A temperature prediction and fire warning method for exposed area of coal seam in mining area

Abstract

The application discloses a kind of temperature prediction and fire warning method of exposed area of coal seam in mining area, and the method comprises: obtaining the multi-source monitoring data of research area including ground station temperature detection data, satellite remote sensing monitoring data;Temperature prediction model including temperature inversion physical module, Gaussian process regression module is constructed, temperature inversion physical module is obtained by satellite remote sensing monitoring data inversion calculation and obtains the preliminary ground surface temperature of each pixel, Gaussian process regression module constructs physical characteristic guide vector and ground station temperature detection data are handled by regression and finally the ground surface temperature of each pixel is obtained by prediction;Fire warning grade is set and the fire warning grade of output pixel is judged.The temperature prediction model of the application can obtain spatial continuous high-precision final ground surface temperature;ADM-TCN time series prediction model predicts the predicted ground surface temperature of future time period, which provides technical support for the advanced identification of fire hazards.

Description

Technical Field

[0001] This invention relates to the field of temperature prediction and fire prevention and early warning in exposed coal seams, and particularly to a method for temperature prediction and fire early warning in exposed coal seams in mining areas. Background Technology

[0002] Currently, coal seam temperature detection largely relies on manual inspections or automated sensors (such as thermocouples and infrared thermometers) to collect data at specific points. However, mining areas have undulating terrain, frequent face movement, and harsh working environments, making sensor deployment and maintenance extremely costly. Furthermore, discrete point data struggles to reflect the spatial heterogeneity of temperature across complex terrains and different geological structures, failing to meet the need for a comprehensive and dynamic assessment of fire risk across the entire mining area. In addition, existing methods lack effective means to predict long-term trends in surface temperature in exposed coal seams, making it difficult to eliminate spontaneous combustion hazards at their initial stage. Therefore, how to comprehensively utilize multi-source monitoring information to achieve accurate and dynamic global monitoring and long-term trend prediction of surface temperature in exposed coal seams has become a core issue for early warning of spontaneous combustion in coal mines and for transforming passive fire suppression into proactive prevention and control. Summary of the Invention

[0003] The purpose of this invention is to solve the technical problems pointed out in the background art and provide a method for temperature prediction and fire early warning in exposed coal seam areas of mining areas. The Gaussian process regression module takes environmental physical parameters closely related to the cause of error, such as vegetation index, elevation, and emissivity index, as inputs. It uses temperature data from ground measurement stations to learn the nonlinear spatial distribution law of the residuals and achieves high-precision temperature error field prediction for the entire area under limited station sample conditions, thereby obtaining the final surface temperature with continuous spatial accuracy.

[0004] The objective of this invention is achieved through the following technical solution:

[0005] A method for temperature prediction and fire early warning in exposed coal seam areas of a mining area, the method comprising:

[0006] S1. Acquire multi-source monitoring data for the study area, including ground station temperature detection data and satellite remote sensing monitoring data;

[0007] S2. Construct a temperature prediction model that includes a temperature inversion physics module and a Gaussian process regression module. The temperature inversion physics module uses satellite remote sensing monitoring data to invert and calculate the preliminary surface temperature of each pixel. The Gaussian process regression module constructs a physical feature guiding vector and performs regression processing with the temperature detection data of ground stations to predict the final surface temperature of each pixel.

[0008] S3. Construct fire warning levels based on the temperature range of the ground surface temperature, and output the fire warning level of the pixel based on the final ground surface temperature of the pixel according to the fire warning level classification.

[0009] To better implement this invention, in method S3, the fire warning levels include no risk, low risk, medium risk, and high risk, and the final surface temperature of pixel i. The method for determining the fire warning level is as follows: If the final surface temperature of pixel i and the average final surface temperature of neighboring pixels are both less than or equal to the temperature threshold. If the final surface temperature of pixel i and the average final surface temperature of its neighboring pixels are both within the range of risk, then pixel i is considered risk-free. Within the interval, pixel i is considered to be at a low risk; if the final surface temperature of pixel i and the average final surface temperature of its neighboring pixels are both within the range... Within the interval, pixel i is classified as moderate risk; if the final surface temperature of pixel i and the average final surface temperature of its neighboring pixels are both greater than the threshold temperature... If pixel i is at high risk, then pixel i is at high risk. , These are the temperature thresholds.

[0010] In a further technical solution, the present invention also includes the following method:

[0011] S4. All pixels in the study area are stratified and grouped according to fire warning level to form a risk-free set, a low-risk set, a medium-risk set, and a high-risk set. Pixels in the low-risk set are continuously monitored and monitored, pixels in the medium-risk set are investigated for potential hazards, and pixels in the high-risk set are given priority for emergency treatment. The temperature prediction model contains a geographic GIS map, in which the fire warning level of the pixels is displayed in a stratified and visual manner.

[0012] Preferably, the study area is equipped with several ground stations, each equipped with a temperature detector, which is used to detect and obtain the measured surface temperature corresponding to the ground station. The data is used as ground station temperature detection data; the satellite remote sensing monitoring data is Landsat series satellite remote sensing image data, which includes preprocessing such as radiometric calibration, atmospheric correction, and geometric correction. The remote sensing monitoring data also includes digital elevation model data, and all data in the remote sensing monitoring data are registered to the same coordinate system according to their location.

[0013] Preferably, in method S2, the temperature inversion physical module obtains the preliminary surface temperature of each pixel using the following method:

[0014] S201. Extract the radiance values ​​of two thermal infrared bands from satellite remote sensing monitoring data. , Construct the emissivity index according to the following formula : ;

[0015] S202. Obtain the preliminary surface temperature of each pixel according to the following inversion formula. : ,in , These are the brightness temperatures obtained from two thermal infrared bands in satellite remote sensing data. , , These are the adaptive coefficients, , , , , These are constants related to the atmosphere.

[0016] Preferably, the Gaussian process regression module obtains the final surface temperature of the pixel. The method is as follows:

[0017] S211. Obtain the measured surface temperature from the ground station temperature detection data. Preliminary surface temperature at the same location Calculate residuals , ;

[0018] S212. Obtain the vegetation index NDVI and digital elevation model (DEM) from satellite remote sensing monitoring data, and construct a model including the vegetation index NDVI, DEM, and emissivity index. Preliminary surface temperature The physical feature guiding vector X, including the physical feature guiding vector X, is combined with the residual. The corresponding pairs constitute the dataset, and the samples in the dataset are the physical feature guiding vector X and the residual. For the corresponding pairing combinations, the Gaussian process regression module uses the following kernel function to capture the smooth changes of residuals with environmental characteristics: ,in Let p and q be the kernel function values. Let be the feature vectors corresponding to samples p and q, respectively. It is a diagonal matrix with a length scale. This is the transpose of the difference vector between the feature vectors corresponding to sample p and sample q. For signal variance, To observe the noise variance; For the Kronecker function; The hyperparameters are optimized by maximizing the log-marginal likelihood function, as shown in the following expression: ,in For residuals Physical feature guiding vector X, hyperparameters The probability, For probability Find the natural logarithm. This is the transpose of the residual vector. The total number of samples in the dataset. Let covariance matrix be the variance matrix. , It is the identity matrix. It is a noise-free kernel covariance matrix;

[0019] S213. Obtain the prediction residuals for each pixel to be predicted using the following formula. : ,in The kernel vector is the pixel to be predicted and all samples; the Gaussian process regression module obtains the final surface temperature of each pixel according to the following formula. : .

[0020] Preferably, the present invention further includes the following method:

[0021] S5. Construct an ADM-TCN time-series prediction model including multi-scale dilated convolutional blocks and a dynamic memory update module. Following methods S1 and S2, obtain the final land surface temperature time-series data with pixels arranged in a time series. Input the final land surface temperature time-series data into the ADM-TCN time-series prediction model. The multi-scale dilated convolutional blocks capture the short-term, medium-term, and long-term features of the final land surface temperature time-series data and fuse them to obtain comprehensive time-series features. The dynamic memory update module includes an update gate and a reset gate. The update gate determines the number of historical memory states to be retained, and the reset gate controls the influence of historical memory information on the current candidate states. It fuses the filtered historical memory information with the input features to generate candidate memory states. Then, it uses linear interpolation to weightedly fuse the historical memory states and candidate states to form candidate features and dynamically update the memory. The ADM-TCN time-series prediction model outputs the predicted land surface temperature for a future period. A fire warning level is set to determine the fire warning level for the predicted land surface temperature.

[0022] Preferably, the ADM-TCN time-series prediction model is trained using land surface temperature time-series sample data. The ADM-TCN time-series prediction model uses a weighted Huber loss function as the total loss function, as expressed below: ,in The Huber loss function is expressed as follows: , This represents the actual surface temperature at a future time m. For the predicted surface temperature at a future time m, For threshold parameters, The time decay weight for future time m, The total future time is predicted by the ADM-TCN time series forecasting model.

[0023] Preferably, the fire warning levels include no risk, low risk, medium risk, and high risk. The predicted surface temperature of pixel i at a future time m and the average predicted surface temperature of neighboring pixels are obtained. The fire warning level is determined as follows: if the predicted surface temperature of pixel i at a future time m and the average predicted surface temperature of neighboring pixels are both less than or equal to a temperature threshold... Then pixel i is risk-free; if the predicted surface temperature of pixel i at future time m and the average predicted surface temperature of its neighboring pixels are both within the range... Within the interval, pixel i is considered to have a mild risk; if the predicted surface temperature of pixel i at future time m and the average predicted surface temperature of its neighboring pixels are both within the range... Within the interval, pixel i is considered to be at medium risk; if the predicted surface temperature of pixel i at a future time m and the average predicted surface temperature of its neighboring pixels are both greater than the threshold temperature... If pixel i is at high risk, then pixel i is at high risk. , These are the temperature thresholds.

[0024] Preferably, the study area includes exposed coal seams and other areas besides exposed coal seams. The ground stations are mainly located in various positions of exposed coal seams in the study area, and some ground stations also cover other areas of the study area.

[0025] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0026] (1) This invention constructs a temperature prediction model that integrates physical feature-guided vectors and data-driven approaches. The inversion coefficients are dynamically adjusted according to the emissivity index of coal type and land cover. The system deviation caused by the heterogeneity of complex surface materials is corrected from the physical mechanism, which significantly improves the inversion accuracy of the initial surface temperature of exposed coal seams in mining areas. Furthermore, a physical feature-guided Gaussian process regression residual correction model (i.e., Gaussian process regression module) is constructed. The Gaussian process regression module takes environmental physical parameters closely related to the cause of error, such as vegetation index, elevation, and emissivity index, as inputs. It uses the temperature data of ground measurement stations to learn the nonlinear spatial distribution law of the residuals. Under the condition of limited station samples, it achieves high-precision temperature error field prediction for the whole area, thereby obtaining the final surface temperature with continuous spatial accuracy. This effectively solves the technical problem that traditional point sensors are difficult to reflect the spatial heterogeneity of temperature.

[0027] (2) This invention constructs an ADM-TCN time series prediction model for long-term trend prediction, which realizes multi-scale dynamic and accurate prediction of surface temperature in exposed coal seam areas. By extracting short-term, medium-term and long-term time series features through multi-scale dilated convolutional blocks and fusing them, it can simultaneously capture multi-scale temperature change patterns such as daily variation, weather fluctuations and seasonal trends. At the same time, through the gating mechanism of the dynamic memory update module, it adaptively filters and fuses historical information, explicitly maintains the long-term memory state across long time steps, and combines the weighted Huber loss function and time decay weight for model training, which can prioritize the accuracy of recent predictions while taking into account the reliability of long-term trends, providing technical support for the advanced identification of fire hazards.

[0028] (3) This invention creates a hierarchical fire early warning method that integrates neighborhood spatial constraints, which improves the scientific nature and robustness of the early warning. When determining the fire early warning level, it not only considers the final or predicted surface temperature of the pixel itself, but also introduces the average temperature of the neighboring pixels as a joint judgment basis, which effectively avoids false alarms or missed alarms caused by abnormal fluctuations of individual pixels, and enhances the spatial consistency and reliability of the early warning results. At the same time, it uses a geographic GIS map to display the risk levels of no risk, mild risk, moderate risk and high risk in a hierarchical visualization, and takes hierarchical processing measures for pixels of different risk levels, which provides intuitive and efficient decision support for transforming passive fire extinguishing into active and precise prevention and control of spontaneous combustion fires in mining areas. Attached Figure Description

[0029] Figure 1 This is a flowchart of the temperature prediction and fire early warning method of the present invention;

[0030] Figure 2 This is a simplified schematic diagram illustrating the principles of the temperature inversion physics module and the Gaussian process regression module in the temperature prediction model of the example embodiment.

[0031] Figure 3 This is a schematic diagram illustrating the structure and principle of the ADM-TCN time series prediction model in the embodiment. Detailed Implementation

[0032] The present invention will be further described in detail below with reference to embodiments:

[0033] Example 1

[0034] like Figure 1 As shown, a method for temperature prediction and fire early warning in exposed coal seam areas of a mining area includes the following steps:

[0035] S1. Acquire multi-source monitoring data for the study area, including temperature probe data from ground stations and satellite remote sensing data. Several ground stations are deployed in the study area, each equipped with a temperature detector used to measure the actual surface temperature at each station. This data serves as temperature detection data from ground stations. In this embodiment, the study area includes exposed coal seams and other areas excluding exposed coal seams. Ground stations are mainly deployed at various locations within the exposed coal seams of the study area, with some stations also covering other areas. Satellite remote sensing monitoring data comprises Landsat series satellite remote sensing imagery, including thermal infrared, visible, and near-infrared bands. The Landsat series satellite remote sensing imagery includes preprocessing such as radiometric calibration, atmospheric correction, and geometric correction (eliminating geometric distortion). The remote sensing monitoring data also includes digital elevation model data. All data in the remote sensing monitoring data are registered to the same coordinate system (preferably the WGS84 UTM projection coordinate system in this invention) according to their location.

[0036] S2. Construct a temperature prediction model including a temperature inversion physics module and a Gaussian process regression module. The temperature inversion physics module uses satellite remote sensing monitoring data to invert and calculate the preliminary surface temperature of each pixel. The Gaussian process regression module constructs a physical feature-guided vector and performs regression processing with ground station temperature detection data to predict the final surface temperature of each pixel. This invention's temperature prediction model uses a multi-source monitoring sample dataset for temperature prediction training. The multi-source monitoring sample dataset stores multi-source monitoring sample data and corresponding temperature labels and fire event labels. The temperature label is the measured surface temperature, and the fire event label is events such as smoke and open flame. The multi-source monitoring sample data includes ground station temperature detection sample data, satellite remote sensing monitoring sample data, and digital elevation model sample data for the study area.

[0037] The method by which the temperature inversion physics module in the temperature prediction model of this invention obtains the preliminary surface temperature of each pixel is as follows:

[0038] S201. Extract the radiance values ​​of two thermal infrared bands from satellite remote sensing monitoring data. , The surface reflectance and radiance values ​​of two thermal infrared bands were obtained from satellite remote sensing data, and the emissivity index was constructed according to the following formula. (Also known as the SBR index, it is the emissivity of the two thermal infrared bands): The emissivity index of a cell can vary depending on whether it is a non-coal body or a coal body (including different types of coal, such as lignite, bituminous coal, and anthracite). This allows for the construction of a coefficient lookup table (recording the corresponding emissivity index according to the type of non-coal body and coal body). In some embodiments, such as... Figure 2As shown, the temperature inversion physics module uses an adaptive split-window algorithm to perform a preliminary physical inversion of the land surface temperature, and the emissivity index can be used as a priori parameter for the split-window algorithm.

[0039] S202. The temperature inversion physics module obtains the preliminary surface temperature of each pixel according to the following inversion formula. : ,in , The brightness temperatures of the two thermal infrared bands are obtained from satellite remote sensing monitoring data. The surface reflectance and radiance values ​​of the two thermal infrared bands are obtained from the satellite remote sensing monitoring data, and then the brightness temperatures of the two thermal infrared bands are obtained. , , These are the adaptive coefficients, , , , , These are constants related to the atmosphere (obtained through atmospheric radiative transfer simulation). This is a dimensionless coefficient (used to quantify the combined effect of atmospheric water vapor on the difference in atmospheric transmittance between the two thermal infrared bands), with a value range of 0.98 to 1.02. The unit is Kelvin (used to characterize the contribution of atmospheric thermal radiation to the inversion of surface temperature). The value range is -2.0 to 2.0.

[0040] The Gaussian process regression module of this invention (see...) Figure 2 The Gaussian process regression module (also known as the PhysStat-GPR residual correction module) obtains the final land surface temperature of the pixels. The method is as follows:

[0041] S211. Obtain the measured surface temperature from the ground station temperature detection data. Preliminary surface temperature at the same location Calculate residuals , .

[0042] S212. Obtain the vegetation index NDVI and digital elevation model (DEM) from satellite remote sensing data. The NDVI (Normalized Difference Vegetation Index) is calculated using near-infrared and red band data and can be used to generate a binary vegetation mask for the study area. Construct a model including the NDVI, DEM, and emissivity index. Preliminary surface temperature The physical feature guiding vector X, including the physical feature guiding vector X, is combined with the residual. The corresponding pairs constitute the dataset, and the samples in the dataset are the physical feature guiding vector X and the residual. The corresponding pairing combinations, residuals Obeying the Gaussian process prior, The Gaussian process regression module uses the following kernel function to capture the smooth changes in residuals with environmental characteristics: ,in Let p and q be the kernel function values. Let be the feature vectors corresponding to samples p and q, respectively. It is a diagonal matrix with a length scale. This is the transpose of the difference vector between the feature vectors corresponding to sample p and sample q. For signal variance, To observe the noise variance. Let Kronecker function be used. When sample p is the same as sample q, then... =1, when sample p and sample q are not the same, then =0. With The hyperparameters are optimized by maximizing the log-marginal likelihood function, as shown in the following expression: ,in For residuals Physical feature guiding vector X, hyperparameters The probability, For probability Find the natural logarithm. For residual vectors transpose, The total number of samples in the dataset. , The residuals for samples n and N are respectively. , , These are the measured surface temperature and the preliminary surface temperature of sample n, respectively. Let covariance matrix be the variance matrix. , It is the identity matrix. It is a noise-free kernel covariance matrix.

[0043] S213. Obtain the prediction residuals for each pixel to be predicted using the following formula. : ,in For the pixel to be predicted With all samples (including) , ... The kernel vector of ) The Gaussian process regression module obtains the final surface temperature of each pixel according to the following formula. (like Figure 2 As shown, the final surface temperature is also known as the high-precision surface temperature. .

[0044] S3. Construct fire warning levels based on surface temperature ranges, and output the fire warning level of a pixel based on its final surface temperature. Preferably, the fire warning levels include no risk, low risk, moderate risk, and high risk, and the final surface temperature of pixel i... The method for determining the fire warning level is as follows: If the final surface temperature of pixel i and the average final surface temperature of neighboring pixels are both less than or equal to the temperature threshold. If the final surface temperature of pixel i and the average final surface temperature of its neighboring pixels are both within a certain range, then pixel i is considered risk-free. Within the interval, pixel i is considered to be at a low risk. If the final surface temperature of pixel i and the average final surface temperature of its neighboring pixels are both within the range... Within the interval, pixel i is classified as moderate risk. If both the final surface temperature of pixel i and the average final surface temperature of its neighboring pixels are greater than the threshold temperature... If pixel i is at high risk, then pixel i is at high risk. , These are the temperature thresholds.

[0045] In some embodiments, the present invention further includes the following method:

[0046] S4. All pixels in the study area are stratified and grouped according to fire warning levels to form risk-free, low-risk, medium-risk, and high-risk sets. Pixels in the low-risk set are continuously monitored, pixels in the medium-risk set are investigated for potential hazards, and pixels in the high-risk set are given priority for emergency handling. The temperature prediction model includes a geographic GIS map, which visualizes the fire warning levels of pixels in a stratified manner. Different risk levels of pixels are treated with tiered measures, providing intuitive and efficient decision support for transforming passive fire suppression into proactive and precise prevention and control of spontaneous combustion fires in mining areas.

[0047] Example 2

[0048] This embodiment is completely identical to Embodiment 1 in methods S1 to S4, except that this embodiment also includes the following technical content:

[0049] S5. Construct an ADM-TCN temporal prediction model that includes multi-scale dilated convolutional blocks and a dynamic memory update module. The structure and principle of the ADM-TCN temporal prediction model are as follows: Figure 3As shown; following methods S1 and S2, the final land surface temperature time-series data, arranged in a time series by pixels, is obtained. This final land surface temperature time-series data is input into the ADM-TCN time-series prediction model. Multi-scale dilated convolutional blocks capture the short-term, medium-term, and long-term features of the final land surface temperature time-series data and fuse them to obtain a comprehensive time-series feature. Specifically, the multi-scale dilated convolutional block has three independent convolutional paths with different dilation levels. These three paths extract short-term features (focusing on temperature fluctuations within a single day or several days), medium-term features (focusing on weekly scale changes), and long-term features (focusing on seasonal and annual temperature trends), respectively. The features output from the three paths are then concatenated to form a comprehensive time-series feature that fuses multiple time scales. The dynamic memory update module includes an update gate and a reset gate. The update gate is used to determine the number of historical memory states that need to be retained, expressed as follows: , Activated for Sigmoid. To update the gate weight parameters, For the memory of time t-1, The features of time t (corresponding to comprehensive time series features). To update the gate's bias term; The larger the value, the more preferentially the model stores the new features; The smaller the value, the more long-term historical temperature patterns are preserved. The reset gate controls the degree to which historical memory information influences the current candidate state, and its expression is as follows: , Activated for Sigmoid. To reset the weight parameters of the door, For the memory of time t-1, The features of time t (corresponding to comprehensive time series features). To reset the gate bias parameters; the candidate memory states are generated by fusing the filtered historical memory information with the input features. The candidate memory expression is as follows: , is the activation function (compresses features to [-1, 1]). For bias parameters, For weight parameters, For the output candidate memory, The Hadamard is multiplied element-wise; then, the historical memory state and the candidate state are weighted and fused using linear interpolation to form candidate features, and the memory is dynamically updated. The expression is as follows: The ADM-TCN time series prediction model forward propagates to obtain the memory state of the last updated time step. (It condenses the long-term and short-term information of the entire input sequence, i.e., the final memory vector after the entire time series is calculated), and performs mapping processing through a full-validation layer to output the predicted land surface temperature for a future time period, the predicted land surface temperature at future time m. The expression is as follows: , For weight parameters, This is the output layer bias vector. A fire warning level is set to determine the fire warning level based on the predicted surface temperature.

[0050] In some embodiments, the ADM-TCN time series prediction model is trained using land surface temperature time series sample data. The land surface temperature time series sample data are selected as measured temperature sequence data detected by ground stations. The ADM-TCN time series prediction model uses the weighted Huber loss function as the total loss function, as shown in the following expression: ,in The Huber loss function is expressed as follows: , This represents the actual surface temperature at a future time m. For the predicted surface temperature at a future time m, The preset threshold parameter, The time decay weight for future time m, the time decay weight The value can be obtained using the following formula: ,in The range of values ​​is: Following the principle that the reliability of recent forecasts is higher than that of long-term forecasts, while ensuring the accuracy of the overall trend, priority is given to optimizing the accuracy of recent forecasts. This refers to the total future time predicted by the ADM-TCN time series forecasting model. When Using formula Calculate the Huber loss to facilitate faster model convergence; when Using formula Calculating Huber loss can reduce the interference of outlier data points on the training process.

[0051] In some embodiments, fire warning levels include no risk, low risk, medium risk, and high risk. The predicted surface temperature of pixel i at a future time m and the average predicted surface temperature of neighboring pixels are obtained. The fire warning level is determined as follows: if the predicted surface temperature of pixel i at a future time m and the average predicted surface temperature of neighboring pixels are both less than or equal to a temperature threshold... If pixel i is risk-free within the future time range m, then pixel i is risk-free. If the predicted surface temperature of pixel i at the future time m and the average predicted surface temperature of its neighboring pixels are both within the range... Within the interval, pixel i is at a slight risk within the future time range m. If the predicted surface temperature of pixel i at the future time m and the average predicted surface temperature of its neighboring pixels are both within the range... Within the interval, pixel i is at a moderate risk within the future time range m. If the predicted surface temperature of pixel i at the future time m and the average predicted surface temperature of its neighboring pixels are both greater than the threshold temperature... If pixel i is at high risk within the future time range m, then pixel i is at high risk. , These are the temperature thresholds.

[0052] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for temperature prediction and fire early warning in exposed coal seam areas of mining areas, characterized in that: The methods include: S1. Acquire multi-source monitoring data for the study area, including ground station temperature detection data and satellite remote sensing monitoring data; S2. Construct a temperature prediction model including a temperature inversion physics module and a Gaussian process regression module. The temperature inversion physics module uses satellite remote sensing monitoring data to invert and calculate the preliminary surface temperature of each pixel. The Gaussian process regression module constructs a physical feature guiding vector and performs regression processing with ground station temperature detection data to predict the final surface temperature of each pixel. The method by which the temperature inversion physics module obtains the preliminary surface temperature of each pixel is as follows: S201. Extract the radiance values ​​of the two thermal infrared bands from the satellite remote sensing monitoring data. , Construct the emissivity index according to the following formula : ; S202. Obtain the preliminary surface temperature of each pixel according to the following inversion formula. : ,in , These are the brightness temperatures obtained from two thermal infrared bands in satellite remote sensing data. , , These are the adaptive coefficients, , , , , These are constants related to the atmosphere; the Gaussian process regression module obtains the final surface temperature of the pixel. The method is as follows: S211. Obtain the measured surface temperature from the ground station temperature detection data. Preliminary surface temperature at the same location Calculate residuals , ; S212. Obtain the vegetation index NDVI and digital elevation model (DEM) from satellite remote sensing data, and construct a model including the vegetation index NDVI, DEM, and emissivity index. Preliminary surface temperature The physical feature guiding vector X, including the physical feature guiding vector X and the residual The corresponding pairs constitute the dataset, and the samples in the dataset are the physical feature guiding vector X and the residual. For the corresponding pairing combinations, the Gaussian process regression module uses the following kernel function to capture the smooth changes of residuals with environmental characteristics: ,in Let p and q be the kernel function values. Let be the feature vectors corresponding to samples p and q, respectively. It is a diagonal matrix with a length scale. This is the transpose of the difference vector between the feature vectors corresponding to sample p and sample q. For signal variance, To observe the noise variance; For the Kronecker function; The hyperparameters are optimized by maximizing the log-marginal likelihood function, as shown in the following expression: ,in For residuals Physical feature guiding vector X, hyperparameters The probability, For probability Find the natural logarithm. This is the transpose of the residual vector. The total number of samples in the dataset. Let covariance matrix be the variance matrix. , It is the identity matrix. It is a noise-free kernel covariance matrix; S213. Obtain the prediction residuals for each pixel to be predicted using the following formula. : ,in The kernel vector is the pixel to be predicted and all samples; the Gaussian process regression module obtains the final surface temperature of each pixel according to the following formula. : ; S3. Construct fire warning levels based on the temperature range of the ground surface temperature, and output the fire warning level of the pixel based on the final ground surface temperature of the pixel according to the fire warning level classification.

2. The method for temperature prediction and fire early warning in exposed coal seam areas of a mining area according to claim 1, characterized in that: In method S3, fire warning levels include no risk, low risk, medium risk, and high risk, and the final surface temperature of pixel i. The method for determining the fire warning level is as follows: If the final surface temperature of pixel i and the average final surface temperature of neighboring pixels are both less than or equal to the temperature threshold. If the final surface temperature of pixel i and the average final surface temperature of its neighboring pixels are both within the range of risk, then pixel i is considered risk-free. Within the interval, pixel i is considered to be at a low risk; if the final surface temperature of pixel i and the average final surface temperature of its neighboring pixels are both within the range... Within the interval, pixel i is classified as moderate risk; if the final surface temperature of pixel i and the average final surface temperature of its neighboring pixels are both greater than the threshold temperature... If pixel i is at high risk, then pixel i is at high risk. , These are the temperature thresholds.

3. The method for temperature prediction and fire early warning in exposed coal seam areas of a mining area according to claim 2, characterized in that: It also includes the following methods: S4. All pixels in the study area are stratified and grouped according to fire warning level to form risk-free set, low-risk set, medium-risk set, and high-risk set. Pixels in the low-risk set are continuously monitored and monitored, pixels in the medium-risk set are investigated for potential hazards, and pixels in the high-risk set are given priority for emergency treatment. The temperature prediction model contains a geographic GIS map, and the fire warning level of the pixels is displayed in a stratified and visual manner in the geographic GIS map.

4. The method for temperature prediction and fire early warning in exposed coal seam areas of a mining area according to claim 1, characterized in that: Several ground stations are deployed in the study area, and each ground station is equipped with a temperature detector used to detect and obtain the measured surface temperature corresponding to the ground station. The data is used as ground station temperature detection data; the satellite remote sensing monitoring data is Landsat series satellite remote sensing image data, which includes preprocessing such as radiometric calibration, atmospheric correction, and geometric correction. The remote sensing monitoring data also includes digital elevation model data, and all data in the remote sensing monitoring data are registered to the same coordinate system according to their location.

5. The method for temperature prediction and fire early warning in exposed coal seam areas of a mining area according to claim 1, characterized in that: It also includes the following methods: S5. Construct an ADM-TCN time-series prediction model including multi-scale dilated convolutional blocks and a dynamic memory update module. Following methods S1 and S2, obtain the final land surface temperature time-series data with pixels arranged in a time series. Input the final land surface temperature time-series data into the ADM-TCN time-series prediction model. The multi-scale dilated convolutional blocks capture the short-term, medium-term, and long-term features of the final land surface temperature time-series data and fuse them to obtain comprehensive time-series features. The dynamic memory update module includes an update gate and a reset gate. The update gate determines the number of historical memory states to be retained, and the reset gate controls the influence of historical memory information on the current candidate states. It fuses the filtered historical memory information with the input features to generate candidate memory states. Then, it uses linear interpolation to weightedly fuse the historical memory states and candidate states to form candidate features and dynamically update the memory. The ADM-TCN time-series prediction model outputs the predicted land surface temperature for a future period. A fire warning level is set to determine the fire warning level for the predicted land surface temperature.

6. The method for temperature prediction and fire early warning in exposed coal seam areas of a mining area according to claim 5, characterized in that: The ADM-TCN time-series prediction model is trained using land surface temperature time-series sample data. The ADM-TCN time-series prediction model uses the weighted Huber loss function as the total loss function, as expressed below: ,in The Huber loss function is expressed as follows: , This represents the true surface temperature at a future time m. For the predicted surface temperature at a future time m, For threshold parameters, The time decay weight for future time m, The total future time is predicted by the ADM-TCN time series forecasting model.

7. A method for temperature prediction and fire early warning in exposed coal seam areas of a mining area according to claim 5 or 6, characterized in that: The fire warning levels include no risk, low risk, medium risk, and high risk. The predicted surface temperature of pixel i at a future time m and the average predicted surface temperature of its neighboring pixels are obtained. The fire warning level is determined as follows: if the predicted surface temperature of pixel i at a future time m and the average predicted surface temperature of its neighboring pixels are both less than or equal to a temperature threshold... Then pixel i is risk-free; if the predicted surface temperature of pixel i at future time m and the average predicted surface temperature of its neighboring pixels are both within the range... Within the interval, pixel i is considered to have a mild risk; if the predicted surface temperature of pixel i at future time m and the average predicted surface temperature of its neighboring pixels are both within the range... Within the interval, pixel i is considered to be at medium risk; if the predicted surface temperature of pixel i at a future time m and the average predicted surface temperature of its neighboring pixels are both greater than the threshold temperature... If pixel i is at high risk, then pixel i is at high risk. , These are the temperature thresholds.

8. A method for temperature prediction and fire early warning in exposed coal seam areas of a mining area according to claim 4, characterized in that: The study area includes exposed coal seams and other areas. The ground stations are mainly located in various locations of exposed coal seams in the study area, and some ground stations also cover other areas of the study area.

Images