A high-precision coal calorific value prediction method based on machine learning

Abstract

The present application relates to the technical field of data prediction, and discloses a coal calorific value high-precision prediction method based on machine learning; the method comprises the following steps: obtaining coal quality detection data and performing windowing processing to form a coal calorific value sequence and a multi-source coal quality component sequence; a disturbance embedding vector group is constructed according to adjacent measurement value changes; statistical features and dominant amplitudes of coal quality component variables are extracted, a signal-to-noise factor is constructed, and a selection weight is obtained through mapping to generate a variable-level coal quality feature vector group; the disturbance embedding vector group and the variable-level coal quality feature vector group are combined in time sequence to form a coal quality correlation modeling set, and a sparse correlation structure field is constructed based on a training set; a comprehensive state representation is obtained through correlation propagation and steady-state consistency constraints; a test set is input into the mechanism to obtain a coal calorific value prediction result sequence; the method can strengthen disturbance expression, improve feature fusion quality, and maintain prediction structure stability, and a coal calorific value high-precision prediction result is obtained.

Description

Technical Field

[0001] This invention relates to the field of data prediction technology, and in particular to a high-precision prediction method for coal calorific value based on machine learning. Background Technology

[0002] The calorific value of coal is an important indicator for evaluating coal quality and utilization efficiency, and it is also an important basis for optimizing the operation of coal-fired boilers, regulating coal production, and calculating energy costs. Continuous monitoring of changes in coal composition parameters and calorific value can provide support for fuel blending, production organization, and energy efficiency assessment, thereby improving coal utilization efficiency and combustion stability. Therefore, constructing an accurate and reliable method for predicting the calorific value of coal is of great significance for ensuring the economy and safety of industrial combustion processes.

[0003] To achieve rapid estimation and quality analysis of coal calorific value, researchers have proposed various detection and calculation methods; among existing technologies, patent CN114472205A provides a method that utilizes... A method for determining ash content using X-rays and estimating the calorific value of coal based on the linear relationship between ash content and calorific value involves detecting the transmitted photon signal to obtain the ash content, which is then substituted into a linear model to achieve rapid calculation of the calorific value of coal; patent CN116136505A utilizes... X-ray fluorescence spectrometry was used to determine the C, O, and S content in coal samples, and a multiple linear regression model was constructed to estimate the calorific value of coal, thereby reducing the detection time of traditional oxygen bomb calorimeters. These methods have already shown some effectiveness in the rapid estimation of single batches of samples, providing a foundation for online coal detection.

[0004] However, the aforementioned methods still have shortcomings in terms of multi-source data fusion and prediction capabilities. CN114472205A establishes a linear calorific value model based on ash content as a single variable, which is difficult to reflect different coal types, multi-parameter disturbances, and complex coupling relationships. Although CN116136505A considers components such as C / O / S, the model structure is still a static linear regression, which is difficult to handle nonlinear correlations between multiple components and structural differences between different batches, and lacks dynamic modeling capabilities under continuous time series. Therefore, existing technologies still have problems of insufficient accuracy and poor generalization when dealing with multi-source heterogeneous variables of coal quality, batch disturbance changes, and prediction of future coal calorific value trends. There is an urgent need for a coal calorific value prediction method that can integrate multi-source coal quality component information and combine dynamic evolution laws of disturbance characteristics, and has high-precision prediction capabilities, so as to improve the prediction stability and reliability under complex coal quality conditions. Summary of the Invention

[0005] This invention provides a high-precision coal calorific value prediction method based on machine learning. Firstly, addressing the problems of complex coal composition correlations, insufficient utilization of feature perturbations, and limited prediction accuracy of traditional regression models, a multi-stage coal calorific value prediction scheme integrating perturbation features and multi-source coal composition is proposed. The method includes: windowing coal quality detection data to obtain a windowed coal calorific value sequence and a windowed multi-source coal composition sequence; calculating perturbation amplitude and perturbation relativity control factors to generate a coal calorific value perturbation embedding vector group; extracting statistical features and dominant amplitudes to construct a signal-to-noise factor and weighting it to obtain a variable-level coal quality feature vector group; combining the two types of features to form a coal quality correlation modeling set; constructing a sparse correlation structure field based on the training set; executing correlation propagation and steady-state consistency constraints to form a coal quality correlation evolution prediction mechanism; and inputting the test set into the prediction mechanism to obtain a coal calorific value prediction result sequence, thus achieving high-precision coal calorific value prediction under multi-source fusion, perturbation-driven, and machine learning modeling.

[0006] A high-precision prediction method for coal calorific value based on machine learning includes:

[0007] S1. Obtain coal quality testing data, construct the original coal quality dataset, and perform windowing processing on the original coal quality dataset to obtain the windowed coal calorific value sequence and the windowed multi-source coal quality composition sequence.

[0008] S2. Calculate the perturbation amplitude and perturbation relativity control factor of adjacent time steps of the windowed coal calorific value sequence, and perform expression transformation on the windowed coal calorific value sequence accordingly to obtain the perturbation embedding vector group.

[0009] S3. Extract the statistical features and dominant amplitudes of the windowed multi-source coal quality component sequence, construct the signal-to-noise factor and map it as the signal-to-noise driven selection weight, and then perform weighted combination of the windowed multi-source coal quality component sequence to obtain the variable-level coal quality feature vector group.

[0010] S4. Stack and combine the perturbation embedding vector group and the variable-level coal quality feature vector group in chronological order to obtain the coal quality correlation modeling set, which includes the training set and the test set.

[0011] S5. Based on the training set, a coal quality correlation evolution prediction mechanism is constructed. The coal quality correlation evolution prediction mechanism establishes a sparse correlation structure field based on the training set, performs correlation propagation and steady-state consistency constraints on the sparse correlation structure field, generates a comprehensive state representation, and outputs the coal calorific value prediction result sequence of the training set.

[0012] S6. Input the test set into the coal quality correlation evolution prediction mechanism to obtain the coal calorific value prediction result sequence of the test set.

[0013] Preferably, the steps for constructing the original coal quality dataset and windowing it to obtain the windowed coal calorific value sequence and the windowed multi-source coal quality composition sequence are as follows: First, the calorific value and multi-source coal quality composition of the target coal sample in the detection batch are collected. The multi-source coal quality composition data includes carbon content, hydrogen content, oxygen content, nitrogen content, sulfur content, mineral composition ratio, and ash fusion characteristic parameters. All collected records are labeled with a unified time index and batch index to ensure that the sequence data has a consistent input structure during the modeling stage.

[0014] Subsequently, the raw coal quality test data underwent format standardization processing, mapping the field structures output by different experimental equipment to a unified format; missing records were filled using interpolation based on neighboring samples from the same batch; deviations caused by detection errors were corrected by amplitude limiting and sequence smoothing to ensure the continuity and stability of the coal calorific value sequence and the multi-source coal quality composition sequence within the time domain; for coal quality composition data with significant dimensional differences, interval normalization was performed on all coal quality composition values ​​to ensure that all coal quality composition data are within a uniform numerical range, thereby guaranteeing the stability and comparability of subsequent perturbation amplitude calculation and signal-to-noise factor mapping.

[0015] After data standardization and quality improvement, a sliding window slicing operation is performed on the coal calorific value sequence and the multi-source coal quality composition sequence according to the preset window length P and sliding step size S, dividing the continuous observation sequence into multiple window segments. Each window segment contains continuous coal calorific value observations and corresponding multi-source coal quality composition values ​​for the corresponding time period, thus forming a windowed coal calorific value sequence and a windowed multi-source coal quality composition sequence. The above windowed sequences serve as the input basis for perturbation embedding vector construction and signal-to-noise factor construction, providing structured and continuous feature inputs for subsequent coal quality correlation modeling and coal calorific value prediction mechanisms.

[0016] Preferably, in step S2, the specific process of constructing the perturbation embedding vector set includes:

[0017] In each window of the windowed coal calorific value sequence, the absolute amplitude of the difference between adjacent measurements is taken and statistically analyzed to obtain the perturbation amplitude of each window.

[0018] Calculate the overall average disturbance of the entire set of window disturbance amplitudes, and obtain the relative disturbance coefficient based on the ratio of the current window disturbance amplitude to the overall average disturbance.

[0019] By introducing gating mapping parameters and performing nonlinear mapping on the relative disturbance coefficients, the disturbance gating coefficients are obtained.

[0020] Perform dimensionality-upgrading, nonlinear, and compression representation transformations on the windowed coal calorific value sequence to generate a fixed-length pre-embedded representation;

[0021] The perturbation gating coefficients and the fixed-length pre-embedded representation are scalar-scaled to synthesize the perturbation embedding vector, and the perturbation embedding vectors are then assembled in chronological order to form a perturbation embedding vector group.

[0022] Furthermore, addressing the issues of uneven fluctuation amplitude, local anomalies easily masking structural trends, and insufficient sequence representation dimensions in coal calorific value sequences in coal quality testing data, this invention proposes a perturbation embedding vector group construction method. This method enhances the ability to distinguish key changes and improves the stability of the predicted structure by introducing perturbation intensity measurement and gating adjustment mechanisms. Specifically, it includes the following steps: First, in each window of the windowed coal calorific value sequence, the absolute amplitude of the difference between adjacent measurements is calculated and the perturbation amplitude is statistically obtained to characterize the short-term variation intensity of coal calorific value in continuous measurements; then, the overall average perturbation of the entire set of window perturbation amplitudes is calculated, and using this as a reference, the ratio of the current window perturbation amplitude to the overall average perturbation is used to obtain the relative perturbation coefficient, enabling cross-batch processing. The scale is unified to avoid amplitude shifts caused by different coal sources. Then, a gating mapping parameter is introduced to apply a nonlinear mapping to the relative perturbation coefficient to obtain the perturbation gating coefficient, which compresses the perturbation value to a bounded interval and enhances the structural response under weak perturbation, so that subtle coal quality changes are not drowned out by strong noise. Subsequently, the windowed coal calorific value sequence is subjected to dimensionality-upgrading, nonlinear and compression expression transformation to generate a fixed-length pre-embedded representation, which reduces the interference caused by redundant fluctuations while maintaining the coal calorific value change pattern. Finally, the perturbation gating coefficient and the fixed-length pre-embedded representation are scalar-scaled to synthesize the perturbation embedding vector, and the vectors are aggregated in time order to form a perturbation embedding vector group, so that the dynamic perturbation of coal calorific value is stably expressed in the embedding space and provides structured input for subsequent coal quality correlation modeling.

[0023] Preferably, in step S3, the specific process of constructing the variable-level coal quality feature vector group includes:

[0024] The preset statistical terms of all coal quality component variables in the windowed multi-source coal quality component sequence are aggregated into a windowed statistical feature vector. The preset statistical terms include average level, fluctuation intensity, skewness and kurtosis.

[0025] The dominant amplitudes of all coal quality component variables in the windowed multi-source coal quality component sequence are extracted, and the signal-to-noise factor is constructed based on the dominant amplitudes and the fluctuation intensity in the window statistical feature vector.

[0026] By introducing mapping coefficients and reference levels, the signal-to-noise factor is mapped to an activation factor to obtain the signal-to-noise driven selection weight.

[0027] Based on the signal-to-noise ratio driven selection weight, the window statistical feature vector and the dominant amplitude are weighted and combined to obtain the variable-level coal quality feature vectors of all coal quality components, and then aggregated in chronological order to form a variable-level coal quality feature vector group.

[0028] Furthermore, addressing the issues of dimensional differences, uneven fluctuations in coal quality components, susceptibility of key features to noise interference, and the inadequacy of traditional statistical expressions to characterize structural changes among components in multi-source coal quality composition data across different batches, this invention proposes a variable-level coal quality feature vector group construction method. This method achieves differentiated expression of major coal quality components through a selective enhancement mechanism driven by the signal-to-noise factor. Specifically, it includes the following steps: First, in a windowed multi-source coal quality composition sequence, four statistical indicators—average level, fluctuation intensity, skewness, and kurtosis—are calculated for all coal quality component variables. All statistical items within the same window are then combined to form a windowed statistical feature vector, used to describe the distribution structure and morphological changes of coal quality components within a fixed window. Subsequently, the dominant amplitude of each coal quality component variable within the window is extracted; its value reflects the maximum effective intensity of that coal quality component in the current window. The signal-to-noise factor (SNR) is constructed by combining the dominant amplitude with the corresponding fluctuation intensity of the coal composition, enabling simultaneous characterization of both the amplitude and fluctuation stability of the coal composition. Next, mapping coefficients and reference levels are introduced to activate the SNR factor, resulting in SNR-driven selection weights. The weights of high-SNR coal components are increased while those of low-SNR coal components are suppressed, allowing for differentiated representation of the contribution of different coal components to the calorific value of coal under complex coal quality conditions. Subsequently, based on the SNR-driven selection weights, a weighted combination of the window statistical feature vector and the dominant amplitude is performed to obtain variable-level coal quality feature vectors, enhancing high-quality information and suppressing noise-dominated information shifts. Finally, the variable-level coal quality feature vectors of all windows are aggregated in chronological order to obtain a variable-level coal quality feature vector group, providing a structurally stable input foundation with coal composition discrimination capabilities for subsequent coal quality correlation modeling.

[0029] Preferably, the specific process of obtaining the coal quality correlation modeling set in step S4 includes:

[0030] The endogenous perturbation embedding matrix is ​​obtained by stacking the perturbation embedding vector groups in chronological order.

[0031] The exogenous component signal-noise embedding matrix is ​​obtained by stacking the variable-level coal quality feature vector groups in chronological order.

[0032] The coal quality correlation modeling set is obtained by combining the endogenous perturbation embedding matrix and the exogenous component signal-noise embedding matrix. The coal quality correlation modeling set includes a training set and a test set.

[0033] Furthermore, addressing the issues of strong structural dispersion in coal quality testing data across different batches and combinations of coal composition, difficulty in uniformly representing coal calorific value perturbations and coal composition differences, and the lack of a unified correlation structure in model inputs, this invention proposes a coal quality correlation modeling set construction method in step S4. This method forms a model input structure with temporal consistency by uniformly stacking and jointly organizing two types of features. Specifically, it includes the following steps: First, the perturbation embedding vector groups are stacked in chronological order to generate an endogenous perturbation embedding matrix, ensuring that the perturbation expression of windowed coal calorific value maintains temporal index consistency, thus forming a continuous feature sequence reflecting the dynamic changes in coal calorific value; then, the variable-level coal quality feature vector groups are stacked in chronological order to generate an externally generated sub-signal-noise embedding matrix. A matrix is ​​used to present the statistical structure and signal-to-noise difference of multi-source coal quality components in different windows, so that the influence of coal quality component level can be input into the subsequent modeling mechanism in a stable matrix form. Then, the endogenous perturbation embedding matrix and the exogenous component signal-to-noise embedding matrix are combined according to the time dimension and sample order to form a coal quality correlation modeling set containing the coal calorific value perturbation structure and the coal quality component signal-to-noise structure. This allows the subsequent model to receive both types of features simultaneously and establish cross-window correlation expressions. Finally, according to the sample size and time division strategy, the obtained coal quality correlation modeling set is divided into a training set and a test set to ensure that the correlation structure can be learned in the training phase and the prediction ability can be verified in the testing phase. This provides a structured input basis for subsequent correlation propagation modeling and coal calorific value prediction.

[0034] Preferably, in step S5, the specific process of obtaining the sparse correlation structure field, performing correlation propagation and steady-state consistency constraints on the sparse correlation structure field, generating a comprehensive state representation, and outputting the training set coal calorific value prediction result sequence includes:

[0035] Based on the endogenous perturbation embedding matrix and the exogenous component signal-noise embedding matrix in the training set, the perturbation similarity kernel, the signal-noise contrast kernel and the temporal position coupling kernel are calculated.

[0036] A fusion weight is introduced to perform weighted fusion of the perturbation similarity kernel, signal-to-noise contrast kernel, and temporal-location coupling kernel to obtain a joint perturbation graph structure, and a joint perturbation graph structure matrix is ​​constructed based on the joint perturbation graph structure.

[0037] Normalize the joint perturbation graph structure matrix and perform edge pruning to obtain a sparse correlated structure field;

[0038] The multi-order graph power set of the sparse correlation structure field is calculated, and the tensor perturbation extension is obtained based on the multi-order graph power set and the endogenous perturbation embedding matrix in the training set.

[0039] We obtain a multi-order perturbation map expansion tensor by weighting and superimposing the exponential expansion of the tensor perturbation expansion with factorial decay.

[0040] The average correlation strength of each coal quality component variable across all time windows is calculated based on the joint perturbation graph structure, and the perturbation deviation factor is obtained by comparing the average correlation strength with the joint perturbation graph structure of each window.

[0041] The structural jump control factor is obtained by applying a monotonically bounded fractional mapping to the disturbance deviation factor.

[0042] Under the influence of the structural jump control factor, steady-state consistency alignment is performed between the multi-order perturbation map expansion tensor and the training set endogenous perturbation embedding matrix to obtain the steady-state comprehensive state tensor.

[0043] The mean convergence of the steady-state integrated state tensor yields the fused state vector;

[0044] By introducing prediction mapping parameters and mapping the fusion state vector, a sequence of coal calorific value prediction results for the training set with a prediction step size of F is obtained, and a coal quality correlation evolution prediction mechanism is constructed.

[0045] Furthermore, the specific process of calculating the perturbation similarity kernel, signal-to-noise contrast kernel, and temporal-location coupling kernel includes:

[0046] By introducing a bandwidth parameter, the Euclidean distance between any two window perturbation embedding vectors is calculated in the endogenous perturbation embedding matrix, and the perturbation similarity kernel is obtained based on the Gaussian similarity function.

[0047] In the externally generated signal-to-noise embedding matrix, the L1 norm of the externally generated signal-to-noise embedding matrix is ​​obtained by performing element-wise addition of the coal composition feature vectors of any two windows.

[0048] The signal-to-noise cointegration difference rate of coal quality components was constructed based on the L1 norm;

[0049] By introducing growth adjustment parameters and suppression parameters, a nonlinear mapping is performed on the signal-to-noise cointegration difference rate of coal quality components to obtain a signal-to-noise comparison kernel;

[0050] Extract the range of the perturbation embedding vector in each dimension and construct the perturbation range amplitude tensor; define the position coupling index factor based on the perturbation range amplitude tensor, and construct the fractional decay function based on the position coupling index factor to obtain the time-position coupling kernel.

[0051] Furthermore, addressing the issues of varying perturbation strengths across different windows in coal quality testing data, limited coupling between coal composition signal-to-noise patterns, temporal imbalances leading to unbalanced correlation propagation, and the need for a stable comprehensive state representation during training, this invention proposes a sparse correlation structure field construction method and a correlation propagation mechanism under steady-state consistency constraints in step S5 to ensure the model accurately characterizes cross-window perturbation patterns and multi-source component correlation structures. Specifically, this includes the following steps: First, a bandwidth parameter is introduced into the endogenous perturbation embedding matrix to perform Gaussian similarity mapping on the Euclidean distance between any two window perturbation embedding vectors. The perturbation similarity kernel is obtained to characterize the similarity of the coal calorific value perturbation patterns. Then, in the externally generated component signal-to-noise embedding matrix, element-wise addition of the coal quality component feature vectors of any two windows is performed to obtain the L1 norm of the externally generated component signal-to-noise embedding matrix. Based on the L1 norm, the coal quality component signal-to-noise cointegration difference rate is constructed, and then a logarithmic mapping is performed using growth adjustment parameters and suppression parameters to obtain the signal-to-noise contrast kernel, enabling the synergistic and differential relationships of coal quality component variables to be expressed at a stable scale. Next, the range of the perturbation embedding vector in each dimension is extracted to construct the perturbation range amplitude tensor, and a position coupling index factor is defined. Based on this, a... The fractional decay function yields the temporal-position coupling kernel, allowing the temporal distance between windows and the perturbation amplitude to jointly influence the association strength. Then, a fusion weight is introduced to weight and fuse the perturbation similarity kernel, signal-to-noise contrast kernel, and temporal-position coupling kernel to obtain a joint perturbation graph structure, forming a joint perturbation graph structure matrix. This matrix is ​​then normalized and edge-pruned to obtain a sparse association structure field. Next, the multi-order graph power set of the sparse association structure field is calculated and combined with the endogenous perturbation embedding matrix of the training set to obtain a tensor perturbation expansion. This expansion is then exponentially expanded using factorial decay to form a multi-order perturbation graph expansion tensor. Finally, based on the joint perturbation... The graph structure is used to calculate the average correlation strength of each coal quality component variable in all windows and compare it with the joint perturbation graph structure of each window to obtain the perturbation deviation factor. Then, a bounded fractional mapping is applied to the perturbation deviation factor to generate a structural jump control factor. Under the action of the structural jump control factor, the multi-order perturbation graph expansion tensor and the training set endogenous perturbation embedding matrix are aligned to obtain the steady-state comprehensive state tensor. Finally, the steady-state comprehensive state tensor is averaged and converged to generate a fused state vector. Under the action of the prediction mapping parameter, the prediction result sequence of coal calorific value of the training set with a prediction step size of F is obtained, and a coal quality correlation evolution prediction mechanism is constructed based on this.

[0052] In a second aspect, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the machine learning-based high-precision prediction method for coal calorific value, which is any one of the above-mentioned methods.

[0053] Thirdly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is running, it controls the device on which the computer-readable storage medium is located to execute any of the above-mentioned high-precision prediction methods for coal calorific value based on machine learning.

[0054] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0055] 1. This invention addresses the problem that coal calorific value in coal quality testing data has a large sequence span, significant fluctuations in perturbation amplitude, and local jumps that cannot be effectively characterized by traditional regression models. It constructs a perturbation embedding vector group and introduces a perturbation gating coefficient to structurally express the perturbation differences between windows. Based on dimensionality enhancement, nonlinearity, and compression transformation, this design modulates the pre-embedded representation with perturbation gating coefficients, clearly presenting the strength of perturbations in coal calorific value within the expression space. This improves the ability to identify short-term fluctuations in coal calorific value, addresses the insufficient response of traditional models to local perturbations, and enhances the accuracy and stability of describing coal quality sequence changes.

[0056] 2. This invention addresses the problems of inconsistent fluctuation patterns of multi-source coal quality components within different windows, significant signal-to-noise differences among coal quality component variables, and the difficulty of maintaining structural relationships between variables using traditional weighting methods. It constructs a signal-to-noise factor by using statistical features and dominant amplitudes, and obtains a signal-to-noise driven selection weight after mapping. Based on this weight, the window statistical feature vector and dominant amplitude are weighted and combined to form a variable-level coal quality feature vector group. This mechanism presents the structural differences and dominant contributions among coal quality component variables under a unified expression scale, maintaining clear distinguishability and hierarchy in the coal quality component data fusion process, improving the stability and expression quality of the coal quality feature vectors, and providing reliable input for subsequent correlation modeling.

[0057] 3. This invention addresses the problems of uneven coal composition correlation patterns between windows, difficulty in characterizing perturbation propagation paths by linear models, and insufficient structural consistency in the prediction stage. It constructs a joint perturbation graph structure using a perturbation similarity kernel, a signal-to-noise comparison kernel, and a temporal position coupling kernel. After edge pruning, a sparse correlation structure field is obtained. Subsequently, the perturbation graph is expanded based on a multi-order graph power set. Finally, steady-state consistency alignment is performed using a perturbation deviation factor and a structural jump control factor to obtain a fused state vector. This mechanism describes the correlation propagation path from three dimensions: perturbation intensity, signal-to-noise structure between coal components, and temporal position relationship. This ensures the continuity and stability of the prediction input in terms of structural relationships, improving the overall prediction reliability and expression completeness of the model under complex coal quality conditions, and achieving high-precision prediction of coal calorific value. Attached Figure Description

[0058] Figure 1This is a flowchart of a high-precision prediction method for coal calorific value based on machine learning provided by the present invention.

[0059] Figure 2 This is a structural diagram of the perturbation embedding vector group construction method provided by the present invention.

[0060] Figure 3 This is a structural diagram of the variable-level coal quality feature vector group construction method provided by the present invention.

[0061] Figure 4 This is a structural diagram of the coal quality correlation modeling set construction method provided by the present invention.

[0062] Figure 5 This is a structural diagram of the coal quality correlation evolution prediction mechanism provided by the present invention.

[0063] Figure 6 This is a line graph showing the correlation structure stability score provided by the present invention.

[0064] Figure 7 This is a comparison chart of the predicted calorific value of coal provided by the present invention. Detailed Implementation

[0065] This invention provides a high-precision coal calorific value prediction method based on machine learning. Addressing the problems of complex coal composition correlations, insufficient utilization of feature perturbations, and limited accuracy of traditional regression models in predicting coal calorific value, this invention proposes a multi-stage coal calorific value prediction scheme that integrates perturbation features and multi-source coal composition. The method includes: windowing coal quality inspection data to obtain a windowed coal calorific value sequence and a windowed multi-source coal composition sequence; calculating perturbation amplitude and perturbation relativity control factors to generate a perturbation embedding vector group; extracting statistical features and dominant amplitudes to construct a signal-to-noise factor and weighting it to obtain a variable-level coal quality feature vector group; combining the two types of features to form a coal quality correlation modeling set; constructing a sparse correlation structure field based on the training set; executing correlation propagation and steady-state consistency constraints to form a coal quality correlation evolution prediction mechanism; and inputting the test set into the prediction mechanism to obtain a coal calorific value prediction result sequence, thus achieving high-precision coal calorific value prediction under multi-source fusion, perturbation-driven, and machine learning modeling.

[0066] S1. Obtain coal quality testing data, construct the original coal quality dataset, and perform windowing processing on the original coal quality dataset to obtain the windowed coal calorific value sequence and the windowed multi-source coal quality composition sequence.

[0067] Please see Figure 1As shown in the embodiments of this application, a high-precision prediction method for coal calorific value based on machine learning includes the following steps: First, coal quality testing data from consecutive batches is collected, and the coal calorific value sequence and multi-source coal quality component sequence are obtained respectively, ensuring that all samples are consistent in time index; then, quality screening is performed on the testing data, missing records are filled in, and abnormal offset values ​​are limited and corrected, so that the coal calorific value sequence and multi-source coal quality component sequence remain continuous within the value range; next, normalization processing is performed on coal quality components with large dimensional differences, so that all coal quality component variables have uniform values. The scale facilitates subsequent perturbation amplitude statistics and feature mapping. In the time dimension, all records are rearranged according to the collection order, so that each coal calorific value record is strictly aligned with the corresponding multi-source coal quality composition record. During the window division process, the above sequence is sequentially truncated with a preset window length P and sliding step size S to form time windows of consistent length. Each window contains continuous coal calorific value observations and corresponding multi-source coal quality composition observations, ultimately forming two windowed input sequences arranged in chronological order, which serve as inputs for subsequent perturbation embedding construction and variable-level coal quality feature embedding calculation, respectively.

[0068] It should be noted that the specific process of windowing the original coal quality dataset to obtain the windowed coal calorific value sequence and the windowed multi-source coal quality composition sequence includes: setting the window length parameter P and the sliding step size parameter S, and calculating the total number of windows based on the total length T of all time series in the original coal quality dataset. ,in, Indicates rounding down; determines the starting index of each window. And, under the starting index, extract observation segments with a length of window length parameter P, where, For window indexing; during the extraction process, the coal calorific value sequence in the original coal quality dataset is divided into window segments to obtain the windowed coal calorific value sequence. The original coal quality dataset is divided into window segments for each coal quality component variable j in the multi-source coal quality component sequence, resulting in a windowed multi-source coal quality component sequence. ;

[0069] S2. Calculate the perturbation amplitude and perturbation relativity control factor of adjacent time steps of the windowed coal calorific value sequence, and perform expression transformation on the windowed coal calorific value sequence accordingly to obtain the perturbation embedding vector group.

[0070] Furthermore, in step S2, the perturbation embedding vector set is obtained, and its flowchart is as follows. Figure 2 As shown, the specific steps to obtain the perturbation embedding vector group are as follows: In the windowed coal calorific value sequence In each window, the absolute magnitude of the difference between adjacent measurements is taken. The data was then statistically analyzed to obtain the disturbance amplitude for each window. ; Calculate the overall average perturbation of the entire set of window perturbation amplitudes. The relative disturbance coefficient is obtained based on the ratio of the current window disturbance amplitude to the overall average disturbance. In this embodiment Introducing gating mapping parameters and Furthermore, a nonlinear mapping is performed on the relative disturbance coefficients to obtain the disturbance gating coefficients. In this embodiment, the nonlinear mapping performed on the relative perturbation coefficient is a hyperbolic tangent mapping, wherein, The gating mapping parameters are learnable parameters, with an initial value of , ; Calcium wattage sequence of coal Execution Upgrade Nonlinear With compression The expression transformation generates a fixed-length pre-embedded representation. ,in, , , , For learnable parameters, , All are initialized to 0. , The Xavier-Uniform uniform distribution is used for initialization; the perturbation gating coefficients and the fixed-length pre-embedded representation are synthesized by scalar scaling to obtain the perturbation embedding vector. The perturbation embedding vectors are then aggregated in chronological order to form a perturbation embedding vector group. .

[0071] It should be noted that Xavier-Uniform uniform distribution initialization refers to automatically calculating the value range based on the input dimension P and output dimensions h and d, and then setting the weight parameters accordingly. , Randomly initialized to , ,in This indicates uniform sampling with symmetrical upper and lower bounds centered at zero. This initialization method ensures that the variance of the linear layer output matches the variance of the input, mitigates the gradient vanishing problem, and improves the numerical stability and training convergence speed of the dimensionality-upgrading and compression mapping.

[0072] S3. Extract the statistical features and dominant amplitudes of the windowed multi-source coal quality component sequence, construct the signal-to-noise factor and map it as the signal-to-noise driven selection weight, and then perform weighted combination of the windowed multi-source coal quality component sequence to obtain the variable-level coal quality feature vector group.

[0073] Furthermore, in step S3, the variable-level coal quality feature vector set is obtained, and its flowchart is as follows. Figure 3 As shown, the specific steps to obtain the variable-level coal quality feature vector group are as follows: Calculate the windowed multi-source coal quality composition sequence. All coal quality component variables The preset statistical items are aggregated into a window statistical feature vector. The preset statistical items include the average level. Fluctuation intensity skewness With peak state In this embodiment Extract the dominant amplitudes of all coal quality component variables from the windowed multi-source coal quality component sequence. The signal-to-noise factor is constructed based on the dominant amplitude and the fluctuation intensity in the window statistical feature vector. Introducing mapping coefficient α and reference level The signal-to-noise ratio (SNR) factor is mapped to the SNR factor to obtain the SNR-driven selection weights. , where the mapping coefficient Used to control the smoothness level, in this embodiment , The total number of coal quality composition variables; weights are selected based on signal-to-noise ratio. By weighting and combining the window statistical feature vector with the dominant amplitude, we obtain the variable-level coal quality feature vector for all coal quality components. They are then compiled in chronological order to form a variable-level coal quality feature vector group. .

[0074] S4. Stack and combine the perturbation embedding vector group and the variable-level coal quality feature vector group in chronological order to obtain the coal quality correlation modeling set, which includes the training set and the test set.

[0075] Furthermore, in step S4, the coal quality correlation modeling set is obtained, and its flowchart is as follows: Figure 4 As shown, the specific steps to obtain the coal quality correlation modeling set are as follows: For the perturbation embedding vector group... The endogenous perturbation embedding matrix is ​​obtained by stacking them in chronological order. ; For the variable-level coal quality feature vector group The externally generated signal-to-noise embedding matrix is ​​obtained by stacking the matrices in chronological order. The coal quality correlation modeling set is obtained by combining the endogenous perturbation embedding matrix and the exogenous component signal-noise embedding matrix. The coal quality correlation modeling set includes a training set and a test set.

[0076] S5. Based on the training set, a coal quality correlation evolution prediction mechanism is constructed. The coal quality correlation evolution prediction mechanism establishes a sparse correlation structure field based on the training set, performs correlation propagation and steady-state consistency constraints on the sparse correlation structure field, generates a comprehensive state representation, and outputs the coal calorific value prediction result sequence of the training set.

[0077] Furthermore, in step S5, a coal quality-related evolution prediction mechanism is constructed, the flowchart of which is as follows: Figure 5 As shown, the specific steps for constructing a coal quality correlation evolution prediction mechanism are as follows: Based on the endogenous perturbation embedding matrix in the training set... With the externally generated signal-to-noise embedding matrix Calculate the perturbation similarity kernel Signal-to-noise ratio comparison kernel Coupled with time and position kernel ,in, Indicates the first A time window segment, Indicates the first A time window segment, , They all belong to sets The corresponding time segment index of the sample; introduce fusion weights. , Perturbation similarity kernel Signal-to-noise ratio comparison kernel Coupled with time and position kernel The joint perturbation graph structure is obtained by weighted fusion. And based on the joint perturbation graph structure, a joint perturbation graph structure matrix is ​​constructed, in which the fusion weights are... , Based on empirical settings, considering the embedding matrix due to endogenous perturbations... It can directly characterize the temporal variation of coal calorific value, determine the dominant topology of the correlation field, and double the weight can highlight the contribution of perturbation features to the propagation of coal quality state, and enhance the sensitivity to short-term trend changes. In this embodiment, it is set as follows: Simultaneous settings Maintaining the standard weights of the temporal and positional coupling kernels allows them to smooth the association strength between neighboring windows and maintain the coherence of the graph structure, while preventing non-physical connections from forming in distant windows due to excessively high weights, thus jointly perturbing the graph structure. It can be viewed as the first in the joint perturbation graph structure matrix. Each element represents a window. and window The structural connection strength between them;

[0078] Normalization and edge pruning are performed on the joint perturbation graph structure matrix to obtain a sparse correlated structure field. The edge cropping operation will be below the preset threshold. The correlation is set to zero, and the preset threshold is used. Based on experience;

[0079] It should be noted that, in this embodiment, to balance computational efficiency and network sparsity, the threshold is... Based on experience and combined with sample statistics, it is set as a linear combination of the mean and standard deviation of the elements of the joint correlation matrix. ,in , , This is the adjustment coefficient; verified by multiple sets of samples, when Stable performance can be achieved in this way. This embodiment takes ;

[0080] Computing sparse correlation structure fields Multi-order graph power set And based on the multi-order graph power set and the endogenous perturbation embedding matrix in the training set, Tensor perturbation extension ,in Represents a sparse correlation field of The power, i.e., the power over the sparse correlation field conduct The propagation is performed using the next multiplication operation, and the number of multiplication operations is... , For the maximum propagation order, first-order diffusion captures direct adjacency relationships, second-order diffusion introduces indirect adjacency information, and third-order diffusion can perceive mid-range temporal dependencies; when When the diffusion range is too large, the global smoothing effect is enhanced, the local disturbance signal is diluted, and the prediction accuracy decreases. In this embodiment, a maximum propagation order is set. ; Extension of tensor perturbation Weighted summation of exponential expansions with factorial decay yields the multi-order perturbation spectrum expansion tensor. Based on joint perturbation graph structure Calculate the average correlation strength of each coal quality component variable across all time windows. The perturbation deviation factor is obtained by comparing the average correlation strength with the joint perturbation map structure of each window. Applying a monotonically bounded fractional mapping to the disturbance deviation factor yields the structural jump control factor. ; In the structural jump control factor, the structural jump control factor is described Under the influence of the perturbation, the tensor is expanded on the multi-order perturbation spectrum. With the training set endogenous perturbation embedding matrix Perform steady-state consistency alignment to obtain the steady-state comprehensive state tensor. ; for steady-state comprehensive state tensor Performing mean convergence yields the fused state vector. Introducing prediction mapping parameters , And the fusion state vector By performing mapping, a sequence of predicted coal calorific value results for the training set with a prediction step size of F is obtained. Finally, a coal quality correlation evolution prediction mechanism was constructed to predict mapping parameters. , Here are the learning parameters, where Initialize to 0, The Xavier-Uniform uniform distribution initialization was used.

[0081] It should be noted that the specific process of calculating the perturbation similarity kernel, signal-to-noise contrast kernel, and temporal-position coupling kernel in step S5 includes:

[0082] Introducing bandwidth parameters The intrinsic perturbation embedding matrix in the training set In this process, the Euclidean distance between the perturbation embedding vectors of any two windows is calculated. Using the Euclidean distance as input, the perturbation similarity kernel is obtained based on a Gaussian similarity function. Among them, bandwidth parameter The larger the value, the less the weights will decay rapidly, even if the Euclidean distance between the perturbation embedding vectors of any two windows is large. The smaller the value, the less likely there will be strong connections between nodes that are slightly further apart. In this embodiment... , The number of endogenous variables in this example. The only variable is the calorific value of coal;

[0083] In the externally generated sub-signal embedding matrix, the L1 norm of the externally generated sub-signal embedding matrix is ​​obtained by performing element-wise addition of the coal quality component eigenvectors of any two windows.

[0084] The signal-to-noise cointegration difference rate of coal quality components was constructed based on the L1 norm. ;

[0085] Introducing growth regulation parameters and inhibition parameters , The signal-to-noise comparison kernel is obtained by nonlinearly mapping the signal-to-noise cointegration difference rate of coal quality components. ;in, This is used to regulate the initial growth rate, ensuring that samples with high cointegration difference rates receive stronger weighted responses in the kernel function, thereby highlighting the impact of significant differences in multi-source coal composition on the correlation structure. In this embodiment... It can achieve acceleration but avoids excessive amplification. Used to adjust the intensity of distal inhibition in this embodiment This ensures that the suppression strength is inversely proportional to the difference rate, guaranteeing that long-distance differences do not dominate the global topology and reducing the interference of noise windows on the propagation path;

[0086] Extract the range of the perturbation embedding vector in each dimension and construct the perturbation range magnitude tensor. Define a position coupling exponential factor based on the perturbation range amplitude tensor. Based on the position coupling exponential factor, a fractional decay function is constructed to obtain the time-position coupling kernel. .

[0087] S6. Input the test set into the coal quality correlation evolution prediction mechanism to obtain the coal calorific value prediction result sequence of the test set.

[0088] Furthermore, in step S6, the test set is input into the coal quality correlation evolution prediction mechanism to obtain the coal calorific value prediction result sequence of the test set. ,in, Indicates the first One prediction window, This indicates the prediction step size within the window.

[0089] To address the significant differences in the strength of coal calorific value perturbations, the complex signal-to-noise structure of multi-source coal quality components, and the evolution of inter-window correlations in coal quality testing data across multiple batches, this invention constructs a coal calorific value prediction model based on perturbation embedding vector groups and variable-level coal quality feature vector groups. During the training phase, the windowed coal calorific value sequence and the windowed multi-source coal quality component sequence are used as joint inputs. The model sequentially performs perturbation amplitude statistics and gating mapping, signal-to-noise factor extraction and weight selection, construction of endogenous perturbation embedding matrices and exogenous component signal-to-noise embedding matrices, and generation and sparsification of the joint perturbation graph structure. The model is developed through multiple stages, including the formation of the correlation structure field, the expansion of multi-order perturbation maps and the alignment with steady-state consistency, and the output of the fused state vector mapping. By minimizing the coal calorific value prediction error, the perturbation structure preservation error, the signal-noise correlation preservation error and the structure jump control regularization term, the model is guided to learn the perturbation embedding expression, the signal-noise driven selection mechanism, the cross-window correlation structure propagation path and the prediction mapping parameters during the iterative update process. After training convergence, a coal calorific value prediction model is obtained that can stably characterize the trend of coal calorific value changes under complex coal quality conditions and has high-precision prediction capabilities.

[0090] The proposed high-precision coal calorific value prediction model based on machine learning is implemented in Python. It utilizes the PyTorch deep learning framework to construct a perturbation embedding module, a signal-to-noise feature modeling module, and a coal quality correlation evolution prediction module. The model input is tensor data stacked by time windows, with joint encoding along the channel dimension. The experimental environment is configured with an Intel i9 CPU, 64GB of RAM, and an NVIDIA RTX 3090 GPU. The coal quality correlation modeling set is divided into training and validation sets in a 7:3 ratio. The optimizer uses the Adam algorithm, with an initial learning rate of 0.001 dynamically adjusted using a segmented decay strategy during training. The batch size is set to 32, and the total number of training rounds is set to 180. The loss function uses a weighted sum of calorific value prediction error, perturbation structure preservation loss, signal-to-noise correlation preservation loss, and steady-state consistency regularization term during backpropagation to ensure that the perturbation embedding expression, signal-to-noise driving weights, and joint perturbation graph structure converge collaboratively during training and achieve stable prediction performance.

[0091] Furthermore, the collected coal quality testing data is input into the constructed coal quality correlation evolution prediction model for training and inference. The model's correlation structure stability score during the training phase is as follows: Figure 6 As shown in the figure, as the number of training rounds gradually increases, the stability score of the association structure gradually improves from a low level in the initial stage along a identifiable trend, and then tends to be relatively stable in the middle and later stages. This change reflects that under the combined effect of the perturbation similarity kernel, the signal-to-noise contrast kernel, and the temporal coupling kernel, the convergence process of the internal structural association of the model is relatively smooth, the structural fluctuations in the training stage are effectively controlled, and the association propagation path gradually stabilizes; the prediction effect of coal calorific value is as follows. Figure 7 As shown in the figure, the horizontal axis represents the time series window index, and the vertical axis represents the coal calorific value, reflecting the actual changes of the test data over time. The figure shows that the actual coal calorific value sequence exhibits slight local fluctuations and jumps in some windows, reflecting the natural perturbation characteristics of coal quality testing data in real-world scenarios. The predicted sequence given by the model generally matches the actual sequence's trajectory well, maintaining stable following within fluctuation ranges and showing reasonable response adjustments near jump points. The predicted curve does not show significant drift, indicating that the proposed coal quality correlation evolution prediction mechanism can maintain good structural consistency and prediction continuity under the joint constraints of perturbation-driven and multi-source coal quality component fusion, effectively expressing the actual coal calorific value variation law. Experimental results verify the applicability and stability of the model in coal calorific value prediction tasks.

[0092] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a coal quality correlation evolution prediction mechanism program. When the processor executes the computer program, it implements the steps in the above-described embodiments of the high-precision prediction method for coal calorific value based on machine learning, for example... Figure 1 Step S1 is shown.

[0093] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A high-precision prediction method for coal calorific value based on machine learning, characterized in that, include: Coal quality testing data is obtained, an original coal quality dataset is constructed, and the original coal quality dataset is windowed to obtain a windowed coal calorific value sequence and a windowed multi-source coal quality composition sequence. The absolute amplitude of the difference between the measured values ​​at adjacent times within each window of the windowed coal calorific value sequence is statistically analyzed to obtain the perturbation amplitude. The relative perturbation coefficient is obtained based on the ratio of the perturbation amplitude of each window to the average perturbation of all windows. The windowed coal calorific value sequence is then subjected to dimensionality-upgrading, nonlinear, and compression transformations based on the relative perturbation coefficient to obtain a perturbation embedding vector group. Calculate the preset statistical terms of all coal quality component variables in the windowed multi-source coal quality component sequence, and aggregate the preset statistical terms into a window statistical feature vector. Extract the maximum amplitude of all coal quality component variables in the windowed multi-source coal quality component sequence within the corresponding window as the dominant amplitude. Construct a signal-to-noise factor based on the dominant amplitude and the fluctuation intensity in the window statistical feature vector, and map the signal-to-noise factor to a signal-to-noise driven selection weight. Based on the signal-to-noise driven selection weight, perform a weighted combination of the window statistical feature vector and the dominant amplitude to form a variable-level coal quality feature vector group. The perturbation embedding vector group and the variable-level coal quality feature vector group are stacked and combined in chronological order to obtain a coal quality correlation modeling set, which includes a training set and a test set. Based on the training set, a coal quality correlation evolution prediction mechanism is constructed. The coal quality correlation evolution prediction mechanism obtains a joint perturbation graph structure by weighted fusion of perturbation similarity kernel, signal-to-noise comparison kernel and time-position coupling kernel. Based on the joint perturbation graph structure, a joint perturbation graph structure matrix is ​​constructed. The joint perturbation graph structure matrix is ​​normalized and edge pruning is performed to obtain a sparse correlation structure field. The correlation propagation and steady-state consistency constraints are performed on the sparse correlation structure field to generate a comprehensive state representation and output the coal calorific value prediction result sequence of the training set. The test set is input into the coal quality correlation evolution prediction mechanism to obtain the coal calorific value prediction result sequence of the test set.

2. The high-precision prediction method for coal calorific value based on machine learning according to claim 1, characterized in that, The process of constructing the perturbation embedding vector set includes: In each window of the windowed coal calorific value sequence, the absolute value of the difference between adjacent measurements is statistically analyzed to obtain the perturbation amplitude of each window. Calculate the overall average perturbation of the entire set of window perturbation amplitudes, and obtain the relative perturbation coefficient, which characterizes the degree of deviation of the current window perturbation intensity from the overall average level, based on the ratio of the current window perturbation amplitude to the overall average perturbation. A gated mapping parameter is introduced to adjust the mapping amplitude and offset of the relative disturbance coefficient, and the relative disturbance coefficient is nonlinearly mapped to obtain the disturbance gated coefficient. The windowed coal calorific value sequence is subjected to dimensionality-upgrading, nonlinear activation, and compression representation transformations to generate a fixed-length pre-embedded representation with a fixed dimension. The perturbation gating coefficients and the fixed-length pre-embedded representation are scalar-scaled and synthesized to obtain the perturbation embedding vectors corresponding to each window. The perturbation embedding vectors are then aggregated in chronological order to form a perturbation embedding vector group.

3. The high-precision prediction method for coal calorific value based on machine learning according to claim 1, characterized in that, The preset statistical items include average level, fluctuation intensity, skewness, and kurtosis; Mapping the signal-to-noise factor to a signal-to-noise driven selection weight includes: introducing a mapping coefficient and a reference level to adjust the smoothness of the mapping, performing activation factor mapping on the signal-to-noise factor, and obtaining a signal-to-noise driven selection weight to characterize the retention degree of each coal quality component variable. Based on this weight, the window statistical feature vector and the dominant amplitude are weighted and combined to form a variable-level coal quality feature vector group, including: weighting and combining the window statistical feature vector and the dominant amplitude according to the signal-noise driven selection weight to obtain the variable-level coal quality feature vectors of all coal quality component variables, and collecting them in chronological order to form a variable-level coal quality feature vector group.

4. The high-precision prediction method for coal calorific value based on machine learning according to claim 1, characterized in that, The construction process of the coal quality correlation modeling set includes: The endogenous perturbation embedding matrix is ​​obtained by stacking the perturbation embedding vector groups in chronological order. The externally generated signal-to-noise embedding matrix is ​​obtained by stacking the variable-level coal quality feature vector groups in chronological order; The coal quality correlation modeling set is obtained by combining the endogenous perturbation embedding matrix and the exogenous component signal-noise embedding matrix, and the training set and test set are divided based on the coal quality correlation modeling set.

5. The high-precision prediction method for coal calorific value based on machine learning according to claim 4, characterized in that, The construction process of the sparse correlation structure field includes: Based on the endogenous perturbation embedding matrix and the exogenous component signal-noise embedding matrix in the training set, the perturbation similarity kernel, the signal-noise contrast kernel and the time-position coupling kernel are calculated according to the distance relationship between any two window perturbation embedding vectors, the signal-noise difference relationship between coal quality component feature vectors and the time position difference relationship between windows, respectively. A fusion weight is introduced to perform weighted fusion of the perturbation similarity kernel, the signal-to-noise contrast kernel, and the temporal position coupling kernel to obtain a joint perturbation graph structure that characterizes the structural connection strength between any two windows, and a joint perturbation graph structure matrix is ​​constructed based on the joint perturbation graph structure. The joint perturbation graph structure matrix is ​​normalized and edge pruning is performed to obtain a sparse correlation structure field.

6. The high-precision prediction method for coal calorific value based on machine learning according to claim 5, characterized in that, The calculation process of the perturbation similarity kernel, the signal-to-noise comparison kernel, and the temporal position coupling kernel specifically includes: A bandwidth parameter is introduced to control the distance decay rate. The Euclidean distance between any two window perturbation embedding vectors is calculated in the endogenous perturbation embedding matrix, and the perturbation similarity kernel is obtained based on the Gaussian similarity function. In the externally generated signal-to-noise embedding matrix, element-wise addition and element-wise subtraction operations are performed on the coal quality component feature vectors of any two windows, respectively. The L1 norm of the addition result and the L1 norm of the subtraction result are calculated, and the coal quality component signal-to-noise cointegration difference rate is constructed based on the ratio of the L1 norm of the addition result to the L1 norm of the subtraction result. By introducing a growth adjustment parameter for adjusting the intensity of the initial growth and a suppression parameter for controlling the intensity of the distal suppression, a nonlinear mapping is performed on the signal-to-noise cointegration difference rate of the coal composition to obtain a signal-to-noise comparison kernel; Extract the amplitude difference between the maximum and minimum differences of any two window perturbation embedding vectors in each dimension to construct a perturbation range amplitude tensor; define a position coupling index factor to adjust the degree of influence of window position differences based on the perturbation range amplitude tensor, and construct a fractional decay function that decays as the window spacing increases based on the position coupling index factor to obtain the time-position coupling kernel.

7. The high-precision prediction method for coal calorific value based on machine learning according to claim 5, characterized in that, The specific process of performing correlation propagation and steady-state consistency constraints on the sparse correlated structure field includes: The graph power results of the sparse correlation structure field under different propagation orders are calculated to obtain a multi-order graph power set. Based on the multi-order graph power set and the endogenous perturbation embedding matrix in the training set, a propagation operation is performed to obtain the tensor perturbation expansion. The tensor perturbation expansion is weighted and superimposed according to the factorial decay weight corresponding to the propagation order to obtain a multi-order perturbation map expansion tensor; The average correlation strength of the nodes corresponding to each time window across all windows is calculated based on the joint perturbation graph structure, and the average correlation strength is compared with the joint perturbation graph structure of the corresponding window to obtain the perturbation deviation factor. By applying a monotonically bounded fractional mapping to the disturbance deviation factor, a structural jump control factor is obtained to adjust the degree of structural state jump. Under the influence of the structural jump control factor, steady-state consistency alignment is performed on the multi-order perturbation map expansion tensor and the training set endogenous perturbation embedding matrix to obtain a steady-state comprehensive state tensor that takes into account both the propagation results and the stability of the original perturbation embedding.

8. The high-precision prediction method for coal calorific value based on machine learning according to claim 7, characterized in that, The specific process of generating a comprehensive state representation and outputting the training set coal calorific value prediction result sequence includes: The steady-state integrated state tensor is subjected to mean convergence along the time window dimension to obtain the fused state vector; A prediction mapping parameter is introduced to map the fused state vector to the predicted coal calorific value, and the fused state vector is linearly mapped to obtain a training set of predicted coal calorific value results with a prediction step size of F. A coal quality correlation evolution prediction mechanism is then constructed.

Images