the indicator for roof pressure determination. Based on this indicator, identify the historical pressure locations and timings for each support.

First, it is necessary to calculate the average value of the cycle end resistance (P_t). Assuming there are n sets of cycle end resistance data (P_t1, P_t2, …, P_tn), the average value of the cycle end resistance can be calculated using the following formula:

$$ ()

where P_ti is the end resistance value of the i^th cycle.

Next, calculate the standard deviation of the cycle end resistance (σ_p):

$$ ()

where n is the number of cycles, P_ti is the end resistance value of each cycle, and P_t is the mean cycle end resistance.

Finally, the sum of the mean cycle end resistance and its standard deviation is used as the roof pressure determination indicator (P^′) [18]:

$$ ()

If the end resistance value (P_t) of a cycle exceeds this indicator (P’), then the timing and location of roof pressure in that cycle can be determined.

Step 2: Calculate the average interval between pressures for each support over the past 10 days or 10 cycles, removing any outliers. The prediction formula for the next pressure location is as follows:

$$ ()

where M_i represents the interval for the i^th pressure event (m); W_z is the most recent pressure location (m); W_l is the predicted location of the next event (m).

Step 3: Following the above method, calculate the historical pressure intervals for all supports over the past 10 days or 10 cycles.

Step 4: The predicted range for the next pressure interval is determined by adding and subtracting the standard deviation of the average pressure interval. Adding this range to the most recent pressure location provides the predicted location for the next pressure event, as illustrated in Figure 2. In Figure 2, X_1–X₃ denote measured periodic weighting intervals, with X₄ representing the corresponding predicted value; L_1–L₄ indicate the measured weighting influence ranges, while L₅ signifies the predicted weighting range; Y_1–Y₄ characterize the measured support loading ranges under periodic weighting effects, and Y₅ corresponds to the predicted support loading range.

2.3. Deep Learning Prediction Method

2.3.2. Steps in the Method

After training the deep learning model on extensive roof pressure data, the model can automatically, continuously, and in real-time predict roof changes over a future period, with predictions showing important information such as future pressure values, as illustrated in Figure 3.

Step 1: Data preprocessing: Collect historical roof pressure (mining pressure) data from each support over the past 2 months. Preprocess the data to reduce noise and anomalies, enhancing data quality, and store it in an InfluxDB database for deep learning model training. Data preprocessing and smoothing results after noise reduction are shown in Figure 4.

Step 2: The deep learning model was trained based on historical ground pressure data from the working face, with model performance enhanced through hyperparameter optimization. The training process commenced with learning rate adjustment, where the learning rate was increased when observing slow convergence of training loss, while decreased in cases of loss oscillation. To accelerate the training process, the batch size was carefully controlled. Regarding the LSTM architecture, the optimization strategy initiated with a single-layer LSTM, followed by performance evaluation on the validation set. Subsequent layers were incrementally added based on empirical results, ultimately yielding a model capable of capturing periodic variations, trend patterns, and latent influencing factors.

Step 3: Periodic prediction of roof pressure: For each support on the coal mining face, use the past 48 h of pressure data as input to the LSTM model, which predicts the pressure trend for the next 6 h. The model’s predictions are updated every half hour in a continuous loop to ensure accuracy. The predicted results for each support are shown in Figure 5.

When pressure on supports in a certain region of the coal mining face increases, the overall pressure forms an arch-like distribution, indicating the presence of a “pressure arch.” Analyzing the pressure cloud chart for the face, as shown in Figure 6, can help determine if a “pressure arch” has formed as a pressure warning indicator. When the number of over-pressured supports in the arch region exceeds a threshold, a pressure warning is issued, and an alert is sent via software message, email, or other methods to inform relevant personnel of the impending pressure event, allowing timely response. It is important to highlight that this function requires effective filtering of false positives caused by anomalies in some supports to avoid unnecessary labor.

3.1. Software Function Design

The software platform consists of five main functional modules: data collection and processing module, predictive analysis module, visualization module, alert and warning module, and decision support module.

3.2. System Architecture Design

The software platform is designed with a distributed architecture, comprising data collection, server, and user components, as shown in Figure 7. The data collection component, installed on hydraulic supports underground, is responsible for real-time data collection and uploading to the server. The server component consolidates the data, stores it, performs analysis, runs predictive algorithms, and manages the database. The user component interacts with the server through a web interface to retrieve real-time monitoring information and prediction results for roof pressure.

3.3. Data Management and Processing

Due to the real-time nature and high frequency of support pressure data, the system must manage and store large volumes of data. We chose InfluxDB as the data storage solution, which is a high-performance time-series database suitable for storing and querying high-frequency sensor data [22].

The historical rock pressure data from each support in the working face over the past 2 months were collected and preprocessed. The local outlier factor (LOF) algorithm was applied to detect outliers, with clearly erroneous records removed and extreme outliers truncated. Additionally, linear interpolation was used to fill in missing values within a reasonable range, ensuring data continuity and integrity. Moreover, moving average filtering was employed to reduce noise interference and improve data quality. The processed data were then utilized for the training and optimization of the deep learning model.

3.4. Prediction Algorithm Implementation

The platform’s prediction algorithms include both statistical methods and a LSTM-based deep learning model. The statistical method analyzes historical data to predict future pressure intervals and values, while the LSTM model is trained on historical data to forecast support pressure over the next 6 h.

The statistical algorithm uses historical pressure and advancement data to predict the next pressure interval and advancement position, following the steps illustrated in Figure 8.

The deep learning algorithm learns and remembers long-term trend changes in historical roof pressure data, enabling it to predict future pressure values for each support. Each prediction uses the past 48 h of roof pressure data, covering at least one complete pressure cycle, allowing the machine learning algorithm to capture periodic variations in roof pressure data. The algorithm predicts each support’s pressure trend over the next 6 h. The specific process is illustrated in Figure 9.

For large-scale data from multiple mining faces or sensors, the software will focus on data preprocessing, utilize statistical models for preliminary analysis, and employ deep learning models to handle complex nonlinear relationships. The models will be updated in real-time to enhance the efficiency of large-scale data processing. Ultimately, the software aims to achieve efficient, real-time, and scalable data processing and predictive analysis.

3.5. User Interface Design

The platform’s user interface is designed with a simple and clear graphical layout, allowing users to observe the pressure trend of each support, upcoming pressure predictions, and alert notifications issued by the system in real time. The core principle of the user interface design is to enhance usability and intuitive information display, ensuring that on-site management personnel can quickly grasp pressure information. The system primarily consists of six main functions: stress analysis, stress inquiry, monitoring alarm, log management, organization management, and system monitoring. Figure 10 shows the user interface design.

4.1. Prediction Results of the Statistical Method

Using the 15,219 coal mining face at Hongliulin Coal Mine as an example, roof pressure data from July 1 to July 31, 2024, was collected and analyzed. A total of 20 pressure events were recorded on this coal mining face, as detailed in Table 1. After removing abnormal data, the actual periodic pressure interval on the 15,219 coal mining face ranged from 12.3 to 24.25 m, with an average pressure interval of 17.15 m. Based on the statistical prediction method, the forecasted pressure interval ranged from 12.04 to 24.48 m, with an average predicted distance of 17.29 m. The interval range error was 0.26 m, and the average interval error was 0.14 m.

Based on this data, we further classified and analyzed the prediction errors. Among the predictions, 10 cases had errors within the range of 0–0.8 m, accounting for 50% of the total predictions. Predictions with errors between 0.8 and 1.6 m occurred in five cases, accounting for 25% of the total. Predictions with errors greater than 1.6 m also accounted for 25%, with five cases. We consider an error less than 1.6 m, equivalent to the distance of two mining cycles, to be acceptable, reaching a satisfactory accuracy rate of 75%.

4.2. Statistical Analysis of Pressure Prediction Based on Deep Learning

Similarly, using the 15,219 coal mining face at Hongliulin Coal Mine, roof pressure data from July 1 to July 31, 2024, was analyzed. A total of 20 pressure events were recorded on this coal mining face, as detailed in Table 2. After removing abnormal data, the actual periodic pressure interval on this face ranged from 12.3 to 24.25 m, with an average distance of 17.15 m. The deep learning-based prediction gave an interval range from 12.5 to 24.05 m, with an average distance of 17.04 m. The interval range error was 0.2 m, and the average interval error was 0.11 m.

From further classification of this data, we observed 10 cases with prediction errors in the range of 0–0.8 m, making up 50% of the total predictions. Predictions with errors between 0.8 and 1.6 m occurred in eight cases, making up 40%. Predictions with errors greater than 1.6 m accounted for 10%, with two cases. Predictions with errors less than 1.6 m made up 90% of the cases.

According to the statistical results, the model achieves a prediction accuracy of 85.71%, a recall rate of 90%, and an F1 score of 87.95%. The high F1 score indicates excellent model performance, demonstrating its effectiveness and suitability for rock pressure prediction in this specific scenario.