Coal spontaneous combustion incubation period CO concentration and coal temperature prediction method

Abstract

The present application relates to the technical field of coal spontaneous combustion monitoring and early warning, in particular to a kind of coal spontaneous combustion incubation period CO concentration and coal temperature prediction method, to solve the problems such as insufficient accuracy and prediction difficulty in the prediction of coal spontaneous combustion incubation period, by the improved POA algorithm introducing dynamic convergence coefficient, memory pool, mechanism and mixed search strategy, effectively accelerate the convergence process of model, reduce the reasoning time, and enhance the robustness of model, the present application is combined with LSTM neural network, the combination of model hyperparameters is optimized, the prediction accuracy is improved, and scientific basis can be provided for accurate prediction of coal spontaneous combustion disaster.

Description

Technical Field

[0001] This invention relates to the field of coal spontaneous combustion monitoring and early warning technology, and in particular to a method for predicting CO concentration and coal temperature during the coal spontaneous combustion latency period. Background Technology

[0002] Coal, as one of the world's most important energy sources, plays an irreplaceable role in industrial production, transportation, and daily life. As a crucial energy raw material, the mining, transportation, storage, and utilization of coal are vital to the stability of energy supply. However, coal mining and use also face numerous safety and environmental challenges. In particular, spontaneous combustion of coal is a frequent occurrence during mine production, potentially leading to gas explosions, dust explosions, and other serious safety accidents, threatening mine production safety, environmental protection, and the safety of personnel. To effectively prevent the frequent occurrence of gas explosions and dust explosions caused by spontaneous combustion of coal, the industry has proposed numerous monitoring and early warning methods, among which gas testing and temperature testing are the most widely used. Gas testing uses abnormalities in the volume fraction of single or complex gases to warn of the risk of spontaneous combustion of coal, while temperature testing predicts hazards based on the heat release characteristics generated during the heating process of spontaneous combustion of coal. However, the incubation period of spontaneous combustion of coal is usually long, and there are no obvious visible symptoms in the early stages. Therefore, relying on traditional monitoring methods is difficult to achieve timely and effective early warning, especially in the absence of obvious gas anomalies, where the potential fire risk in coal mines is easily overlooked.

[0003] In the early stages of coal spontaneous combustion, changes in CO concentration and coal temperature are the most sensitive indicators, reflecting the initial signals of spontaneous combustion. CO, in particular, as a crucial gas in the spontaneous combustion process, shows that increases in its concentration are closely related to the pyrolysis and oxidation reactions of coal. Changes in coal temperature are directly linked to the heat release process of spontaneous combustion. Therefore, real-time monitoring and prediction of CO concentration and coal temperature help identify the risk of spontaneous combustion early, providing timely warnings to coal mines and preventing catastrophic safety accidents.

[0004] With the rapid development of big data and machine learning technologies, their application to the prediction and early warning of coal spontaneous combustion has become a trend. Machine learning methods can deeply explore the nonlinear relationships in monitoring data such as gas concentration and temperature, providing more accurate predictive capabilities than traditional methods. This is especially true when the latency period for coal spontaneous combustion is long and early warning is difficult, effectively improving the accuracy of early warnings. Combined with traditional monitoring methods, machine learning provides strong technical support for the accurate prediction and risk warning of coal spontaneous combustion disasters.

[0005] Although the theoretical research on coal spontaneous combustion prediction and forecasting is quite mature, the research on predicting the gas and coal temperature in the coal spontaneous combustion process using machine learning and big data technology is still in its early stages. For example, the monitoring error is relatively large: sensors are prone to calibration errors and drift due to environmental influences during long-term use, and even data loss may occur. Existing sensors often lack sensitivity when there are low concentrations of gas or slight temperature changes, which affects the accuracy and reliability of the data. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention proposes a method for predicting CO concentration and coal temperature during the coal spontaneous combustion latency period. This method aims to address the challenges of insufficient accuracy and difficulty in predicting the latency period of coal spontaneous combustion. By introducing a dynamic convergence coefficient, memory pool, mechanism, and hybrid search strategy through an improved POA algorithm, the convergence process of the model is effectively accelerated, inference time is reduced, and the robustness of the model is enhanced. Furthermore, by combining this method with an LSTM neural network, the hyperparameter combination of the model is optimized, improving prediction accuracy and providing a scientific basis for the accurate prediction of coal spontaneous combustion disasters.

[0007] The technical solution adopted by this invention to solve its technical problem is as follows: A method for predicting CO concentration and coal temperature during the spontaneous combustion latency period of coal, comprising the following steps:

[0008] S1: Import historical data on spontaneous combustion of coal in the goaf of the mine into the prediction system to construct a time series dataset of CO concentration, coal temperature and preset indicators;

[0009] S2: Perform data preprocessing on the dataset, including missing value imputation and data normalization, to improve model accuracy;

[0010] S3: Use the Spearman correlation coefficient method to determine the correlation between preset indicators, and select features based on the magnitude of the correlation coefficient.

[0011] S4: Perform feature importance analysis on the feature selection metrics of the model, and use the SHAP algorithm to calculate the Shapley value of the features to obtain different metrics;

[0012] S5: The preprocessed dataset is divided into training, validation, and test sets in an 8:1:1 ratio. An LSTM neural network model is constructed, and the model parameters are initialized.

[0013] S6: Optimize the LSTM model parameters using the improved POA algorithm. When the improved POA-LSTM optimization model meets the preset conditions, select the hyperparameter configuration.

[0014] S7: Construct an LSTM model to predict CO concentration and coal temperature based on the optimal parameter combination of the output, and retrain and test the dataset;

[0015] S8: Train and build LSTM, WOA-LSTM, PSO-LSTM, POA-LSTM and IPOA-LSTM algorithms in sequence. By comparing evaluation metrics, judge the prediction results and prediction accuracy, and verify the advantages of the IPOA-LSTM model.

[0016] Specifically, step S2 includes:

[0017] S21: Missing value imputation. The imputation method uses linear interpolation. In data preprocessing, linear interpolation assumes that the missing data point changes linearly between its preceding and following known data points. There are two known points... and Estimate a point located between these two points. value at The formula for linear interpolation is:

[0018] ;

[0019] in: These are the points that need interpolation. and These are the values ​​of the known data points. and These are the independent variable values ​​of the known data points;

[0020] S22: Outlier Handling, wherein the IQR-based outlier handling method detects and handles outliers during the data preprocessing stage; this includes calculating the quartiles of the data and using IQR to determine whether outliers exist in the data, wherein the calculation steps are as follows:

[0021] Calculate the first quartile (Q1, 25th quartile) and the third quartile (Q3, 75th quartile) of the data;

[0022] Calculate IQR = Q3 - Q1, where IQR represents the distribution range of the middle 50% of the data in the dataset;

[0023] The low and high thresholds for outliers are set as follows: Lower Bound = Q1 - 1.5 × IQR, Lower Bound = Q3 + 1.5 × IQR. Values ​​below the low threshold or above the high threshold are considered outliers.

[0024] S23: Data normalization, used to reduce the range of model parameter updates during training, including the following calculation formula:

[0025] ;

[0026] in, It is the raw data. and These are the minimum and maximum values ​​of the data, respectively. This is the normalized data.

[0027] Specifically, step S3 includes:

[0028] S31: Collect multi-dimensional data related to spontaneous combustion of coal, including CO concentration, coal temperature, carbon dioxide concentration, oxygen concentration, methane concentration, and olefin concentration.

[0029] S32: Set up a data matrix, where each row represents a sample and each column represents an indicator or target variable, where the target variable is the risk of spontaneous combustion of coal or the spontaneous combustion index;

[0030] S33: Use the Spearman correlation coefficient to calculate the correlation between each indicator and the risk of spontaneous combustion of coal, and quantitatively assess the linear association between the relevant indicators; wherein, the Spearman correlation coefficient between each indicator is calculated using the following formula:

[0031] ;

[0032] in: It is the difference in the ranking of each pair of variables. It is the sample size. It is the Spearman correlation coefficient;

[0033] S34: Generating a heatmap: The heatmap function in MATLAB is used to visualize the weights of the indicators. Based on the normalized results, the Spearman correlation coefficient between each indicator is obtained.

[0034] Specifically, step S4 includes:

[0035] S41: Quantify the contribution of each feature to a single prediction;

[0036] S42: Given a prediction, calculate the marginal contribution of each feature using the Shapley value, including calculations based on all possible feature subsets, assuming there are... The Shapley value is calculated using the following formula for each feature:

[0037] ;

[0038] in, Represents the set of all features. Indicates that features are not included. of a subset of Indicates the number of elements in the set. Indicates a subset of features and characteristics Time sample points The predicted value, Indicates a subset of features Time sample points The predicted value, Measured in the control feature subset At that time, characteristics The magnitude of the impact on the predicted value, sample points Features The SHAP value is for all possible subsets of features. Lower features The effects are weighted and summed;

[0039] S43: Obtain the importance of different predictors in the prediction model by calculating feature importance. Mean absolute SHAP value across all sample points ;

[0040] ;

[0041] in, Represents the sample set, express The absolute value, Represents the total number of samples. That is, characteristics The overall feature importance in the prediction model The larger the value, the more characteristic it indicates. The more important it is.

[0042] Specifically, step S5 includes:

[0043] S51: Input layer dimension;

[0044] S52: Output layer dimension;

[0045] S53: Hidden layer dimension;

[0046] S54: Activation function

[0047] S55: Number of neurons in the hidden layer

[0048] S56: Input layer time steps

[0049] S57: Learning rate;

[0050] S58: Number of iterations.

[0051] Specifically, step S6 includes:

[0052] S61: Run the POA algorithm, calculate the objective function value, randomly generate prey, update the position during the search phase, and update the position to save the optimal solution during the hunting phase, until the number of iterations is reached;

[0053] This includes population initialization, used to generate the initial population, and the random value rand ensures that the initial positions of the pelicans are evenly distributed between the upper and lower bounds of the solution, covering the entire solution space;

[0054] ;

[0055] in, It is the first The Pelicans' first Position in dimensions; rand It is a random number, ranging from [0,1]. Indicates the first In dimension, the upper boundary of the problem, The lower boundary; This refers to the pelican population size. The dimensions of the problem;

[0056] S62: Matrixing of Pelican Populations in POA This indicates that the current solution for all pelicans is recorded for subsequent iterations, updates, and optimizations.

[0057] ;

[0058] in, It is the first The Pelicans' position It is a matrix representation of the pelican population;

[0059] S63: The objective function value vector is in the form of a vector used to measure the quality of the solution for each pelican. Each individual pelican represents a set of LSTM parameters, which are optimized by an optimization algorithm to find the optimal solution.

[0060] ;

[0061] in, It is the first The objective function value for each pelican. It is the objective function value vector of the entire pelican population;

[0062] S64: The search phase, i.e., the way the pelican approaches its prey, is precisely modeled.

[0063] S65: Hunting phase;

[0064] S66: The improved POA algorithm incorporates dynamic convergence coefficients, a memory pool mechanism, and a hybrid search strategy into the Pelican optimization algorithm.

[0065] The beneficial effects of this invention are as follows:

[0066] The present invention proposes a method for predicting CO concentration and coal temperature during the coal spontaneous combustion latency period. This method is based on the joint prediction of CO concentration and coal temperature by multiple factors to provide early warning of the coal spontaneous combustion process. It has high accuracy and stability in predicting CO concentration and coal temperature, can identify the risk of coal spontaneous combustion in advance, and improves the accuracy of predicting the coal spontaneous combustion latency period. Attached Figure Description

[0067] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0068] Figure 1 Flowchart for improving the POA algorithm and LSTM neural network model prediction;

[0069] Figure 2 This is a diagram of the architecture of a coal mine bundled tube monitoring system.

[0070] Figure 3 The original data plot in the visualization of CO concentration data;

[0071] Figure 4 A data plot for handling missing and outlier values ​​in CO concentration data visualization;

[0072] Figure 5 Normalized data plots for CO concentration data visualization;

[0073] Figure 6 A correlation coefficient graph for preset indicators;

[0074] Figure 7 This is a CO concentration prediction plot from the SHAP global interpretability feature summary.

[0075] Figure 8 Temperature prediction plot in the SHAP global interpretability feature summary;

[0076] Figure 9 Ordination plot of the average influence of the input variable SHAP;

[0077] Figure 10 Here is a diagram of the LSTM network structure;

[0078] Figure 11 Training curves (step state) for the IPOA-LSTM model.

[0079] Figure 12 Compare the predicted CO concentration curves with the model prediction curves;

[0080] Figure 13 The temperature prediction plot is compared with the model prediction curve;

[0081] Figure 14 Evaluation map of CO concentration predicted by different models in radar chart;

[0082] Figure 15 Evaluation charts of coal temperature predicted by different models are generated from radar charts for different model predictions. Detailed Implementation

[0083] To make the technical means, creative features, objectives and effects of this invention easier to understand, the implementation and application process of the method for predicting CO concentration and coal temperature during the spontaneous combustion period of coal proposed in this invention will be further elaborated and explained below in conjunction with specific embodiments.

[0084] like Figures 1 to 15 As shown, the implementation process of this invention is as follows:

[0085] S1: Import historical data on spontaneous combustion of coal in the goaf of the mine into the prediction system to construct a time series dataset of CO concentration, coal temperature and preset indicators (oxygen concentration, olefin concentration, etc.). The dataset is derived from 507 publicly available coal mine experimental data, including data on CO concentration, coal temperature, oxygen concentration, C2H4 concentration, C2H4 / C2H6, etc. The corresponding program modules are developed using the MATLAB platform to analyze and process the data, providing reliable data support for the prediction and prevention of spontaneous combustion risk of coal.

[0086] S2: Perform data preprocessing on the dataset, including missing value imputation and data normalization, to improve model accuracy;

[0087] In the specific implementation of step S2 of the present invention, the following are included:

[0088] S21: Missing value imputation. The imputation method uses linear interpolation. In data preprocessing, linear interpolation assumes that the missing data point changes linearly between its preceding and following known data points. There are two known points... and Estimate a point located between these two points. value at ， The formula for linear interpolation is:

[0089] ;

[0090] in: These are the points that need interpolation. and These are the values ​​of the known data points. and These are the independent variable values ​​of the known data points;

[0091] S22: Outlier Handling, wherein the IQR-based outlier handling method detects and handles outliers during the data preprocessing stage; this includes calculating the quartiles of the data and using IQR to determine whether outliers exist in the data, wherein the calculation steps are as follows:

[0092] Calculate the first quartile (Q1, 25th quartile) and the third quartile (Q3, 75th quartile) of the data;

[0093] Calculate IQR = Q3 - Q1, where IQR represents the distribution range of the middle 50% of the data in the dataset;

[0094] The low and high thresholds for outliers are set as follows: Lower Bound = Q1 - 1.5 × IQR, Lower Bound = Q3 + 1.5 × IQR. Values ​​below the low threshold or above the high threshold are considered outliers.

[0095] S23: Data normalization, used to reduce the range of model parameter updates during training, including the following calculation formula:

[0096] ;

[0097] in, It is the raw data. and These are the minimum and maximum values ​​of the data, respectively. Normalized data is used to represent the normalized data. Machine learning algorithms are very sensitive to the scale of input data. If the scale difference of the input data is too large, the model may experience gradient vanishing or gradient exploding during training, which will affect the learning process, slow down the model's convergence speed, or even prevent it from converging. By normalizing, the range of model parameter updates during training is reduced, making the training process more stable. Usually, the values ​​of all feature data are scaled to a uniform range [0,1], which can effectively avoid this situation. In addition, during the inference stage, normalized data allows the model to make predictions in a shorter time, supporting real-time monitoring and early warning.

[0098] Taking CO concentration data as an example, linear interpolation was used to fill in missing values. The processed data showed the same trend as the original data, greatly reducing data noise. For example, CO concentration data... Figure 3-5 As shown.

[0099] S3: Use the Spearman correlation coefficient method to determine the correlation between preset indicators, and select features based on the magnitude of the correlation coefficient. Because the coal spontaneous combustion time series dataset contains many indicators that are weakly correlated with coal spontaneous combustion, we choose to use the Spearman correlation coefficient method to determine the correlation between preset indicators, select features based on the magnitude of the correlation coefficient, and remove indicators with weak correlation so that the model training only includes the most important features, thereby optimizing the model structure and making it easier to extract the patterns of time series data and improve the accuracy and stability of prediction.

[0100] In the specific implementation of step S3 of the present invention, the following are included:

[0101] S31: Collect multi-dimensional data related to spontaneous combustion of coal, including but not limited to CO concentration, coal temperature, carbon dioxide concentration, oxygen concentration, methane concentration, olefin concentration, etc.

[0102] S32: Set up a data matrix, where each row represents a sample and each column represents an indicator or target variable. The target variable is usually the risk of spontaneous combustion of coal or the spontaneous combustion index.

[0103] S33: Use the Spearman correlation coefficient to calculate the correlation between each indicator and the risk of spontaneous combustion of coal, and quantitatively assess the linear association between the relevant indicators; wherein, the Spearman correlation coefficient between each indicator is calculated using the following formula:

[0104] ;

[0105] in: It is the difference in the ranking of each pair of variables. It is the sample size. It is the Spearman correlation coefficient;

[0106] S34: Generate a heatmap: The heatmap function in MATLAB is used to visualize the weights of the indicators. Based on the normalized results, the Spearman correlation coefficient between each indicator is obtained.

[0107] The Spearman correlation coefficient method was used to compare the correlation coefficients between CO concentration and experimental temperature and other preset indicators. The results of the correlation coefficient calculation are shown in Figure 6.

[0108] CO, CO2, CO / O2, CO2 / O2, and C2H4 / C2H6 are positively correlated with the experimental temperature, while O2 is negatively correlated with the experimental temperature. The reason for this may be that as the experimental temperature increases, the redox reaction is intensified, consuming a large amount of oxygen and producing a large amount of CO. At the same time, the high temperature promotes the oxidation of organic matter in coal to generate more carbon oxides.

[0109] The experimental temperature, CO2, C2H4, CO / O2, CO2 / O2, CH4, C2H4, and C2H6 were positively correlated with CO concentration, while O2 was negatively correlated with CO concentration. Temperature and pyrolysis gases had a significant impact on CO, mainly because higher temperatures lead to faster coal oxidation rates and higher CO concentrations. At the same time, a large amount of pyrolysis gases, which are byproducts of CO, are also produced. O2 acts as an inhibitory factor because incomplete combustion is more likely to occur in low-oxygen environments, producing a large amount of CO and thus increasing CO concentration.

[0110] S4: Perform feature importance analysis on the feature selection metrics of the model, and use the SHAP (SHapley Additive ex Planations) algorithm to calculate the Shapley value of the features to obtain different metrics;

[0111] The importance of a feature in a predictive model is determined by the weight of that feature as an input to the machine learning model, using the mean absolute SHAP value of the feature as a relevant metric.

[0112] In the specific implementation of step S4 of the present invention, the following are included:

[0113] S41: Quantifying the contribution of each feature in a single prediction; SHAP is a game theory-based model interpretation method used to quantify the contribution of each feature in a single prediction. It draws on the idea of ​​Shapley values, which are mainly used to allocate the contributions of each participant in a cooperative game. The core idea of ​​the SHAP method is to obtain the importance of different indicators in the prediction model by calculating the marginal contribution of each feature to the prediction result.

[0114] S42: Given a prediction model, calculate the marginal contribution of each feature using the Shapley value, including calculations based on all possible feature subsets, assuming there are... The Shapley value is calculated using the following formula for each feature:

[0115] ;

[0116] in, Represents the set of all features. Indicates that features are not included. of a subset of Indicates the number of elements in the set. Indicates a subset of features and characteristics Time sample points The predicted value, Indicates a subset of features Time sample points The predicted value, Measured in the control feature subset At that time, characteristics The magnitude of the impact on the predicted value, sample points Features The SHAP value is for all possible subsets of features. Lower features The effects are weighted and summed;

[0117] like Figure 7-8 As shown in the global interpretability feature summary diagram, each point represents the influence of a sample's feature value on the prediction. The color represents the magnitude of the feature value, with red indicating high and blue indicating low. The magnitude represents the influence of each feature variable on the prediction of CO concentration and coal temperature.

[0118] Regarding the degree of influence on CO concentration prediction, the coal temperature at the top has the widest range of SHAP values, indicating that it has the most significant impact on CO concentration prediction. In addition, the O2 concentration, CO2 concentration, C2H4, CH4, and C2H6 values ​​also have relatively high SHAP values, indicating that carbon oxides and pyrolysis gases have a significant impact on CO concentration prediction. Finally, the CO / O2, CO2 / O2, and C2H4 / C2H6 values ​​are relatively small, indicating that the Graham coefficient and the ratio of pyrolysis gases have almost no impact on CO concentration prediction.

[0119] In predicting coal temperature, CO concentration had the largest SHAP value, indicating its most significant impact on coal temperature prediction. This was followed by O2, CO2, CO / O2, and CO2 / O2, which also had relatively large SHAP values, suggesting that carbon oxides and the Graham coefficient had a significant impact on the model output. A clear negative correlation existed between O2 concentration and SHAP value, indicating that increasing O2 concentration inhibited the rise in coal temperature and also affected coal temperature prediction. Finally, C2H6, C2H4, C2H4 / C2H6, and CH4 had relatively small values, indicating that pyrolysis gases had a smaller impact on CO concentration prediction.

[0120] S43: Obtain the importance of different predictors in the prediction model by calculating feature importance. Mean absolute SHAP value across all sample points ;

[0121] ;

[0122] in, Represents the sample set, express The absolute value, Represents the total number of samples. That is, characteristics The overall feature importance in the prediction model The larger the value, the more characteristic it indicates. The more important;

[0123] Significance analysis of indicators in a predictive model can help determine the degree of influence of each predictor on the outcome. Therefore, it can be calculated using the steps described above. By using the SHAP algorithm, the contribution of each input variable to the model's prediction can be quantified, revealing the internal correlation logic of the model and improving its interpretability. Figure 12-13 The contribution of each feature variable to the prediction result is clearly shown in the figure, with the mean absolute SHAP value as an example. The weights of these indicators can be used as input features for prediction models. By optimizing the weight ratios of these indicators, the problems of class imbalance and noisy data can be solved, making the model focus more on key information, improving the recognition accuracy of a few classes and the performance of the model, and increasing the reliability of CO concentration and coal temperature predictions.

[0124] S5: After completing the data preprocessing, the dataset is divided into training, validation, and test sets in an 8:1:1 ratio. An LSTM neural network model is constructed, and the model parameters are initialized. The core structure of LSTM consists of three main parts: the Forget Gate, the Input Gate, and the Output Gate. These gates control the flow and retention of information, thereby effectively preserving and updating the network's memory. The LSTM network structure is as follows: Figure 10 As shown;

[0125] First, the cell state is a key part of LSTM, which can be understood as the network's "long-term memory". It flows throughout the sequence and carries information. LSTM uses these gates to decide which information needs to be retained and which needs to be discarded.

[0126] Second, LSTM uses a gating mechanism to determine which information should be discarded or retained. The output of each gate is a value between 0 and 1, representing the proportion of information that should be retained or discarded. The input of each gate comes from the output of the previous layer and the input of the current time step.

[0127] Specifically, the forget gate vector of the model is represented as:

[0128] ;

[0129] in, It is the output of the forget gate. It is the Sigmoid activation function. It is the weight matrix of the forget gate. It is a concatenation of the previous implicit state and the current input. It is the bias of the forgetting gate;

[0130] The input gate vector of the model is represented as:

[0131] ;

[0132] in, Indicates the output of the input gate. It is the weight matrix of the input gate. It is the input gate bias;

[0133] The candidate cell states of the model are:

[0134] ;

[0135] in, This represents the new candidate cell state, and tanh is the tanh activation function. It is the weight matrix of cell states. It is a cell state bias;

[0136] The output gate vector of the model is represented as:

[0137] ;

[0138] in, This indicates the output of the output gate. It is the hidden state of the previous time step. This is the input for the current time step. It is the weight matrix of the output gate. It is the bias of the output gate;

[0139] Third: Cell state updating, in LSTM, Cell state updates are determined by the forgetting gate, the input gate, and the candidate cell state:

[0140] ;

[0141] in, Control the discarded information, The table controls the added information. It is a candidate cell state;

[0142] Fourth, the final output in the hidden state. It is calculated through the output gate:

[0143] ;

[0144] in, It is the implicit state at the current moment. This indicates the output of the output gate. This is the current state of the cell.

[0145] In the specific implementation of step S5 of the present invention, the following are included:

[0146] S51: Input layer dimension; For this model, multi-dimensional data related to coal spontaneous combustion will be input, including temperature, CO2, O2, C2H4, etc., that is, the input layer determines the specific input value based on the number of input indicators;

[0147] S52: Output layer dimension; This model aims to predict the spontaneous combustion process of coal, which is determined based on CO concentration and coal temperature, therefore the dimension is 2;

[0148] S53: Hidden layer dimension; Since the problem solved by this invention is a multi-input single-output problem, which is relatively simple, the hidden layer is set to 1;

[0149] S54: Activation function. For non-linear data, networks using non-linear activation functions such as ReLU perform better.

[0150] S55: Number of hidden layer neurons. The number of hidden layer neurons determines the capacity and complexity of the LSTM model. It represents the number of features that the model can learn and remember. For simple tasks, it is usually started with a small value (such as 64) and gradually increased to 128, 256, etc., to observe the training speed and performance of the model.

[0151] S56: Input layer time steps. The time step count refers to the number of consecutive time points observed by the LSTM model in each iteration. For short-term prediction tasks, a smaller time step count of 5 can be used; typically, the step size constraint is set to 5-10.

[0152] S57: Learning rate; The learning rate determines the step size of parameter adjustment during each gradient update. If the learning rate is too high, the model may oscillate during training and fail to converge; if the learning rate is too low, the convergence speed may be very slow or it may get stuck in a local optimum. Generally, a small initial learning rate is used, i.e., 0.001-0.01.

[0153] S58: Number of iterations; The number of iterations refers to the number of times the model is fully trained on the entire training set; At the end of each epoch, all data is fed into the model for one forward and one backward propagation. Generally, we start trying from 50 to 100 epochs, and combine it with early stopping. When the performance on the validation set no longer improves, we can stop training to avoid wasting computing resources.

[0154] S6: Optimize the LSTM model parameters using the improved POA algorithm. When the improved POA-LSTM optimization model meets the preset conditions, select the optimal hyperparameter configuration.

[0155] In the specific implementation of step S6 of the present invention, the following are included:

[0156] S61: Run the POA algorithm, calculate the objective function value, randomly generate prey, update the position during the search phase, and update the position to save the optimal solution during the hunting phase, eventually reaching the maximum number of iterations;

[0157] This includes population initialization, used to generate the initial population, and the random value rand ensures that the initial positions of the pelicans are evenly distributed between the upper and lower bounds of the solution, thus covering the entire solution space;

[0158] ;

[0159] in, It is the first The Pelicans' first Position in dimensions; rand It is a random number, ranging from [0,1]. Indicates the first In dimension, the upper boundary of the problem, The lower boundary; This refers to the pelican population size. The dimensions of the problem;

[0160] S62: Matrixing of Pelican Populations in POA This indicates that the current solution for all pelicans is recorded for subsequent iterations, updates, and optimizations.

[0161] ;

[0162] in, It is the first The Pelicans' position It is a matrix representation of the pelican population;

[0163] S63: The objective function value vector is in the form of a vector used to measure the quality of the solution for each pelican. Each individual pelican represents a set of LSTM parameters, which are optimized by an optimization algorithm to find the optimal solution.

[0164] ;

[0165] in, It is the first The objective function value for each pelican. It is the objective function value vector of the entire pelican population;

[0166] S64: The search phase, where the pelican approaches its prey, is precisely modeled so that the POA algorithm can effectively scan the search space and fully utilize its ability to explore different areas. Simultaneously, the prey's position within the search space is randomly assigned, a mechanism that further enhances the POA algorithm's exploration efficiency when handling precise search problems. This characteristic constitutes an important component of the POA algorithm. The following is a mathematical approximation of the above principles and hunting strategies:

[0167] ;

[0168] in, This is the first update after the first phase (exploration phase). The Pelicans in The position of the dimension It is an integer value in the range [1,2]. For the prey in the The position of the dimension; The objective function value for the prey;

[0169] If the objective function value decreases at the new position, the POA algorithm adopts the new position of the pelican. This update is considered effective and can prevent the pelican from moving to a non-optimal position. This process can be described by the following formula:

[0170] ;

[0171] In the formula, For the first A new position for the Pelicans for The objective function value;

[0172] S65: During the hunting phase, this planar surface flight strategy helps pelicans expand their attack range during the development phase. By simulating this behavior of pelicans, individual pelicans can be moved to more advantageous positions within the prey-catching area, thereby enhancing their local search capabilities. Subsequently, the program needs to check whether the position of the pelican has changed during the hunting process in order to move to a more suitable position. The mathematical model of pelican behavior is specifically described as follows.

[0173] ;

[0174] in, For the second phase update The Pelicans in The position of the dimension; A random integer between [0, 2]. for The radius of the domain; This represents the current iteration number; This represents the maximum number of iterations.

[0175] The mathematical model for updating the pelican's position during the capture phase is as follows:

[0176] ;

[0177] In the formula, For the first A new position for the Pelicans; for The objective function value;

[0178] S66: The improved POA algorithm introduces dynamic convergence coefficients, memory pool mechanism and hybrid search strategy into the Pelican Optimization Algorithm (POA). The three optimization strategies work together to find the best balance between global and local search, accelerate the search process, avoid premature convergence and improve the quality of the optimal solution.

[0179] In traditional Proof-of-Action (POA) algorithms, the search process often lacks a balance between global search and local exploitation, especially in the early and late stages of optimization, where the step size remains constant. This leads to insufficient exploration of the solution space by individuals in the early stages if the step size is small, making it difficult to cover a wide range of regions and prone to getting trapped in local optima. In the later stages of optimization, a large step size can cause individuals to skip high-quality solutions, hindering effective fine-grained local exploitation. This paper proposes a dynamic convergence coefficient mechanism, which dynamically adjusts the search range and step size of individuals according to different stages of the optimization process. By introducing dynamic convergence coefficients, the algorithm can achieve a smooth transition from the initial broad global search to the later fine-grained local exploitation. Definition of dynamic convergence coefficient:

[0180] ;

[0181] In the formula, This is the current iteration number; This represents the maximum number of iterations. This is a parameter that controls the convergence speed, with a value range of [0.05, 0.2]. The larger the value, the faster the convergence.

[0182] Furthermore, in the traditional POA algorithm, individual position updates rely solely on information from the current population, neglecting the utilization of high-quality solutions found during historical iterations. This often leads to high-quality solutions failing to continue contributing to the optimization process, resulting in local optima. To address this, this invention introduces a memory pool mechanism to store historical high-quality solutions generated during optimization, preventing the algorithm from losing valuable information. The solutions in the memory pool are not only used for population updates but also guide individuals towards better search directions, enhancing the synergistic effect of global and local searches. Through the memory pool mechanism, historical information can be fully utilized, improving search efficiency and overall algorithm performance. The update rules for the memory pool are as follows:

[0183] ;

[0184] In the formula, Individuals are randomly selected from the memory pool; It is the perturbation coefficient, which controls the intensity of the introduction of solutions into the memory pool;

[0185] Furthermore, in traditional POA algorithms, individual updates rely on searching for solutions within the population. However, this approach often fails to effectively utilize global information from high-quality solutions, leading to a singular search direction, insufficient solution diversity, and an increased probability of getting trapped in local optima. To further improve population diversity and optimization capabilities, this invention introduces the concept of Differential Evolution (DE) into the individual update process. The hybrid search strategy combines the current optimal solution, the memory pool solution, and the differential information between individuals through differential evolution, prompting individuals to form diverse exploration paths in the search space. Through this hybrid search strategy, the algorithm can more efficiently balance global search and local exploitation, significantly improving search capabilities during the optimization process. The differential evolution formula is as follows:

[0186] ;

[0187] In the formula, The differential scaling factor has a value range of [0.5, 1]; xbest1 and xbest2 are two different individuals randomly selected from the memory pool;

[0188] Furthermore, during the search phase, the algorithm needs to balance the ability to explore globally and exploit locally. By dynamically adjusting the search step size of each individual and introducing a memory pool mechanism, individuals can utilize historical high-quality solutions for more efficient searching, ensuring a balance between global search and local exploitation, and improving the algorithm's convergence accuracy and efficiency. The improved formula for the search phase is as follows:

[0189] ;

[0190] In the formula, A solution randomly selected from the memory pool; The memory pool contribution weight controls the strength of the memory pool's effect, with a value range of [0.1, 0.5].

[0191] Furthermore, the hunting phase primarily focuses on in-depth exploration of local areas while maintaining a certain level of global exploration capability. By comparing high-quality solutions and mining discrepancies, the granularity and diversity of the search are improved. The individual update formula for the hunting phase is as follows:

[0192] ;

[0193] In the formula, and These are two solutions randomly selected from the memory pool.

[0194] S7: Construct an LSTM model to predict CO concentration and coal temperature based on the optimal parameter combination output, and retrain and test the dataset; use the constructed dataset to train and validate the model, and output the best LSTM parameter combination through the IPOA algorithm. The initial range and optimal value of the optimized parameters are shown in Table 1 below.

[0195] Table 1 shows the model hyperparameter values.

[0196]

[0197] To intuitively compare the performance of the four optimization algorithms—WOA-LSTM, PSO-LSTM, POA-LSTM, and IPOA-LSTM—such as... Figure 11 As shown, fitness curves for four different models in the process of finding the optimal parameter combination are plotted. MSE represents the average of the squares of the errors between the model's predicted values ​​and the actual values; a smaller value indicates a smaller prediction error and better model performance. Analysis shows that in hyperparameter optimization with MSE as the objective function, the convergence curves of all algorithms undergo three stages: a sharp decrease, fluctuations, and stable convergence. Specifically, in the early stages of training, the algorithm globally explores the search space, and the MSE decreases sharply; as the number of iterations increases, the algorithm approaches the optimal solution region, and the MSE fluctuates due to the complexity of the objective function and the local sensitivity of the parameters; finally, under the adjustment of the dynamic convergence factor and search strategy of the optimization algorithm, the hyperparameters tend to stabilize, and the MSE tends to plateau, marking the end of the optimization process. The local plot shows that the MSE of the four algorithms is ranked as follows:

[0198] WOA-LSTM>PSO-LSTM>POA-LSTM>IPOA-LSTM,

[0199] Among them, IPOA-LSTM has the smallest MSE and requires the fewest iterations, converging in only 136 iterations, while the other three algorithms converged after 175 iterations. WOA-LSTM and PSO-LSTM are prone to getting trapped in local optima, with large and fluctuating MSEs; while POA-LSTM shows some optimization, its convergence speed is slow. In contrast, IPOA-LSTM, through reasonable parameter settings, dynamic convergence coefficients, and a hybrid strategy, not only quickly finds a parameter combination close to the global optimum, reducing training time, but also improves the model's prediction accuracy and stability.

[0200] S8: Train and build LSTM, WOA-LSTM, PSO-LSTM, POA-LSTM, and IPOA-LSTM algorithms sequentially, and compare them by evaluation metrics (MAE, RMSE, R). 2 The prediction results are compared to the prediction accuracy to verify the advantages of the IPOA-LSTM model.

[0201] To evaluate the feasibility and accuracy of this model's predictions, the test set results of the trained model are further calculated using several evaluation metrics. For regression models, common evaluation metrics include Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination. wait;

[0202] MAE represents the mean of the absolute errors between the actual and predicted values, and can effectively reflect the error of the predicted values. The calculation formula is as follows:

[0203] ;

[0204] RMSE represents the deviation between the model's predicted values ​​and the true values. The closer the value is to 0, the better the model's generalization performance and the higher its accuracy. Its calculation formula is as follows:

[0205] ;

[0206] The coefficient of determination reflects the degree of correlation between the dependent and independent variables. The closer its value is to 1, the better the correlation between the predicted and actual values. The calculation formula is as follows:

[0207] ;

[0208] To verify the performance of the IPOA-LSTM model, this paper establishes prediction models for CO concentration and coal temperature during the coal spontaneous combustion latency period based on LSTM, WOA-LSTM, PSO-LSTM, POA-LSTM, and IPOA-LSTM algorithms, using the optimal parameter combination obtained in step S7. The prediction results of the LSTM model on the training and test sets after optimization by different algorithms are shown below. Figure 9 As shown in the figure, the unoptimized LSTM model performs poorly in both the training and test sets. The prediction trends of the four improved LSTM models are basically consistent with the trends of the original data, indicating that the optimization algorithm has certain advantages. From the magnified view, it can be seen that the prediction results of WOA-LSTM, PSO-LSTM, and POA-LSTM differ significantly from the original values ​​when the CO concentration fluctuates greatly, while the prediction results of IPOA-LSTM are always roughly consistent with the original data, with the smallest error.

[0209] To more intuitively observe the prediction results of the four models and quantitatively evaluate their performance, a radar chart illustrating the model prediction evaluation was created based on the prediction results from the training and test sets. Figure 14-15 It can be observed that, compared to other prediction models, IPOA-LSTM has higher R... 2 The values ​​are 0.95 and 0.97, respectively, which are closest to 1 compared to other models. The MAE is 8 and 12, and the RMSE is 10 and 6. Compared to the other four models, the closer the value is to 0, the better the generalization performance and the higher the accuracy of the model. The above indicators fully demonstrate the effectiveness of the IPOA-LSTM model in improving prediction accuracy. It has high accuracy and stability in the task of predicting CO concentration and coal temperature, can identify the risk of coal spontaneous combustion in advance, and improve the accuracy of predicting the latency period of coal spontaneous combustion.

[0210] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting CO concentration and coal temperature during the coal spontaneous combustion latency period, characterized in that: Includes the following steps: S1: Import historical data on spontaneous combustion of coal in the goaf of the mine into the prediction system to construct a time series dataset of CO concentration, coal temperature and preset indicators; S2: Perform data preprocessing on the dataset, including missing value imputation and data normalization, to improve model accuracy; S3: Use the Spearman correlation coefficient method to determine the correlation between preset indicators, and select features based on the magnitude of the correlation coefficient. S4: Perform feature importance analysis on the feature selection metrics of the model, and use the Shapley algorithm to calculate the Shapley value of the features to obtain different metrics; S5: The preprocessed dataset is divided into training, validation, and test sets in an 8:1:1 ratio. An LSTM neural network model is constructed, and the model parameters are initialized. S6: Optimize the LSTM model parameters using the improved POA algorithm. When the improved POA-LSTM optimization model meets the preset conditions, select the hyperparameter configuration. Step S6 includes: S61: Run the POA algorithm, calculate the objective function value, randomly generate prey, update the position during the search phase, and update the position to save the optimal solution during the hunting phase, until the number of iterations is reached; This includes population initialization, used to generate the initial population, and the random value rand ensures that the initial positions of the pelicans are evenly distributed between the upper and lower bounds of the solution, covering the entire solution space; ； in, It is the first The Pelicans' first Position in dimensions; rand It is a random number, ranging from [0,1]. Indicates the first In dimension, the upper boundary of the problem, The lower boundary; This refers to the pelican population size. The dimensions of the problem; S62: Matrixing of Pelican Populations in POA This indicates that the current solution for all pelicans is recorded for subsequent iterations, updates, and optimizations. ； in, It is a matrix representation of the pelican population; S63: The objective function value vector is in the form of a vector used to measure the quality of the solution for each pelican. Each individual pelican represents a set of LSTM parameters, which are optimized by an optimization algorithm to find the optimal solution. ； in, It is the first The objective function value for each pelican. It is the objective function value vector of the entire pelican population; S64: The search phase, i.e., how the pelican approaches its prey, is precisely modeled; S65: Hunting phase; S66: The improved POA algorithm simultaneously introduces dynamic convergence coefficients, a memory pool mechanism, and a hybrid search strategy into the Pelican optimization algorithm; among them, the dynamic convergence coefficients are used to achieve a smooth transition from the initial global broad search to the later fine-grained local exploration; the definition of the dynamic convergence coefficients is as follows: ； In the formula, This is the current iteration number; This is the total number of iterations; It is a parameter that controls the convergence speed, and its value ranges from [0.05, 0.2]. A hybrid search strategy is used to balance global search and local development, using the following formula: ； In the formula, The differential scaling factor has a value range of [0.5, 1]. , Two different individuals were randomly selected from the memory pool; Among them, the improved formula for the search phase, combining dynamic convergence coefficients and the memory pool mechanism, is as follows: ； In the formula, It is an integer value in the range [1, 2]. The objective function value of the prey. The solution is randomly selected from the memory pool; This contributes weight to the memory pool, controlling the strength of the memory pool's effect, with a value range of [0.1, 0.5]. This requires the introduction of a differential evolution strategy to update individuals during the hunting phase. The update formula is as follows: ； In the formula, A random integer between [0, 2]. for The radius of the domain, and These are two solutions randomly selected from the memory pool; S7: Construct an LSTM model to predict CO concentration and coal temperature based on the optimal parameter combination of the output, and retrain and test the dataset; S8: Train and build LSTM, WOA-LSTM, PSO-LSTM, POA-LSTM and IPOA-LSTM algorithms in sequence. By comparing evaluation metrics, judge the prediction results and prediction accuracy, and verify the advantages of the IPOA-LSTM model.

2. The method for predicting CO concentration and coal temperature during the coal spontaneous combustion latency period as described in claim 1, characterized in that: Step S2 includes: S21: Missing value imputation, where the data imputation method uses linear interpolation; S22: Outlier Handling, wherein the IQR-based outlier handling method detects and handles outliers during the data preprocessing stage; this includes calculating the quartiles of the data and using IQR to determine whether outliers exist in the data, wherein the calculation steps are as follows: Calculate the first quartile Q1 and the third quartile Q3 of the data; Calculate IQR = Q3 - Q1, where IQR represents the distribution range of the middle 50% of the data in the dataset; The low and high thresholds for outliers are set as follows: Lower Bound = Q1 - 1.5 × IQR, Lower Bound = Q3 + 1.5 × IQR. Values ​​below the low threshold or above the high threshold are considered outliers. S23: Data normalization, used to reduce the range of model parameter updates during training.

3. The method for predicting CO concentration and coal temperature during the coal spontaneous combustion latency period as described in claim 1, characterized in that: Step S3 includes: S31: Collect multi-dimensional data related to spontaneous combustion of coal, including CO concentration, coal temperature, carbon dioxide concentration, oxygen concentration, methane concentration, and olefin concentration. S32: Set up a data matrix, where each row represents a sample and each column represents an indicator or target variable, where the target variable is the risk of spontaneous combustion of coal or the spontaneous combustion index; S33: Use the Spearman correlation coefficient to calculate the correlation between each indicator and the risk of spontaneous combustion of coal, and quantitatively assess the linear association between the relevant indicators; wherein, the Spearman correlation coefficient between each indicator is calculated using the following formula: ； in: It is the difference in the ranking of each pair of variables. It is the sample size. It is the Spearman correlation coefficient; S34: Generate a heatmap: The heatmap function in MATLAB is used to visualize the weights of the indicators. Based on the normalized results, the Spearman correlation coefficient between each indicator is obtained.

4. The method for predicting CO concentration and coal temperature during the coal spontaneous combustion latency period as described in claim 1, characterized in that: Step S4 includes: S41: Quantify the contribution of each feature to a single prediction; S42: Given a prediction, calculate the marginal contribution of each feature using the Shapley value, including calculations based on all possible feature subsets. The formula for calculating the Shapley value is as follows: ； in, Represents the set of all features. Indicates that features are not included. of a subset of Indicates the number of elements in the set. Indicates that it contains a subset of features and characteristics Time sample points The predicted value, Indicates that it contains a subset of features Time sample points The predicted value, Measured in the control feature subset At that time, characteristics The magnitude of the impact on the predicted value, sample points Features The Shapley value is for all possible subsets of features. Lower features The effects are weighted and summed; S43: Obtain the importance of different predictors in the prediction model by calculating feature importance. Mean absolute Shapley value across all sample points ; ； in Represents the sample set, express The absolute value, Represents the total number of samples. That is, characteristics The overall feature importance in the prediction model The larger the value, the more characteristic it indicates. The more important it is.

5. The method for predicting CO concentration and coal temperature during the coal spontaneous combustion latency period as described in claim 1, characterized in that: Step S5 includes: S51: Input layer dimension; S52: Output layer dimension; S53: Hidden layer dimension; S54: Activation function S55: Number of neurons in the hidden layer S56: Input layer time steps S57: Learning rate; S58: Number of iterations.

Images