Method for predicting air powder generation amount of pulverized coal fired boiler
Through the FlowFusionNet model, the combination of the unified spatiotemporal processing module and the feature fusion prediction module solves the problem that traditional methods are difficult to capture complex spatiotemporal dependencies, and realizes accurate prediction of the air powder generation of coal powder boilers, improving the accuracy and stability of prediction.
Patent Information
- Application Number
- CN202411753215.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-02
- Publication Date
- 2025-03-25
AI Technical Summary
Traditional coal powder generation prediction methods for coal powder boilers are difficult to effectively capture complex space-time dependencies, especially in boiler systems with highly heterogeneous and dynamic changes in spatiotemporal data, resulting in limited prediction performance.
The FlowFusionNet model is proposed, including a unified spatiotemporal processing module and a feature fusion prediction module. The unified spatiotemporal processing module extracts global representations through the backbone network and captures short-term local dependency features through the adaptive compression rank feature layer. The feature fusion prediction module uses parameter normalization and dynamic adjustment, combining global and local features to perform dynamic fusion to output predicted values.
The FlowFusionNet model can effectively handle the spatial and temporal heterogeneity and dynamic changes in the air powder generation data of coal powder boilers, significantly improving the accuracy and stability of prediction, and showing higher prediction accuracy in complex multidimensional data.
Smart Images
Figure CN119691550A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of machine learning, and in particular relates to a method for predicting the amount of air-powdered coal boiler generated. Background Art
[0002] During the operation of pulverized coal boilers, the prediction of air-powder generation is a key link in achieving combustion efficiency optimization and energy consumption control. The air-powder generation is affected by many factors, including fuel type, boiler load, air supply volume, pulverized coal fineness, etc., and has a significant spatiotemporal dependency. Traditional prediction methods usually rely on statistical models or classic machine learning models, which perform well under small-scale data or low-dimensional features, but in complex boiler systems, it is difficult to fully capture the nonlinear spatiotemporal dependency characteristics.
[0003] The prediction of air-dust generation in boiler systems faces the high heterogeneity of spatiotemporal data. The air-dust generation in different areas of the boiler may be affected by both local conditions and global characteristics. In the time dimension, the air-dust generation of the boiler is affected by dynamic external conditions. There are also mixed behaviors of periodicity and non-periodicity. Traditional time series methods are difficult to effectively capture such complex dynamic changes. Such complex spatiotemporal dependencies place higher demands on the model. Traditional prediction methods usually treat the time and space dimensions separately and cannot capture their coupling relationship, resulting in limited prediction performance.
[0004] In recent years, the rapid development of deep learning technology has provided powerful modeling capabilities for complex prediction tasks in the industrial field. The introduction of deep learning has significantly enhanced the learning and modeling capabilities of the dynamic changes of complex systems. The data of industrial systems usually exist in the form of time series and spatial dimensions, such as the air-powder concentration, wind speed, temperature in different areas of the boiler. Deep learning models can integrate multi-source spatiotemporal data, extract nonlinear features and establish complex dependencies in time and space dimensions, thereby achieving accurate predictions. Summary of the invention
[0005] The present invention provides a method for predicting the air-powder generation of a pulverized coal boiler. Aiming at the air-powder generation data of a pulverized coal boiler with spatiotemporal correlation and heterogeneity, a FlowFusionNet model is proposed, which consists of a unified spatiotemporal processing module and a feature fusion prediction module.
[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose specifically includes the following steps: S1. Collect data related to the amount of air and pulverized coal generated by pulverized coal boilers, including features and target variables, and preprocess the collected data on the amount of air and pulverized coal generated by pulverized coal boilers; S2, using the Max-Min method to normalize the air-powder generation data of pulverized coal boilers and divide the data set proportionally; S3. Build a unified spatiotemporal processing module to process short-term local dependencies and long-term global trends in data. The specific steps are as follows: S31, input the pulverized coal boiler air pulverized coal generation data X∈R s×n×c , extract the global representation Z through the backbone network b ; S32, propose an adaptive compression rank feature layer, introduce a weight adaptation mechanism, generate a dynamic weight Δw, perform compression rank operation, and obtain short-term local dependency features With the output feature y i ; S33, model the output features, and obtain the global feature matrix through iterative calculation and update and the local feature matrix S4. Construct a feature fusion prediction module to handle the heterogeneity and dynamic changes in the data. The specific steps are as follows: S41, normalize and dynamically adjust the pulverized coal boiler air pulverized coal generation data X and the shared compression rank parameter matrix W to obtain the final feature representation S42, map the final feature representation to the prediction dimension through the fully connected layer to obtain the intermediate representation Z t ; S43, dynamically fuse the intermediate representation and the global representation through a weighted fusion unit, and output a predicted value Z of the amount of air pulverized coal boiler generated.
[0007] Preferably, in S1, data related to the air-powder generation of the pulverized coal boiler are collected, including air-powder generation data, air-powder concentration data, boiler load data, primary air volume data, secondary air volume data, combustion temperature data and external air pressure data, and the mean method is used to fill the missing values of the air-powder generation data of the pulverized coal boiler. The specific formula is as follows: Where, X ave is the imputed mean, n is the number of missing values, x i is a non-missing value.
[0008] Preferably, in S2, the Max-Min method is used to normalize the air-powder generation data of the pulverized coal boiler, and the training set and the test set are divided proportionally. The specific formula is as follows: Where x is the air pulverized coal boiler generation data, x min is the minimum value in the data of air pulverized coal boiler generation, x max It is the maximum value in the data of air-powder generation of pulverized coal boiler.
[0009] Preferably, the pulverized coal boiler air pulverized coal generation data X∈R is input in S3 and S31. s×n×c , where s is the time step, n is the number of spatial nodes, representing different areas in the boiler, and c is the feature dimension of each spatial node. It is then processed through the backbone network, which is composed of a multi-layer perceptron to extract the global representation. The specific formula is as follows: In the formula, F b (X) is the backbone network, which is composed of a multi-layer sensor. X is the input data of the amount of air and powder generated by the pulverized coal boiler. f h is the h-th time step feature generated by the backbone network, n is the number of spatial nodes, C out It is the feature dimension of each spatial node output by the backbone network.
[0010] Preferably, a backbone network composed of multi-layer sensors is used to extract global feature representations, and global trend information is captured by stacking network layers. This can effectively extract global features from the air-powder generation data of pulverized coal boilers, enabling the model to learn global trends and generate accurate global feature representations. The multi-layer structure of the backbone network enhances the model's ability to capture global trends and significantly improves the modeling effect of long-term dependencies.
[0011] Preferably, the adaptive compression rank feature layer is proposed in S3 and S32, and the air pulverized coal boiler generation data X∈R s×n×c Extract short-term local dependency features. The specific formula is as follows: In the formula, F c (X) is the compressed rank linear transformation function, X is the input pulverized coal boiler air pulverized coal generation data, wh is the feature extracted by the adaptive compressed rank feature layer at the hth time step, n is the number of spatial nodes, m is the dimension of the hidden layer, and then the short-term local dependency features are decomposed by compressed rank matrix. After processing, the weight matrix after compression rank matrix decomposition is obtained. The specific formula is as follows: In the formula, Reshape is the rearrangement operation, F(.) is the compressed rank linear transformation function, is the weight vector of the nth spatial node in the i-th iteration, N is the dimension of the input feature, and M is the dimension of the output feature. Then, the weight adaptation mechanism is introduced to generate updated dynamic weights through compressed rank matrix decomposition. The specific formula is as follows: Where A∈R n×r , B∈R m×ris the compressed rank matrix, where r is the rank of the matrix, n is the number of spatial nodes, m is the dimension of the hidden layer, α is the regularization parameter, and ε is the numerical stability constant. The weight matrix W after decomposing the compressed rank matrix (i) The output feature calculation is completed by interactively combining with the input pulverized coal boiler air pulverized coal generation data X through the dynamic weight Δw. The specific formula is as follows: y i =σ(W (i) X+Δw (i) X+b); Where W (i) is the weight matrix after compressed rank matrix decomposition, X is the air pulverized coal boiler generation data, Δw is the dynamic weight, b is the bias term, and σ is the ReLU activation function.
[0012] Preferably, an adaptive compressed rank feature layer is proposed, and a dynamic weight mechanism is introduced. By decomposing the compressed rank matrix, short-term local dependencies can be captured. The feature weights can be adjusted dynamically, and the features that affect short-term changes can be extracted with emphasis, thereby enhancing the model's sensitivity to local features, effectively capturing the short-term dependency features in time series data, and improving the model's ability to adapt to rapidly changing trends.
[0013] Preferably, the output feature y in S3 and S33 is i Modeling is performed to obtain the global feature matrix and the local feature matrix, and then iterative calculation and update are performed. The specific formula is as follows: In the formula, Initialize the global feature matrix and the local feature matrix simultaneously by outputting the features, X ：t is the spatial node feature of the current time step t, is the local feature matrix obtained after the i-1th iteration, is the global feature matrix obtained after the i-1th iteration, W is the shared compression rank parameter matrix, which is learned during the training process, b is the bias term, Update is the update iteration operation, and then the global feature matrix is updated by fusion and the local feature matrix Generate a shared compression rank parameter matrix. The specific formula is as follows: In the formula, is the global feature matrix, The local feature matrix, is the fusion operation, b is the bias term, Memory is the operation of integrating these global and local feature matrices into the shared compressed rank parameter matrix W, and ΔW is the parameter update value.
[0014] Preferably, the output features are modeled, and the global feature matrix and the local feature matrix are generated through iterative calculation. The short-term and long-term features are dynamically fused, and the global and local information can be dynamically combined, which enhances the model's ability to integrate multidimensional features. The dynamic update mechanism of global and local features significantly improves the generalization ability of the model, enabling it to show higher prediction accuracy in complex multidimensional data.
[0015] Preferably, in S4 and S41, the weight matrix W after decomposing the pulverized coal boiler air pulverized coal generation data X and the compressed rank matrix is (i) Perform parameter normalization and dynamic adjustment to obtain the final feature representation. The specific formula is as follows: Where, X is the air pulverized coal boiler generation data, W (i-1) is the weight matrix after the compression rank matrix decomposition of the i-1th layer, d is the feature dimension, ε is a small positive number used to avoid the denominator being zero, and Norm is the normalization operation. is the global feature matrix of the lth layer, The local feature matrix of layer l, σ is the ReLU activation function, W (i) is the weight matrix after compressed rank matrix decomposition.
[0016] Preferably, the pulverized coal boiler air-powder generation data and the weight matrix after compression rank matrix decomposition are parameterized and dynamically adjusted to generate the final feature representation, which reduces the redundancy of feature dimensions, enhances the model's ability to assign weights to feature importance, and significantly improves the model's adaptability to heterogeneous data, thereby enhancing the feature fusion effect.
[0017] Preferably, in S4 and S42, the final feature representation is mapped to To the prediction dimension, we get the intermediate representation. The specific formula is as follows: In the formula, FC is the fully connected layer, σ is the ReLU activation function, is the final feature representation mapping.
[0018] Preferably, the final feature representation is mapped to the prediction dimension through a fully connected layer to generate an intermediate representation. The use of the ReLU activation function enhances the expressiveness of nonlinear features, achieves efficient mapping of features to target dimensions, provides high-quality intermediate representations for the final prediction, improves the nonlinear expressiveness of features, and enhances the modeling effect of complex feature relationships.
[0019] Preferably, in S4 and S43, the intermediate representation and the global representation are dynamically fused through a weighted fusion unit, and finally the predicted value of the air-powdered coal boiler generation amount is output. The specific formula is as follows: Z=H(z b , z t ;X ：t ,α)=(1-α)⊙F b (X :t )+α⊙Z t ; In the formula, H is the feature fusion function, z b is the global representation, z t is the intermediate representation, X ：t is the spatial node feature of the current time step t, α is the weight of weighted fusion, F b (X ：t ) is the backbone network, ⊙ is the element-by-element multiplication operation, and Z is the predicted value of the air-powder generation of the pulverized coal boiler.
[0020] Preferably, the intermediate representation and the global representation are dynamically fused through a weighted fusion unit to generate the final wind-dust generation prediction value, which combines the advantages of global and local features to ensure the comprehensiveness and accuracy of the final prediction result. Moreover, through the weighted fusion operation, the model excels in the processing ability of heterogeneous features, further improving the stability and accuracy of the prediction.
[0021] In summary, due to the adoption of the present technical solution, the beneficial effects of the present invention are as follows: the present invention proposes a FlowFusionNet prediction model, which is applied to the prediction scenario of air-powder generation in pulverized coal boilers, including a unified space-time processing module and a feature fusion prediction module. The unified space-time processing module processes short-term local dependencies and long-term global trends in the data, and the feature fusion prediction module is used to process the heterogeneity and dynamic changes in the data, and outputs the prediction results of the air-powder generation of pulverized coal boilers. The modules cooperate with each other to achieve accurate prediction of the air-powder generation of pulverized coal boilers. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 This is a step-by-step diagram of a method for predicting the amount of air-powder generated by a pulverized coal boiler.
[0023] Figure 2 This is the structure diagram of the FlowFusionNet prediction model.
[0024] Figure 3 This is the FlowFusionNet prediction model for the prediction of the air pulverized coal boiler generation amount. DETAILED DESCRIPTION
[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0026] See also Figure 1-Figure 3 The present invention provides a technical solution: a method for predicting the amount of air and pulverized coal boiler generated by pulverized coal. By constructing a unified spatiotemporal processing module, the short-term local dependency and long-term global trend in the data can be processed. The feature fusion prediction module can process the heterogeneity and dynamic changes in the data and output the prediction result of the amount of air and pulverized coal boiler generated by pulverized coal. The specific steps are as follows: Figure 1 shown.
[0027] Construct the FlowFusionNet prediction model, whose structure is as follows Figure 2 As shown, the specific steps are as follows:
[0028] S1. Collect data related to the amount of air and pulverized coal boiler generated, including features and target variables, and pre-process the collected data on the amount of air and pulverized coal boiler generated.
[0029] Furthermore, the data related to the air-powder generation of pulverized coal boilers are collected, including air-powder generation data, air-powder concentration data, boiler load data, primary air volume data, secondary air volume data, combustion temperature data and external air pressure data. The mean method is used to fill the missing values of the air-powder generation data of pulverized coal boilers. The specific formula is as follows: In the formula, X ave is the imputed mean, n is the number of missing values, x i is a non-missing value.
[0030] S2. Use the Max-Min method to normalize the air-powder generation data of pulverized coal boilers and divide the data set proportionally.
[0031] Furthermore, the Max-Min method is used to normalize the air-powder generation data of the pulverized coal boiler, and the training set and the test set are divided into a ratio of 7:3. The specific formula is as follows: Where x is the air pulverized coal boiler generation data, x min is the minimum value in the data of air pulverized coal boiler generation, x max It is the maximum value in the data of air-powder generation of pulverized coal boiler.
[0032] S31, input the pulverized coal boiler air pulverized coal generation data X∈R s×n×c, extract the global representation z through the backbone network b .
[0033] Furthermore, a unified spatiotemporal processing module is constructed, and the pulverized coal boiler air pulverized coal generation data X∈R s×n×c , where s is the time step, n is the number of spatial nodes, representing different areas in the boiler, and x is the feature dimension of each spatial node. It is then processed through the backbone network, which is composed of a multi-layer perceptron to extract the global representation. The specific formula is as follows: In the formula, F b (X) is the backbone network, which is composed of a multi-layer sensor. X is the input data of the amount of air pulverized coal boiler generated. f h is the h-th time step feature generated by the backbone network, n is the number of spatial nodes, c out It is the feature dimension of each spatial node output by the backbone network.
[0034] S32, propose an adaptive compression rank feature layer, introduce a weight adaptation mechanism, generate a dynamic weight Δw, perform compression rank operation, and obtain short-term local dependency features With the output feature y i .
[0035] Furthermore, an adaptive compression rank feature layer is proposed, and the air pulverized coal boiler generation data X∈R s×n×c Extract short-term local dependency features. The specific formula is as follows: In the formula, F c (X) is the compressed rank linear transformation function, X is the input pulverized coal boiler air pulverized coal generation data, w h is the feature extracted by the adaptive compressed rank feature layer at the hth time step, n is the number of spatial nodes, m is the dimension of the hidden layer, and then the short-term local dependency features are decomposed by compressed rank matrix After processing, the weight matrix after compression rank matrix decomposition is obtained. The specific formula is as follows: In the formula, Reshape is the rearrangement operation, F(.) is the compressed rank linear transformation function, is the weight vector of the nth spatial node in the i-th iteration, N is the dimension of the input feature, and M is the dimension of the output feature. Then, the weight adaptation mechanism is introduced to generate updated dynamic weights through compressed rank matrix decomposition. The specific formula is as follows: Where A∈R n×r , B∈R m×ris the compressed rank matrix, where r is the rank of the matrix, n is the number of spatial nodes, m is the dimension of the hidden layer, α is the regularization parameter, and ε is the numerical stability constant. The weight matrix W after decomposing the compressed rank matrix (i) The output feature calculation is completed by interactively combining with the input pulverized coal boiler air pulverized coal generation data X through the dynamic weight Δw. The specific formula is as follows: y i =σ(W (i) X+Δw (i) X+b); Where W (i) is the weight matrix after compressed rank matrix decomposition, X is the air pulverized coal boiler generation data, Δw is the dynamic weight, b is the bias term, and σ is the ReLU activation function.
[0036] S33, model the output features, and obtain the global feature matrix through iterative calculation and update and the local feature matrix
[0037] Furthermore, the output feature y i Modeling is performed to obtain the global feature matrix and the local feature matrix, and then iterative calculation and update are performed. The specific formula is as follows: In the formula, Initialize the global feature matrix and the local feature matrix simultaneously by outputting the features, X ：t is the spatial node feature of the current time step t, is the local feature matrix obtained after the i-1th iteration, is the global feature matrix obtained after the i-1th iteration, W is the shared compression rank parameter matrix, which is learned during the training process, b is the bias term, Update is the update iteration operation, and then the global feature matrix is updated by fusion and the local feature matrix Generate a shared compression rank parameter matrix. The specific formula is as follows: In the formula, is the global feature matrix, The local feature matrix, is the fusion operation, b is the bias term, Memory is the operation of integrating these global and local feature matrices into the shared compressed rank parameter matrix W, and ΔW is the parameter update value.
[0038] S41, normalize and dynamically adjust the pulverized coal boiler air pulverized coal generation data X and the shared compression rank parameter matrix W to obtain the final feature representation
[0039] Furthermore, a feature fusion prediction module is constructed to combine the pulverized coal boiler air pulverized coal generation data X with the weight matrix W after the compressed rank matrix decomposition. (i) Perform parameter normalization and dynamic adjustment to obtain the final feature representation. The specific formula is as follows: Where, X is the air pulverized coal boiler generation data, W (i-1) is the weight matrix after the compression rank matrix decomposition of the i-1th layer, d is the feature dimension, ε is a small positive number used to avoid the denominator being zero, and Norm is the normalization operation. is the global feature matrix of the lth layer, The local feature matrix of layer l, σ is the ReLU activation function, W (i) is the weight matrix after compressed rank matrix decomposition.
[0040] S42, map the final feature representation to the prediction dimension through the fully connected layer to obtain the intermediate representation z t .
[0041] Furthermore, the final feature representation is mapped through the fully connected layer To the prediction dimension, we get the intermediate representation. The specific formula is as follows: In the formula, FC is the fully connected layer, σ is the ReLU activation function, is the final feature representation mapping.
[0042] S43, dynamically fuse the intermediate representation and the global representation through a weighted fusion unit, and output a predicted value Z of the amount of air pulverized coal boiler generated.
[0043] Furthermore, the intermediate representation and the global representation are dynamically fused through a weighted fusion unit, and finally the predicted value of the air-powder generation amount of the pulverized coal boiler is output. The specific formula is as follows: Z=H(z b ,z t ;X ：t ,α)=(1-α)☉F b (X ：t )+α⊙z t ; In the formula, H is the feature fusion function, z b is the global representation, z t is the intermediate representation, X ：t is the spatial node feature of the current time step t, α is the weight of weighted fusion, F b (X ：t) is the backbone network, ⊙ is the element-by-element multiplication operation, and Z is the predicted value of the air-powder generation of the pulverized coal boiler.
[0044] Furthermore, the FlowFusionNet prediction model is written in Python. The experiment runs on the Windows operating system. Pytorch is selected as the framework in the CUDA11.27 environment. Training is performed on the NVIDIA RTX A6000 GPU. Adam is selected as the optimizer. The initial learning rate is set to 0.001, the number of training rounds is set to 150, the training batch is set to 64, and the weight decay is set to 0.0005. The data set is 60 days of pulverized coal boiler air pulverized gas generation related data, which is preprocessed and input into the FlowFusionNet prediction model.
[0045] Furthermore, the FlowFusionNet prediction model predicts the air-powder generation of pulverized coal boilers. Figure 3 As shown in the figure, the actual data of the air-powder generation of the pulverized coal boiler is compared with the predicted data. The horizontal axis is time (days), the vertical axis is the air-powder generation (cubic meters), the dotted line represents the actual data, and the solid line represents the predicted data of the model. It can be seen from the figure that the trend of the predicted value is basically consistent with the actual value, especially in most areas with large fluctuations, the predicted value of the model can accurately follow the rising and falling trends of the actual value, indicating that the model has a strong adaptability when dealing with the volatile data of air-powder generation. The chart covers a time range of one month. The model has good predictive ability over a long time span. Whether it is the peak or trough period of generation, the model can better predict the trend of change, indicating that it has good modeling and prediction capabilities for long time series data.
Claims
1. A method for predicting the amount of air-powder generated by a pulverized coal boiler, characterized in that: The following steps are involved: S1. Collect data related to the amount of air and pulverized coal boiler generated, including features and target variables, and pre-process the collected data on the amount of air and pulverized coal boiler generated; S2, using the Max-Min method to normalize the air-powder generation data of pulverized coal boilers and divide the data set proportionally; S3. Build a unified spatiotemporal processing module to process short-term local dependencies and long-term global trends in data. The specific steps are as follows: S31, input the pulverized coal boiler air pulverized coal generation data X∈R s×n×c , extract the global representation Z through the backbone network b ; S32, propose an adaptive compression rank feature layer, introduce a weight adaptation mechanism, generate a dynamic weight Δw, and perform a compression rank operation to obtain a short-term local dependency feature and the output feature yi; S33, model the output features, and obtain the global feature matrix through iterative calculation and update and the local feature matrix S4. Construct a feature fusion prediction module to handle the heterogeneity and dynamic changes in the data. The specific steps are as follows: S41, normalize and dynamically adjust the pulverized coal boiler air pulverized coal generation data X and the shared compression rank parameter matrix W to obtain the final feature representation S42, map the final feature representation to the prediction dimension through the fully connected layer to obtain the intermediate representation Z t ; S43, dynamically fuse the intermediate representation and the global representation through a weighted fusion unit, and output a predicted value Z of the amount of air pulverized coal boiler generated.
2. A method for predicting the amount of air-powder generated by a pulverized coal boiler according to claim 1, characterized in that: In S3 and S31, the pulverized coal boiler air pulverized coal generation data X∈R s×n×c , where S is the time step, n is the number of spatial nodes, representing different areas in the boiler, and C is the feature dimension of each spatial node. It is then processed through the backbone network, which is composed of a multi-layer perceptron to extract the global representation. The specific formula is as follows: In the formula, F b (X) is the backbone network, which is composed of a multi-layer sensor. X is the input data of the amount of air pulverized coal boiler generated. f h is the h-th time step feature generated by the backbone network, n is the number of spatial nodes, c out It is the feature dimension of each spatial node output by the backbone network.
3. A method for predicting the amount of air-powder generated by a pulverized coal boiler according to claim 2, characterized in that: The adaptive compression rank feature layer is proposed in S3 and S32, and the air pulverized coal boiler generation data X∈R s×n×c Extract short-term local dependency features. The specific formula is as follows: In the formula, F c (X) is the compressed rank linear transformation function, X is the input pulverized coal boiler air pulverized coal generation data, W h is the feature extracted by the adaptive compressed rank feature layer at the hth time step, n is the number of spatial nodes, m is the dimension of the hidden layer, and then the short-term local dependency features are decomposed by compressed rank matrix After processing, the weight matrix after compression rank matrix decomposition is obtained. The specific formula is as follows: In the formula, Reshape is the rearrangement operation, F(.) is the compressed rank linear transformation function, is the weight vector of the nth spatial node in the i-th iteration, N is the dimension of the input feature, and M is the dimension of the output feature. Then, the weight adaptation mechanism is introduced to generate updated dynamic weights through compressed rank matrix decomposition. The specific formula is as follows: Where A∈R n×r , B∈R m×r is the compressed rank matrix, where r is the rank of the matrix, n is the number of spatial nodes, m is the dimension of the hidden layer, α is the regularization parameter, and ε is the numerical stability constant. The weight matrix W after decomposing the compressed rank matrix (i) The output feature calculation is completed by interactively combining with the input pulverized coal boiler air pulverized coal generation data X through the dynamic weight Δw. The specific formula is as follows: yes i =σ(W (i) X+Δw (i) X+b); Where W (i) is the weight matrix after compressed rank matrix decomposition, X is the air pulverized coal boiler generation data, Δw is the dynamic weight, b is the bias term, and σ is the ReLU activation function.
4. A method for predicting the amount of air-powder generated by a pulverized coal boiler according to claim 3, characterized in that: In S3 and S33, the output feature yi is modeled to obtain a global feature matrix and a local feature matrix, and iterative calculation and update are performed. The specific formula is as follows: In the formula, Initialize the global feature matrix and the local feature matrix simultaneously by outputting the features, X ：t is the spatial node feature of the current time step t, is the local feature matrix obtained after the i-1th iteration, is the global feature matrix obtained after the i-1th iteration, W is the shared compression rank parameter matrix, which is learned during the training process, b is the bias term, Update is the update iteration operation, and then the global feature matrix is updated by fusion and the local feature matrix Generate a shared compression rank parameter matrix. The specific formula is as follows: In the formula, is the global feature matrix, The local feature matrix, is the fusion operation, b is the bias term, Memory is the operation of integrating these global and local feature matrices into the shared compressed rank parameter matrix W, and ΔW is the parameter update value.
5. A method for predicting the amount of air-powder generated by a pulverized coal boiler according to claim 4, characterized in that: In S4 and S41, the pulverized coal boiler air pulverized coal generation data X and the weight matrix W after the compressed rank matrix is decomposed (i) Perform parameter normalization and dynamic adjustment to obtain the final feature representation. The specific formula is as follows: Where, X is the air pulverized coal boiler generation data, W (i-1) is the weight matrix after the compression rank matrix decomposition of the i-1th layer, d is the feature dimension, ε is a small positive number used to avoid the denominator being zero, and Norm is the normalization operation. is the global feature matrix of the lth layer, The local feature matrix of layer l, σ is the ReLU activation function, W (i) is the weight matrix after compressed rank matrix decomposition.
6. A method for predicting the amount of air-powder generated by a pulverized coal boiler according to claim 5, characterized in that: In S4 and S42, the final feature representation is mapped through a fully connected layer To the prediction dimension, we get the intermediate representation. The specific formula is as follows: In the formula, FC is the fully connected layer, σ is the ReLU activation function, is the final feature representation mapping.
7. A method for predicting the amount of air-powder generated by a pulverized coal boiler according to claim 6, characterized in that: In S4 and S43, the intermediate representation and the global representation are dynamically fused through a weighted fusion unit, and finally the predicted value of the air pulverized coal boiler generation amount is output. The specific formula is as follows: Z=H(Z b ,Z t ;X ：t ,α)=(1-α)⊙F b (X ：t )+α⊙Z t ; In the formula, H is the feature fusion function, Z b is the global representation, Z t is the intermediate representation, X ：t is the spatial node feature of the current time step t, α is the weight of weighted fusion, F b (X ：t ) is the backbone network, ⊙ is the element-by-element multiplication operation, and Z is the predicted value of the air-powder generation of the pulverized coal boiler.
8. The method for predicting the amount of air-powder generated by a pulverized coal boiler according to claim 1, characterized in that: Aiming at the problem of predicting the amount of air and pulverized coal boiler generated, the collected data related to the amount of air and pulverized coal boiler generated include air and pulverized coal generation data, air and pulverized coal concentration data, boiler load data, primary air volume data, secondary air volume data, combustion temperature data and external air pressure data. The collected relevant data are preprocessed to ensure that there are no missing values in the data, and then the processed data are divided into training set and test set for training and evaluating the performance of the prediction model of air and pulverized coal boiler generated.