Test-time adaptive fine coal ash fraction prediction method with hidden space concept drift assessment

Abstract

The application discloses a test-time adaptive fine ash content prediction method for latent space concept drift evaluation, and belongs to the technical field of computers. The application aims to improve the generalization ability and robustness of an artificial intelligence model under non-stationary conditions of heavy medium coal preparation. The application constructs a sample-offset sample pair modeling mechanism driven by time shift, designs a double-input latent space representation network with LSTNet as the backbone, and realizes direct estimation of the dynamic offset of fine ash content based on the difference between the representations of two time points. The application fuses target offset loss and latent space reconstruction constraints to construct a composite loss function, thereby reducing the excessive dependence of the model on a single data distribution by enhancing the stability of the representation space structure. The test-time adaptive strategy proposed by the application can evaluate and compensate for concept drift under new conditions without relying on large-scale retraining, thereby improving the accuracy, stability and reliability of ash content prediction under unknown or extreme conditions.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to a test-time adaptive method for predicting the ash content of clean coal in latent space concept drift assessment. Background Technology

[0002] Ash content in clean coal is a core indicator for evaluating the quality of coal products. Effective control and accurate prediction of ash content in clean coal are of paramount importance for improving product quality, optimizing resource utilization, reducing production costs, and enhancing production stability.

[0003] Current technologies for determining the ash content of clean coal mainly rely on laboratory chemical analysis methods, which calculate the ash content by burning coal samples and measuring the mass of residual minerals. While this method offers high accuracy, it is complex, time-consuming, and exhibits significant time lag, making it difficult to meet the demands of modern coal preparation systems for real-time online monitoring and rapid control.

[0004] Existing data-driven prediction methods are mostly based on direct modeling of collected samples, but this method faces three challenges: First, the internal mechanism of the model is difficult to explain, resulting in a serious "black box" problem; second, the non-stationary nature of the heavy media coal preparation process makes it difficult for the model to guarantee prediction accuracy in the early stages of changes in operating conditions and when the amount of data is insufficient; third, there are significant differences in coal quality characteristics and process parameters among different coal preparation plants, making it difficult to directly transfer and apply high-precision models trained in a single scenario.

[0005] Concept drift, or the change in the mapping relationship between input features and target variables, is an important cause of the above problems; the same process parameters may lead to drastically different sorting results in different coal preparation plants, which directly restricts the model's cross-scenario generalization ability. Summary of the Invention

[0006] To address the aforementioned technical problems in existing technologies, this invention proposes a test-time adaptive method for predicting the ash content of clean coal using latent space concept drift evaluation. By constructing a time-shift driven sample-offset sample pair drift pattern for learning samples, it achieves accurate prediction of the ash content of clean coal, solving the problems of traditional methods being sensitive to changes in operating conditions and having insufficient cross-scenario transfer performance.

[0007] To achieve the above objectives, the present invention adopts the following technical solution;

[0008] A test-time adaptive method for predicting the ash content of clean coal using latent space concept drift assessment, which utilizes sample-offset sample pairs to achieve accurate prediction of the target ash content offset, includes the following steps;

[0009] Step 1: Collect heavy media coal preparation data and construct a coal preparation dataset;

[0010] Step 2: Preprocess the collected data and divide it into training and test sets;

[0011] Step 3: Construct time-shift driven sample-offset sample pairs;

[0012] Step 4: Extract latent space features of samples and offset samples based on the LSTNet model;

[0013] Step 5: Calculate the reconstruction loss and target offset loss using the obtained latent space features;

[0014] Step 6: Evaluation and application of the model.

[0015] Preferably, in step 1, the key process parameters collected include at least the separation density, coal slime content, magnetic material content, raw coal data, hydrocyclone inlet pressure and combined tank liquid level, as well as the corresponding clean coal ash content index.

[0016] Preferably, in step 2, the preprocessing includes at least sensor error correction and missing value processing. After preprocessing, the dataset is divided into a training set and a test set in a 9:1 ratio for subsequent model training and performance evaluation.

[0017] Preferably, in step 3, the time offset is designed using the preprocessed data, and the offset sample is extracted based on the sample and the time offset to construct a time shift driven sample-offset sample pair. The construction process is shown in formula (1).

[0018] (1);

[0019] in, Indicates the offset sample. This represents the sample at the current moment. The time offset represents the sample-offset sample pair in the form of... .

[0020] Preferably, in step 4, latent space features in the samples and offset samples are extracted based on the LSTNet model. The LSTNet model consists of three stages: a convolutional neural network layer, a recurrent neural network layer, and a recurrent skip neural network layer. The specific steps are as follows.

[0021] Step 4.1: Input the obtained sample-offset sample pairs into the model and use a convolutional neural network to extract their local features;

[0022] Step 4.2: Use a recurrent neural network to model short-term dependencies on the local features extracted by the convolutional neural network.

[0023] Step 4.3: Use a recurrent skip neural network to model long-term dependencies on the local features extracted by the convolutional neural network.

[0024] Step 4.4: Combine the short-term dependency modeling implemented by the recurrent neural network with the long-term dependency modeling implemented by the recurrent skip neural network to form the latent space features of the sample.

[0025] Preferably, the LSTNet network structure for samples The calculation process is as follows;

[0026] (2);

[0027] (3);

[0028] (4);

[0029] (5);

[0030] in, Indicates the input sample. This represents the local features extracted by the convolutional neural network. This represents short-term dependencies modeled using recurrent neural networks. This indicates long-term dependencies modeled using a recurrent skip neural network. This represents a two-dimensional convolution operation. Indicates the GRU layer. This indicates skipping GRU layers. This indicates a splicing operation. This indicates the LSTNet encoder.

[0031] Preferably, in step 5, the target offset loss calculation formula is as shown in formula (6) and formula (7);

[0032] (6);

[0033] (7);

[0034] in, Indicates target offset loss. Indicates L1 loss, Represents the total number of samples. and Let represent the two input parameters in the L1 loss, and represent the two vectors that need to be compared. and Representing vectors respectively The first in elements and vectors The first in One element, and These represent the sample at the current time and the offset sample, respectively. and These represent the current clean coal ash content and the offset clean coal ash content, respectively.

[0035] Preferably, the reconstruction loss calculation formula is as shown in formulas (8) to (11);

[0036] (8);

[0037] (9);

[0038] (10);

[0039] (11);

[0040] in, and These represent the addition of two different types of random noise. and These represent the input samples respectively. The output values ​​of the LSTNet encoder after applying different noise levels. Indicates the mean square error loss. and These represent the two input parameters for the mean squared error loss. and Representing vectors respectively The first in elements and vectors The first in One element, Indicates the reconstruction loss. Indicates SIGReg loss. and These represent the weights of the mean squared error loss and the SIGReg loss, respectively.

[0041] Preferably, the final loss is as shown in formula (12);

[0042] (12);

[0043] in, Indicates the final loss. This represents the weight value corresponding to the reconstruction loss.

[0044] Preferably, in step 6, in practical applications, the model adds the offset output by the model at the current moment to the ash content of the clean coal at the previous moment to obtain the accurate ash content value at the current moment.

[0045] The beneficial technical effects brought about by this invention;

[0046] 1. Enhanced model interpretability. Traditional clean coal ash content prediction models struggle to trace the specific causes of fluctuations in predicted values, limiting their practical guidance in actual production. This invention, by predicting the dynamic change patterns of clean coal ash content, visually presents the key factors influencing ash content fluctuations, significantly improving the model's interpretability and thus enhancing its guidance for on-site operations and process adjustments.

[0047] 2. Enhanced proactive adaptability to non-stationary processes. When concept drift occurs, traditional models often require accumulating a large number of new samples to complete parameter adjustments, making it difficult to guarantee prediction accuracy in the early stages of drift. This invention extends the modeling object from static mapping to dynamic learning of drift patterns. Through forward-looking prediction of drift trends, it achieves rapid response to new operating conditions, significantly improving the accuracy and stability of ash content prediction during periods of changing operating conditions, bringing better economic benefits to coal preparation plants.

[0048] 3. Achieving cross-condition knowledge transfer and few-sample learning. Traditional methods are limited by the significant differences in coal quality characteristics, equipment parameters, and process conditions among different coal preparation plants, making effective cross-scenario transfer difficult. This invention utilizes the similarity of the clean coal ash drift mechanism across scenarios, and by modeling the inherent laws of ash change, it achieves model adaptation with only a small number of samples under new operating conditions. This feature effectively improves the model's generalization ability and transfer efficiency, providing strong support for the intelligent promotion and deployment of coal preparation plants. Attached Figure Description

[0049] Figure 1 This is a flowchart of the test-time adaptive clean coal ash content prediction method for implicit space concept drift assessment of the present invention.

[0050] Figure 2 This is a structural diagram of the LSTNet model used in this invention.

[0051] Figure 3 This is a diagram of the GRU structure in the LSTNet model used in this invention.

[0052] Figure 4 This is a graph showing the experimental results of different weight combinations of the present invention on the test set.

[0053] Figure 5 This is a flowchart of the present invention for calculating the ash content of clean coal at the current moment. Detailed Implementation

[0054] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0055] A test-time adaptive method for predicting clean coal ash content based on implicit space concept drift assessment, the process of which is as follows: Figure 1 As shown; the accurate prediction of the target ash offset is achieved using sample-offset sample pairs, including the following steps;

[0056] Step 1: Collect heavy media coal preparation data and construct a coal preparation dataset;

[0057] The key process parameters collected include at least the separation density, coal slime content, magnetic content, raw coal data, hydrocyclone inlet pressure and combined tank liquid level, as well as the corresponding clean coal ash content index.

[0058] Step 2: Preprocess the collected data and divide it into training and test sets;

[0059] The preprocessing includes at least sensor error correction and missing value handling. After preprocessing, the dataset is divided into training and testing sets in a 9:1 ratio for subsequent model training and performance evaluation.

[0060] Step 3: Construct time-shift driven sample-offset sample pairs;

[0061] Using the preprocessed data, a time offset is designed, and offset samples are extracted based on the samples and the time offset to construct a time shift driven sample-offset sample pair. The construction process is shown in formula (1).

[0062] (1);

[0063] in, Indicates the offset sample. This represents the sample at the current moment. The time offset represents the sample-offset sample pair in the form of... .

[0064] Step 4: Extract latent space features from samples and offset samples based on the LSTNet model; The LSTNet model is divided into three stages: convolutional neural network layer, recurrent neural network layer, and recurrent skip neural network layer; The specific steps are as follows;

[0065] Step 4.1: Input the obtained sample-offset sample pairs into the model and use a convolutional neural network to extract their local features;

[0066] Step 4.2: For the local features extracted by CNN, use recurrent neural networks (RNNs) to model short-term dependencies.

[0067] Step 4.3: Use Skip-RNN to model long-term dependencies on the local features extracted by CNN;

[0068] Step 4.4: Combine the short-term dependency modeling implemented by RNN with the long-term dependency modeling implemented by Skip-RNN to form the latent space features of the sample.

[0069] LSTNet network structure for samples The calculation process is as follows;

[0070] (2);

[0071] (3);

[0072] (4);

[0073] (5);

[0074] in, Indicates the input sample. This represents the local features extracted by the CNN. This represents short-term dependencies modeled using RNNs. This indicates long-term dependencies modeled using Skip-RNN. This represents a two-dimensional convolution operation. Indicates the GRU layer. This indicates skipping GRU layers. This indicates a splicing operation. This indicates the LSTNet encoder.

[0075] Step 5: Calculate the reconstruction loss and target offset loss using the obtained latent space features;

[0076] The formulas for calculating target offset loss are shown in formulas (6) and (7);

[0077] (6);

[0078] (7);

[0079] in, Indicates target offset loss. Indicates L1 loss, Represents the total number of samples. and Let represent the two input parameters in the L1 loss, and represent the two vectors that need to be compared. and Representing vectors respectively The first in elements and vectors The first in One element, and These represent the sample at the current time and the offset sample, respectively. and These represent the current clean coal ash content and the offset clean coal ash content, respectively.

[0080] The formulas for calculating reconstruction loss are shown in formulas (8) to (11);

[0081] (8);

[0082] (9);

[0083] (10);

[0084] (11);

[0085] in, and These represent the addition of two different types of random noise. and These represent the input samples respectively. The output values ​​of the LSTNet encoder after applying different noise levels. This represents the mean squared error loss. and These represent the two input parameters for the mean squared error loss. and Representing vectors respectively The first in elements and vectors The first in One element, Indicates the reconstruction loss. Indicates SIGReg loss. and These represent the weights of the mean squared error loss and the SIGReg loss, respectively.

[0086] The final loss is shown in formula (12);

[0087] (12);

[0088] in, Indicates the final loss. This represents the weight value corresponding to the reconstruction loss.

[0089] Step 6: Evaluation and application of the model.

[0090] In practical applications, the model adds the offset output by the model at the current moment to the ash content of the clean coal at the previous moment to obtain the accurate ash content value at the current moment.

[0091] The specific implementation is as follows.

[0092] Step 1: Collect heavy medium coal preparation data and construct a coal preparation dataset.

[0093] The data used in this invention all originate from the actual heavy media separation process of a coal preparation plant under Shandong Energy Zaozhuang Mining Group. The collected data covers the following three categories: First, key process parameters (as shown in Table 1), including sampling date, sampling time, magnetic content, coal slime content, separation density, hydrocyclone inlet pressure, and combined medium tank level, with a sampling frequency of once per minute; second, raw coal ash content and yield (as shown in Table 2), specifically including sampling date, sampling time, raw coal density classification, and the corresponding ash content and yield; third, clean coal ash content data (as shown in Table 3), specifically including sampling date, sampling time, and clean coal ash content value. In terms of production organization, a three-shift system (8 hours per shift) is adopted on-site. Each shift classifies raw coal according to density and conducts multiple raw coal float-sink experiments to obtain raw coal ash content as needed. Based on the above data, this invention systematically organizes and constructs a sample dataset for predicting the time-series offset of clean coal ash content.

[0094] Table 1: Example table of process parameter data;

[0095] .

[0096] Table 2: Example table of raw coal ash content and yield data;

[0097] .

[0098] Table 3: Example table of clean coal ash content data;

[0099] .

[0100] Step 2: Preprocess the collected data and divide it into training and test sets.

[0101] To enhance the model's ability to extract information from the heavy media coal preparation process, this invention first systematically preprocesses the raw collected data. Through operations such as missing value handling, noise and distortion introduced during data acquisition are effectively eliminated, ensuring the accuracy and consistency of the input data. After cleaning and reconstruction, the raw data is transformed into a structured format suitable for modeling input requirements, providing a reliable foundation for the model to fully explore the potential features in the data. After preprocessing, the resulting data is divided into training and testing sets at a 9:1 ratio for model training and performance evaluation, respectively. This preprocessing step effectively reduces prediction bias caused by raw data quality issues, laying a solid foundation for improving the accuracy of clean coal ash content prediction.

[0102] Step 3: Construct time-shift driven sample-offset sample pairs.

[0103] After the data preprocessing in step 2 is completed, offset samples are extracted using the processed samples and time offsets to construct time-shift driven sample-offset sample pairs. The specific method for generating offset samples is as follows: First, a time offset is set; second, the samples are shifted backward using the obtained time offset to obtain offset samples; finally, the samples and offset samples are combined to form sample-offset sample pairs. The calculation process for generating sample-offset sample pairs is as follows;

[0104] (1);

[0105] in, Indicates the offset sample. This represents the sample at the current moment. The time offset represents the sample-offset sample pair in the form of... .

[0106] Step 4: Extract the latent space features of the sample and the offset sample.

[0107] The latent space features of the samples and offset samples are constructed using the sample-offset pairs obtained in step 3 and combined with the LSTNet model. The structure of the LSTNet model is as follows: Figure 2 As shown, it mainly consists of three layers: a convolutional neural network layer, a training neural network layer, and a recurrent skipping neural network layer. The specific working process of this model is as follows:

[0108] Step 4.1: Input the obtained sample-offset sample pairs into the model and use a convolutional neural network (CNN) to extract their local features;

[0109] Step 4.2: Use a recurrent neural network (RNN) to model short-term dependencies on the local features extracted by the CNN.

[0110] Step 4.3: Use a recurrent skip neural network (Skip-RNN) to model long-term dependencies on the local features extracted by the CNN;

[0111] Step 4.4: Combine the short-term dependency modeling implemented by RNN with the long-term dependency modeling implemented by Skip-RNN to form the latent space features of the sample.

[0112] For input samples The overall calculation formula for this model is:

[0113] (2);

[0114] (3);

[0115] (4);

[0116] (5);

[0117] in, Indicates the input sample. This represents the local features extracted by the CNN. This represents short-term dependencies modeled using RNNs. This indicates long-term dependencies modeled using Skip-RNN. This represents a two-dimensional convolution operation. Indicates the GRU layer. This indicates skipping GRU layers. This indicates a splicing operation. This indicates the LSTNet encoder.

[0118] The GRU module structure used in this invention is as follows: Figure 3 As shown, the specific calculation formula is as follows;

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] in, For the present The input value of the time unit, and These are the hidden states at the current time and the previous time, respectively. This indicates the candidate state of the current work unit. and These represent the reset door and the update door, respectively. , and Represents the weight matrix. For the sigmoid function, The tanh activation function is used. This indicates the dot product operation.

[0124] Step 5: Calculate the reconstruction loss and target offset loss using the obtained latent space features.

[0125] The sample-offset sample pairs obtained in step 3 are used as model input, and the clean coal ash content offset is used as model output. The model is trained using the training set to improve its ability to predict the clean coal ash content offset at different times. The model combines the reconstruction loss and the target offset loss to form an overall loss function, which is used to calculate the difference between the predicted value and the actual value.

[0126] The formulas for calculating target offset loss are shown in formulas (6) and (7);

[0127] (6);

[0128] (7);

[0129] in, Indicates target offset loss. Indicates L1 loss, Represents the total number of samples. and Let represent the two input parameters in the L1 loss, and represent the two vectors that need to be compared. and Representing vectors respectively The first in elements and vectors The first in One element, Indicates LSTNet encoder, and These represent the sample at the current time and the offset sample, respectively. and These represent the current clean coal ash content and the offset clean coal ash content, respectively.

[0130] The formulas for calculating reconstruction loss are shown in formulas (8) to (11);

[0131] (8);

[0132] (9);

[0133] (10);

[0134] (11);

[0135] in, Indicates the input sample. Indicates LSTNet encoder, and These represent the addition of two different types of random noise. and These represent the input samples respectively. The output values ​​of the LSTNet encoder after applying different noise levels. This represents the mean square error. Represents the total number of samples. and These represent the two input parameters for the mean squared error loss. and Representing vectors respectively The first in elements and vectors The first in One element, Indicates the reconstruction loss. Indicates SIGReg loss. and These represent the weights of the mean squared error and the SIGReg loss, respectively.

[0136] The final loss is shown in formula (12);

[0137] (12);

[0138] in, Indicates the final loss. and These represent the sample at the current time and the offset sample, respectively. Indicates target offset loss. Indicates the reconstruction loss. This represents the weight value corresponding to the reconstruction loss.

[0139] This invention conducted experiments to determine the optimal weight values, and the results are as follows: Figure 4 As shown. The x-axis represents different weight parameter pairs ( The y-axis represents the mean absolute error of the model on the test set. Based on the above experimental results, in this invention... and Take values ​​of 0.1 and 0.6 respectively.

[0140] Furthermore, to achieve optimal model performance, this invention adjusts the designed hyperparameters. Specifically, the learning rate is 0.001; the Adam optimizer is used; the batch size is 1024; and the number of training epochs is 200. Simultaneously, to prevent overfitting, this invention monitors the changes in various metric values ​​during training, and saves the current optimal model parameters when the loss decreases.

[0141] Step 6: Model Evaluation and Application.

[0142] After model training, the model is evaluated using a test set. This invention uses Mean Absolute Error (MAE) for performance evaluation on the test set, and the obtained MAE is 0.0241, indicating high prediction accuracy.

[0143] To further verify the superiority of the proposed method for predicting the ash content offset of clean coal compared to directly predicting the ash content, comparative experiments were conducted, and the results are listed in Table 4. The results show that the proposed method significantly outperforms existing methods in terms of prediction performance.

[0144] Table 4: Method Comparison Table;

[0145] .

[0146] When applying the evaluated model to the actual coal preparation process, this invention does not directly predict the ash content of the clean coal, but rather predicts its offset. In specific application, the offset predicted at the current moment is added to the ash content of the clean coal at the previous moment to obtain the ash content of the clean coal at the current moment. The application process is as follows: Figure 5 As shown, this method effectively overcomes the problems of traditional prediction methods by predicting ash content offset, achieving accurate prediction of clean coal ash content, and improving the model's cross-scenario transferability, thereby further improving the overall efficiency of the heavy media coal preparation process.

[0147] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A test-time adaptive method for predicting the ash content of clean coal during implicit space concept drift assessment, characterized in that, Accurate prediction of target ash offset is achieved using sample-offset sample pairs, including the following steps: Step 1: Collect heavy media coal preparation data and construct a coal preparation dataset; Step 2: Preprocess the collected data and divide it into training and test sets; Step 3: Construct time-shift driven sample-offset sample pairs; Step 4: Extract latent space features of samples and offset samples based on the LSTNet model; Step 5: Calculate the reconstruction loss and target offset loss using the obtained latent space features; Step 6: Evaluation and application of the model; In step 5, the target offset loss calculation formulas are shown in formulas (6) and (7): (6)； (7)； in, Indicates target offset loss. Indicates L1 loss, Indicates the total number of elements. and Let represent the two input parameters in the L1 loss, and represent the two vectors that need to be compared. and Representing vectors respectively The first in elements and vectors The first in One element, and These represent the sample at the current time and the offset sample, respectively. and These represent the current clean coal ash content and the offset clean coal ash content, respectively. The formulas for calculating reconstruction loss are shown in formulas (8) to (11); (8)； (9)； (10)； (11)； in, and These represent the addition of two different types of random noise. and These represent the input samples respectively. The output values ​​of the LSTNet encoder after applying different noise levels. This represents the mean squared error loss. and These represent the two input parameters for the mean squared error loss. and Representing vectors respectively The first in elements and vectors The first in One element, Indicates the reconstruction loss. Indicates SIGReg loss. and These represent the weights of the mean squared error loss and the SIGReg loss, respectively. The final loss is shown in formula (12): (12)； in, Indicates the final loss. This represents the weight value corresponding to the reconstruction loss.

2. The test-time adaptive clean coal ash content prediction method for implicit space concept drift assessment according to claim 1, characterized in that, In step 1, the key process parameters collected include at least the separation density, coal slime content, magnetic content, raw coal data, hydrocyclone inlet pressure and combined tank liquid level, as well as the corresponding clean coal ash content index.

3. The test-time adaptive clean coal ash content prediction method for implicit space concept drift assessment according to claim 1, characterized in that, In step 2, the preprocessing includes at least sensor error correction and missing value processing. After preprocessing, the dataset is divided into training set and test set in a 9:1 ratio for subsequent training and performance evaluation of the model.

4. The test-time adaptive clean coal ash content prediction method for implicit space concept drift assessment according to claim 3, characterized in that, In step 3, the time offset is designed using the preprocessed data, and the offset sample is extracted based on the sample and the time offset to construct the time shift driven sample-offset sample pair. The construction process is shown in formula (1): （1）； in, Indicates the offset sample. This represents the sample at the current moment. The time offset represents the sample-offset sample pair in the form of... .

5. The test-time adaptive clean coal ash content prediction method for implicit space concept drift assessment according to claim 1, characterized in that, In step 4, latent space features are extracted from samples and offset samples based on the LSTNet model. The LSTNet model consists of three stages: a convolutional neural network layer, a recurrent neural network layer, and a recurrent skip neural network layer. The specific steps are as follows: Step 4.1: Input the obtained sample-offset sample pairs into the model and use a convolutional neural network to extract their local features; Step 4.2: Use a recurrent neural network to model short-term dependencies on the local features extracted by the convolutional neural network. Step 4.3: Use a recurrent skip neural network to model long-term dependencies on the local features extracted by the convolutional neural network. Step 4.4: Combine the short-term dependency modeling implemented by the recurrent neural network with the long-term dependency modeling implemented by the recurrent skip neural network to form the latent space features of the sample.

6. The test-time adaptive clean coal ash content prediction method for implicit space concept drift assessment according to claim 5, characterized in that, LSTNet network structure for samples The calculation process is as follows: (2)； (3)； (4)； (5)； in, Indicates the input sample. This represents the local features extracted by the convolutional neural network. This represents short-term dependencies modeled using recurrent neural networks. This indicates long-term dependencies modeled using a recurrent skip neural network. This represents a two-dimensional convolution operation. Indicates the GRU layer. This indicates skipping GRU layers. This indicates a splicing operation. This indicates the LSTNet encoder.

7. The test-time adaptive clean coal ash content prediction method for implicit space concept drift assessment according to claim 1, characterized in that, In step 6, in practical applications, the model adds the offset output by the model at the current moment to the ash content of the clean coal at the previous moment to obtain the accurate ash content value at the current moment.

Images