Industrial flue gas multi-modal prediction method based on a gated multi-dimensional information fusion transformer

Abstract

The application discloses an industrial flue gas multi-modal prediction method based on a gated multi-dimensional information fusion transformer, comprising: adopting a feature adaptive multi-modal partition method FAMP to perform modal recognition and segmentation on industrial time series data, and extracting data features under different operation modes. The multi-modal data are subjected to deep feature extraction through a GMFTrans encoder, a multi-layer perception module is combined to capture dynamic dependency relationship between cross modes, a gating mechanism is introduced to adaptively fuse multi-dimensional feature representation, and a Transformer decoding structure is used to realize accurate prediction of NOx and concentration in flue gas. According to the feature adaptive modal partition mechanism, the application can dynamically classify according to the importance mode of the inherent characteristics of data, enhance the description ability of the model to the multi-modal characteristics of the complex industrial process, and thus still maintain high prediction consistency in the multi-modal switching scene.

Description

Technical Field

[0001] This invention relates to the fields of petrochemical process modeling and pollutant emission prediction technology, specifically to a multimodal prediction method for industrial flue gas based on a gated multidimensional information fusion transformer. Background Technology

[0002] The rapid development of the petrochemical industry has brought about serious environmental pollution problems, with emissions occurring during desulfurization and denitrification processes. and These harmful gases pose a threat to the ecosystem.

[0003] The desulfurization and denitrification process of catalytic cracking regenerated flue gas (CCRFG) involves complex chemical reactions, and the data exhibits characteristics such as multimodal, strong nonlinearity, and time-series dynamic evolution. Most traditional prediction models are based on single-mode assumptions, which makes it difficult to adapt to the complex multimodal characteristics of actual industrial processes. Furthermore, they rely on manually preset modes and cannot dynamically respond to changes in operating conditions.

[0004] Clustering methods such as K-Means and KFCM are easily limited by preset parameters, LSTM deep learning models struggle to handle frequent data switching, and traditional Transformers suffer from high computational overhead and a lack of adaptive modality adjustment mechanisms. To address these issues, this invention proposes a multimodal prediction framework based on GMFTrans. This framework achieves dynamic modality partitioning of data through the Feature Adaptive Multimodal Partitioning (FAMP) method, extracts modality-specific features using the multi-encoder structure of GMFTrans, captures cross-modal dependencies using a multilayer perceptron (MLP), and dynamically fuses multidimensional information using a gating mechanism. This allows for in-depth mining of intramodal and intermodal connections, resulting in strong prediction accuracy. Summary of the Invention

[0005] The purpose of this invention is to propose a multimodal prediction method for industrial flue gas based on a gated multidimensional information fusion converter to solve the problems mentioned in the background art.

[0006] To achieve the aforementioned objectives, the first technical solution adopted by this invention is: a multimodal prediction method for industrial flue gas based on a gated multidimensional information fusion converter, comprising: Data on flue gas treatment in a refinery is collected, preprocessed, and divided into subsets of windows for each partition based on an adaptive window. Feature data and target variables are extracted from the window subset of each partition. For the feature data in each window, the mutual information value between each feature and the target variable is calculated as the importance score of the feature in the current window. The feature importance scores calculated within the window are normalized, and the features are grouped based on the similarity between their importance scores. The multi-mode partitioning module generates partitioning results, normalizes the new feature data obtained in each mode, generates a time series of length L, and inputs it into the encoder to obtain the encoder output features. The encoder output features corresponding to different modes are concatenated along the feature dimension, and the concatenated features are input into a multilayer perceptron (MLP) to learn the cross-modal correlation between different modes and obtain cross-modal feature information. A gating mechanism is introduced to fuse the output features of different encoders and the information from the output of a multilayer perceptron (MLP) to obtain fused feature information. The fused feature information is input into the decoder, which outputs the predicted target value.

[0007] Furthermore, the adjustment formula for the adaptive window is: ;

[0008] Where int is a rounding function, min_win and max_win are the preset minimum and maximum window sizes, respectively, and target_std is the standard deviation of the target variable in the current data segment.

[0009] Furthermore, feature data matrix X' and target variable Y are extracted from the window subset of each partition. For the feature data within each window, the mutual information value between each feature and the target variable is calculated as the importance score of that feature in the current window. The formula for calculating the importance score is as follows: ; in and These are the marginal probability distributions of variables X' and Y, respectively. It is a joint probability distribution.

[0010] Furthermore, the feature importance scores calculated within the window are normalized, and groups are formed based on the similarity between feature importance scores. Cosine similarity is used to calculate the cosine similarity matrix of feature importance vectors between different windows. The normalization formula is as follows: ; in It is a feature importance vector. It is the normalized feature importance vector; The formula for calculating cosine similarity is: ; ; ; Where C is the centroid of the pattern classification, and k is the number of clusters. is the similarity matrix; D is the distance matrix, and i and j are the sequence numbers of the data segments.

[0011] Furthermore, a suitable sample size M is selected. When the sample size of a certain pattern is less than M, upsampling is used to supplement the data. When the sample size of the pattern is greater than M, downsampling is used to select the M most representative samples as new data samples. For the feature data of each pattern... Its sample is , ,…, If mi < M, use linear interpolation for upsampling. The linear interpolation formula is: ; The original sample and These are two adjacent samples, where j is the new sample index, and j is between l and l+1. This is the new sample after interpolation. After upsampling, a new feature data matrix is ​​obtained.

[0012] If mi > M, then downsampling is used for data alignment, and the calculation formula is: ; ; in The mean is used as a representative measure. The sample was calculated using the Euclidean distance formula. and The distance between them Represents the L2 norm; To ensure the diversity of the downsampled samples, considering the differences between samples and selecting the sample farthest from the selected sample, the formula for calculating the average distance between the candidate sample and the selected sample is as follows: ; Where s is the index of the selected sample. Represents each candidate sample, This indicates the selected sample.

[0013] Furthermore, the new feature data obtained in each pattern Normalization is performed to generate a time series of length L, which is then input into the encoder. The normalization formula is: ; Where i corresponds to different modes, and d is the dimension. and These are the mean and standard deviation of the feature, respectively. These are the normalized feature data; For normalized feature data and target data Generate a time series of length L. and ,in: ; Multiple data inputs are fed into the GMFTrans encoder, which uses a shared encoder strategy to extract feature representations from data in different modes. The encoder's output in different modes is represented as follows: ; Where t is the time step and i corresponds to different modes.

[0014] Furthermore, the formula for the MLP output is: ; Where p is the number of layers in the MLP. This represents the concatenated cross-modal input features. , , It is the output of the hidden layer. , , It's weight. , , It is a bias term. and It is a non-linear activation function of the hidden layer. This represents the final output feature of the multilayer perceptron.

[0015] Furthermore, the fused feature information is input into the decoder to predict the target value. The prediction formula is as follows: ; ;

[0016] in Zy is the dimension of the fused features, and Zy is the representation of the target sequence after positional encoding and self-attention mechanism. It is the fused multimodal feature representation. This represents the cross-attention mechanism. It is the output after cross-attention. This represents the final prediction result of the model. FC is a fully connected layer, LN is a normalization layer, and FFN is a feedforward network layer.

[0017] Due to the application of the above-mentioned technical solution, the present invention has the following advantages compared with the prior art: The GMFTrans-based multimodal prediction framework offers more accurate modality segmentation capabilities. It employs the Feature Adaptive Multimodal Partitioning (FAMP) method, dynamically partitioning data based on feature importance and similarity, significantly improving the adaptability of pattern classification. Furthermore, this framework utilizes GMFTrans's multi-encoder structure to extract modality-specific key features, effectively captures cross-modal dependencies using a multilayer perceptron (MLP), and dynamically fuses multidimensional information through a gating mechanism, fully leveraging intra-modal and inter-modal information. Therefore, the proposed GMFTrans-based multimodal modeling framework demonstrates higher prediction accuracy and robustness in complex petrochemical processes. Attached Figure Description

[0018] Figure 1 This is a flow chart of the catalytic cracking regeneration flue gas purification process in this invention; Figure 2 This is a process flow diagram of the denitrification process in this invention; Figure 3 This is a process flow diagram of the desulfurization process in this invention; Figure 4 This is a flowchart of the multimodal prediction model based on GMFTrans of this invention; Figure 5 This is a diagram of the multimodal prediction model architecture based on GMFTrans of this invention; Figure 6 This is a schematic diagram of data mode partitioning based on the FAMP method of this invention; Figure 7 The graph shows the prediction results of NOx using different methods of this invention. Figure 8 This is a comparison chart of the errors in predicting NOx using different methods of this invention; Figure 9 This invention uses different methods to predict... Marginal histogram; Figure 10 This invention uses different methods to predict... The absolute error distribution diagram. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or system that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or systems.

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] Figure 1 This application describes the relevant process for catalytic cracking regeneration flue gas purification. The flue gas treatment mainly involves two key steps: denitrification and desulfurization. NOx in the flue gas is treated by a reduction unit to generate environmentally harmless nitrogen and water, while acidic components such as sulfur dioxide are removed after passing through an absorption unit. The process involves the removal of NOx and... Predicting the content of [specific substances] is of great significance and necessity in determining whether flue gas meets emission standards.

[0023] Figure 2 This diagram illustrates a simplified representation of the denitrification process unit. The NOx removal procedure for the regenerated waste gas mainly consists of two parts: an ammonia delivery section and a catalytic reduction section. Through a series of chemical reactions, the high concentration of NOx in the regenerated flue gas is ultimately converted into non-toxic nitrogen and water, achieving clean production in the chemical process.

[0024] Figure 3 This illustrates a simplified representation of the desulfurization process device. The sulfur-containing flue gas first undergoes pretreatment processes such as dust removal and cooling, and then enters the absorption tower. Sodium hydroxide is made into a slurry and sprayed on the gas. In the absorption tower, it undergoes an acid-base neutralization reaction with acidic gases such as sulfur dioxide to generate calcium sulfite. The calcium sulfite is eventually oxidized and crystallized to form calcium sulfate.

[0025] Figure 4 The flowchart of the multimodal prediction model based on GMFTrans proposed in this invention is illustrated, including the following steps: S1. Collect data on flue gas treatment at the refinery. Preprocess the collected data to eliminate outliers and perform smoothing. The interquartile range (IQR) method is used for outlier detection, and a Savitzky-Golay filter is used for smoothing. An adaptive window is used to partition the data; the window size can be adjusted according to data volatility. The formula for adjusting the adaptive window is: (1); (2); Where int is a rounding function, min_win and max_win are the preset minimum and maximum window sizes, respectively, and target_std is the standard deviation of the target variable in the current data segment.

[0026] S2. Extract the feature data matrix X' and the target variable Y from the window subset of each partition. For the feature data within each window, calculate the mutual information value between each feature and the target variable, which serves as the importance score for that feature in the current window. The importance score calculation formula is as follows: (3); in and These are the marginal probability distributions of variables X' and Y, respectively. It is a joint probability distribution.

[0027] S3. Normalize the feature importance scores calculated within the window. Features with higher importance scores have a greater impact on the target variable under the production mode represented by the current window. Group features based on the similarity between their importance scores, and calculate the cosine similarity matrix of feature importance vectors between different windows using cosine similarity. The normalization formula is as follows: (4); in It is a feature importance vector. It is the normalized feature importance vector.

[0028] The formula for calculating cosine similarity is as follows: (5); (6); (7); Where C is the centroid of the pattern classification, and k is the number of clusters. It is a similarity matrix; D is the distance matrix, and i and j are the sequence numbers of the data segments.

[0029] The similarity matrix is ​​obtained by calculating the cosine similarity. Based on the similarity matrix, the distance matrix D is calculated. Finally, for j data segments, their distances to the initial class centers are calculated, and they are assigned to the nearest group. Through continuous iterative updates, the final partitioning result is obtained.

[0030] S4. Construct a multi-mode partitioning module (FAMP) based on S1, S2, and S3, as shown in the following diagram. Figure 6 As shown.

[0031] S5. Select an appropriate sample size M. When the sample size of a certain pattern is less than M, upsampling is used to supplement the data; when the sample size of the pattern is greater than M, downsampling is used to select the M most representative samples as new data samples. The dataset X is divided into multiple patterns, and the division result can be represented as X1, X2, ... XN. For the feature data of each pattern... Its sample is , ,…, If mi < M, use linear interpolation for upsampling. The linear interpolation formula is as follows: (8); The original sample and These are two adjacent samples, where j is the new sample index, and j is between l and l+1. This is the new sample after interpolation. After upsampling, a new feature data matrix is ​​obtained.

[0032] If mi > M, then downsampling is used for data alignment, and the mean is calculated for all samples in each pattern. As a representative metric, samples are calculated using Euclidean distance. and The distance between them. The calculation formula is: (9); (10); in The mean is used as a representative measure. The sample was calculated using the Euclidean distance formula. and The distance between them This represents the L2 norm.

[0033] To ensure the diversity of the downsampled samples, after selecting the representative sample closest to the mean, the differences between samples are further considered, and the sample furthest from the selected sample is chosen. The formula for calculating the average distance between the candidate sample and the selected sample is as follows: (11); Where s is the index of the selected sample. Represents each candidate sample, This indicates the selected sample.

[0034] S6. New feature data obtained in each pattern Normalization is performed to generate a time series of length L, which is then input into the encoder. The normalization formula is as follows: (12); Where i corresponds to different modes, and d is the dimension. and These are the mean and standard deviation of the feature, respectively. These are the normalized feature data.

[0035] For normalized feature data and target data Generate a time series of length L. and .in: ; Data containing multiple modes is input into the GMFTrans encoder, and feature representations are extracted from the data across different modes using a shared encoder strategy. The encoder's output in different modes is represented as follows: (13); Where t is the time step and i corresponds to different modes.

[0036] S7. A Multilayer Perceptron (MLP) is introduced to fuse features from data across different modes. To learn cross-modal information, data from different modes are concatenated to obtain an input data vector that integrates multiple modes. This is then input into a multilayer perceptron (MLP). The output formula of the MLP is: (14); (15); (16); Where p is the number of layers in the MLP. This represents the concatenated cross-modal input features. , , It is the output of the hidden layer. , , It's weight. , , It is a bias term. and It is a non-linear activation function of the hidden layer. This represents the final output feature of the multilayer perceptron.

[0037] S8. A gating mechanism is introduced to fuse information from different encoder outputs and multilayer perceptron (MLP) outputs. To fully extract correlations within the same pattern and consider the interactions between different patterns in the process data, a gating mechanism is used to fuse information from different encoder outputs and MLP outputs. A dynamic weight is assigned to each output, and the learned weights are used to perform a weighted summation of the feature representations from each encoder and the MLP output. The fused feature representation is as follows: (17); (18); avgpool performs average pooling on information from different encoders. It uses the sigmoid activation function; FC1 and FC2 are fully connected layers. It is the feature representation after fusion.

[0038] S9. Input the fused feature information into the decoder to predict the target value. The prediction formula is: (19); (20); in Zy is the dimension of the fused features, and Zy is the representation of the target sequence after positional encoding and self-attention mechanism. It is the fused multimodal feature representation. This represents the cross-attention mechanism. It is the output after cross-attention. This represents the final prediction result of the model. FC is a fully connected layer, LN is a normalization layer, and FFN is a feedforward network layer.

[0039] S10. A multimodal prediction model based on GMFTrans is proposed for prediction, and its structure diagram is shown below. Figure 5 As shown.

[0040] By analyzing NOx and other substances in the catalytic cracking regeneration flue gas purification process To predict the content and verify the accuracy and feasibility of the model.

[0041] a. Catalytic cracking flue gas denitrification process: Catalytic cracking is a crucial process in petrochemicals. The flue gas produced during this process contains NOx and SOx, which can cause environmental pollution. Therefore, the flue gas must undergo denitrification and desulfurization treatment before safe discharge. The denitrification process primarily removes NOx from the regeneration waste gas. Figure 2 This is a schematic diagram of the denitrification process, mainly composed of an ammonia transport section and a catalytic reduction section. Through a series of chemical reactions, the high concentration of NOx in the regenerated flue gas is ultimately converted into non-toxic nitrogen and water, achieving clean production in the chemical process. A one-year dataset of regenerated flue gas denitrification from a refinery was selected, with sampling frequency once per hour, totaling over 4300 samples. The first 3700 samples were used as the training set, and the last 680 samples as the test set, facilitating subsequent modeling of the multi-mode process data. Based on the principle of nitrogen oxide reduction, 12 factors related to the denitrification process were selected as key parameters, as shown in Table 1. The parameters of the proposed multi-modal prediction model based on GMFTrans are shown in Table 2.

[0042] Table 1 Input and output variables of the denitrification process.

[0043]

[0044] Table 2. Hyperparameter settings.

[0045]

[0046] b. Analysis of denitrification results: In this study, the FAMP method was used to divide the training set into different modes, and the top seven feature variables in each mode were selected based on feature importance and fed into the model for prediction. To measure the effectiveness of the model, MSE, RMSE, MAE, SMAPE, and R2 were used as evaluation metrics. Under the same experimental conditions, it was compared with nine other models: MrCAN, AR-TCN, FS-PC-Attention, FS-PC-LSTM, CNN-LSTM, GRU-LSTM, LSTM-Transformer, DSTP-RNN, and RBFNN.

[0047] Table 3 Performance metrics of prediction results from different models.

[0048]

[0049] As shown in Table 3, compared with the other nine methods, the multimodal modeling framework based on GMFTrans achieved the best performance in all evaluation metrics in this study. It had the highest R² (0.897) and the lowest MSE (5.343).

[0050] Figure 7The graph shows the prediction results of each model on the test set. It can be seen that the RBFNN, DSTP-RNN, and GRU-LSTM models have poor fitting performance on the flue gas denitrification dataset, with significant deviations between the predicted and actual curves. AR-TCN, CNN-LSTM, LSTM-Transformer, and FS-PC-LSTM generally show better prediction trends, but significant deviations exist in certain areas. The FS-PC-Attention multi-modal prediction framework and MrCAN show relatively high good fits between the actual and predicted values. In this study, the multi-modal modeling framework based on GMFTrans exhibits the best prediction performance. Figure 8 The graph shows the prediction errors of each model, revealing discrepancies between the predicted and actual values. Compared to other models, the multi-model prediction framework based on GMFTrans exhibits smaller fluctuations in its error curves, demonstrating better prediction performance.

[0051] c. Catalytic cracking flue gas desulfurization process: Desulfurization is an essential part of catalytic cracking flue gas emissions. Figure 3 The diagram illustrates the desulfurization process. Sulfur-containing flue gas first undergoes pretreatment processes such as dust removal and cooling, then enters the absorption tower. Sodium hydroxide is prepared as a slurry and sprayed onto the gas. In the absorption tower, it undergoes an acid-base neutralization reaction with acidic gases such as sulfur dioxide to generate calcium sulfite. The calcium sulfite is ultimately oxidized and crystallized to form calcium sulfate, thus ensuring that the flue gas meets emission standards. In this study, two years of desulfurization data from the catalytic cracking regeneration flue gas of a refinery were selected, with a sampling frequency of once per hour, resulting in a total sample size of approximately 8700. The last 2000 data points were selected as the test set. Based on the desulfurization process, 13 variables were selected as feature variables, as shown in Table 4. The parameters of the proposed multimodal prediction model based on GMFTrans are shown in Table 5.

[0052] Table 4 Input and output variables of the desulfurization process.

[0053]

[0054] Table 5. Hyperparameter settings.

[0055]

[0056] d. Analysis of desulfurization results: In this study, the FAMP method was used to classify the desulfurization dataset into patterns. Then, feature selection was performed on the segmented multi-pattern data, selecting the top 8 important variables for each pattern and feeding them into the GMFTrans model. Comparative experiments were conducted with 9 other models. To measure the effectiveness of the model, MSE, RMSE, MAE, SMAPE, and R² were used as evaluation metrics.

[0057] Table 6 Performance metrics of prediction results from different models.

[0058]

[0059] Table 6 shows the prediction results of different models on the desulfurization dataset. As can be seen from the table, compared with other advanced time series prediction models, the framework proposed in this paper has the best prediction performance, with an MSE value of 0.028, an MAE value of 0.088, and an R² as high as 0.991.

[0060] Figure 9 Marginal histograms were plotted between the actual and predicted values ​​for each model, with a control group provided for comparison. The graphs show that RBFNN exhibits relatively high prediction error and poor fitting performance. DSTP-RNN, AR-TCN, GRU-LSTM, CNN-LSTM, and LSTM-Transformer show similar prediction results; the scatter plots of predicted and actual values ​​are closely connected near the fitted line, but the fitted line deviates slightly from the diagonal. FS-PC-LSTM has more points scattered outside the fitted line, leading to a decrease in overall prediction performance. The FS-PC-Attention framework and MrCAN show smaller prediction biases compared to previous models. In contrast, the proposed GMFTrans-based multimodal modeling framework, while exhibiting bias in predictions on a small subset of data, achieves the best overall prediction performance. Figure 10 The absolute error distribution of different models is presented. As can be seen from the figure, the multi-mode prediction framework based on GMFTrans has the best performance and the smallest absolute error, which is close to 0.

[0061] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0062] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to the above embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0063] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for multimodal prediction of industrial flue gas based on a gated multidimensional information fusion transformer, characterized in that, include: Data on flue gas treatment in a refinery is collected, preprocessed, and divided into subsets of windows for each partition based on an adaptive window. In each partition's window subset, feature data and target variables are extracted. For the feature data within each window, the mutual information value between each feature and the target variable is calculated and used as the importance score of the feature in the current window. The feature importance scores calculated within the window are normalized, and the features are grouped based on the similarity between their importance scores. The multi-mode partitioning module generates partitioning results, normalizes the new feature data obtained in each mode, generates a time series of length L, and inputs it into the encoder to obtain the encoder output features. The encoder output features corresponding to different modes are concatenated along the feature dimension, and the concatenated features are input into a multilayer perceptron (MLP) to learn the cross-modal correlation between different modes and obtain cross-modal feature information. A gating mechanism is introduced to fuse the output features of different encoders and the information from the output of a multilayer perceptron (MLP) to obtain fused feature information. The fused feature information is input into the decoder, which outputs the predicted target value.

2. The industrial flue gas multimodal prediction method based on a gated multidimensional information fusion transformer as described in claim 1, characterized in that, The formula for adjusting the adaptive window is: ； Where int is the rounding function, min_win and max_win are the preset minimum and maximum window sizes, respectively, and target_std is the standard deviation of the target variable in the current data segment.

3. The industrial flue gas multimodal prediction method based on a gated multidimensional information fusion transformer as described in claim 1, characterized in that, In each partition's window subset, feature data matrix X' and target variable Y are extracted. For the feature data within each window, the mutual information value between each feature and the target variable is calculated as the importance score of that feature in the current window. The formula for calculating the importance score is as follows: ； in and These are the marginal probability distributions of variables X' and Y, respectively. It is a joint probability distribution.

4. The industrial flue gas multimodal prediction method based on a gated multidimensional information fusion transformer as described in claim 3, characterized in that, The feature importance scores calculated within the window are normalized, and groups are formed based on the similarity between feature importance scores. The cosine similarity matrix of feature importance vectors between different windows is calculated using cosine similarity. The normalization formula is as follows: ； in It is a feature importance vector. It is the normalized feature importance vector; The formula for calculating cosine similarity is: ； ； ； Where C is the centroid of the pattern classification, and k is the number of clusters. It is a similarity matrix. D is the distance matrix, and i and j are the sequence numbers of the data segments.

5. The industrial flue gas multimodal prediction method based on a gated multidimensional information fusion converter as described in claim 4, characterized in that, Choose an appropriate sample size M. When the sample size of a certain pattern is less than M, upsampling is used to supplement the data. When the sample size of the pattern is greater than M, downsampling is used to select the M most representative samples as new data samples. For the feature data of each pattern... Its sample is , ,…, If mi < M, use linear interpolation for upsampling. The linear interpolation formula is: ； The original sample and These are two adjacent samples, where j is the new sample index, and j is between l and l+1. It is the new sample after interpolation. After upsampling, a new feature data matrix is ​​obtained. If mi > M, then downsampling is used for data alignment, calculated using the following formula: ； ； in The mean is used as a representative measure. The sample was calculated using the Euclidean distance formula. and The distance between them Represents the L2 norm; To ensure the diversity of the downsampled samples, considering the differences between samples and selecting the sample farthest from the selected sample, the formula for calculating the average distance between the candidate sample and the selected sample is as follows: ； in This indicates the number of samples selected, and s represents the index of the selected sample. Represents each candidate sample, This indicates the selected sample.

6. The industrial flue gas multimodal prediction method based on a gated multidimensional information fusion transformer as described in claim 1, characterized in that, New feature data obtained in each pattern Normalization is performed to generate a time series of length L, which is then input into the encoder. The normalization formula is: ； Where i corresponds to different modes, and d is the dimension. and These are the mean and standard deviation of the feature, respectively. These are the normalized feature data; For normalized feature data and target data Generate a time series of length L. and ,in: Multiple data inputs are fed into the GMFTrans encoder, which uses a shared encoder strategy to extract feature representations from data in different modes. The encoder's output in each mode is represented as follows: ； Where t is the time step and i corresponds to different modes. This represents the encoder network corresponding to the i-th mode, used to extract the temporal feature representation of that mode.

7. The industrial flue gas multimodal prediction method based on a gated multidimensional information fusion transformer as described in claim 1, characterized in that, The formula output by the MLP is ； Where p is the number of layers in the MLP. This represents the concatenated cross-modal input features. , , It is the output of the hidden layer. , , It's weight. , , It is a bias term. and It is a non-linear activation function of the hidden layer. This represents the final output feature of the multilayer perceptron.

8. The industrial flue gas multimodal prediction method based on a gated multidimensional information fusion converter as described in claim 1, characterized in that, The fused feature information is input into the decoder to predict the target value. The prediction formula is as follows: ； ； in Zy is the dimension of the fused features, and Zy is the representation of the target sequence after positional encoding and self-attention mechanism. It is the fused multimodal feature representation. This represents the cross-attention mechanism. It is the output after cross-attention. This represents the final prediction result of the model. FC is a fully connected layer, LN is a normalization layer, and FFN is a feedforward network layer.

Images