Ultralow-temperature denitration catalyst activity prediction method and system based on multi-modal data

Abstract

The invention discloses an ultralow-temperature denitration catalyst activity prediction method and system based on multi-modal data, and belongs to the technical field of catalyst activity prediction.The method comprises the steps that multi-modal original data of smoke components, reaction temperature and a catalyst microstructure are collected and preprocessed; constructing an activity attenuation model based on a graph neural network, defining a core through graph structure mapping, associating nodes and in-degree and out-degree nodes, designing a node interaction rule, constructing a GNN architecture containing four layers of graph convolution, and training and optimizing; and inputting preprocessed data to obtain the current activity and a future change trend, correcting an error by combining a residual error model, dynamically adjusting a data acquisition period when the error exceeds a limit, outputting a result containing an activity current value, an attenuation rate and residual life, and triggering early warning when the activity reaches a threshold value. The system adapts to the ultralow-temperature denitration working condition, the prediction precision and stability are improved, a scientific basis is provided for catalyst maintenance, and emission exceeding and resource waste are avoided.

Description

Technical Field

[0001] This invention relates to the field of catalyst activity prediction technology, and in particular to a method and system for predicting the activity of ultra-low temperature denitrification catalysts based on multimodal data. Background Technology

[0002] Ultra-low temperature denitrification technology is widely used in flue gas purification in industries such as steel, power, and waste incineration. Its core principle is to convert nitrogen oxides in flue gas into harmless nitrogen and water using a denitrification catalyst at ultra-low temperatures of 80-150℃. The denitrification efficiency of this technology directly depends on the activity state of the catalyst. With the increasing complexity of industrial flue gas composition, dynamic changes in reaction temperature, and the deterioration of the catalyst's microstructure after long-term operation, catalyst activity will continuously decline. If changes in activity cannot be accurately predicted, it can easily lead to excessive nitrogen oxide emissions and resource waste caused by untimely or premature catalyst replacement.

[0003] Existing catalyst activity prediction technologies suffer from the following key problems: First, the data dimension is limited. Most methods rely solely on single-modal data such as flue gas composition or reaction temperature, neglecting the coupling effect between flue gas composition, reaction temperature, and catalyst microstructure. For example, increased sulfur dioxide concentration accelerates the degradation of catalyst microstructure, thereby exacerbating activity decay and resulting in incomplete feature characterization. Second, the model architecture cannot capture the dynamic diffusion law of activity decay. Traditional machine learning models, such as multilayer perceptrons and convolutional neural networks, struggle to simulate the dynamic transfer process of harmful components, exhibiting poor generalization performance. Third, they cannot distinguish the impact of different features on activity, further reducing prediction accuracy.

[0004] Therefore, this invention proposes a method and system for predicting the activity of ultra-low temperature denitrification catalysts based on multimodal data. Summary of the Invention

[0005] To address at least one of the aforementioned technical problems, this invention provides a method and system for predicting the activity of ultra-low temperature denitrification catalysts based on multimodal data.

[0006] In a first aspect, the present invention provides a method for predicting the activity of an ultra-low temperature denitrification catalyst based on multimodal data, the method comprising:

[0007] Step 1: Collect multimodal raw data under ultra-low temperature denitrification scenario, and preprocess the multimodal raw data to obtain a standardized feature set. The multimodal raw data includes flue gas composition data, reaction temperature data, and catalyst microstructure data.

[0008] Step 2: Construct a catalyst activity decay model based on a graph neural network. The model construction process includes:

[0009] Step 2.1: Graph Structure Data Mapping: The standardized feature set is mapped to graph structure data that conforms to the dynamic diffusion law of activity decay. The real-time activity state of the catalyst is used as the core node, and the features corresponding to the flue gas composition, reaction temperature, and catalyst microstructure are respectively used as associated nodes in the graph. Feature nodes that have a direct effect on the activity state of the core node are defined as in-degree nodes, and feature nodes affected by the activity state of the core node are defined as out-degree nodes. A close association distance is assigned to the connection between each associated node and the core node, and the close association distance is determined by the degree of direct correlation between the corresponding feature and the activity decay.

[0010] Step 2.2: Node interaction rule design: Combine the differences between the features of in-degree nodes and core nodes, and the differences between the features of out-degree nodes and core nodes, and adjust the weight of the feature differences based on the closeness of association. At the same time, assign a feature collaboration strength coefficient to the connection between in-degree nodes, and assign a feature action rate coefficient to the connection between out-degree nodes and core nodes.

[0011] Step 2.3: GNN network architecture design: including: input layer, feature interaction layer, graph convolutional layer, fully connected layer and output layer;

[0012] Step 2.4: Model Training and Parameter Optimization: Based on historical activity decay data, the model is divided into training, validation, and test sets. The training set is used to learn the mapping relationship between graph structure nodes and activity states. The validation set is used to adjust the number of graph convolutional layers and feature interaction weights. The test set is used to evaluate the model's generalization performance.

[0013] Step 3: Input the preprocessed data into the trained catalyst activity decay model to obtain the current activity state of the catalyst and the predicted value of the activity change trend in the future preset period.

[0014] Preferably, the flue gas composition data includes the concentration of harmful components that inhibit catalyst activity and the concentration of essential components that support the denitrification reaction; the reaction temperature data includes the real-time temperature of different regions of the catalyst bed and the degree of temperature fluctuation per unit time; and the catalyst microstructure data includes structural parameters reflecting the total amount of active sites and distribution parameters reflecting the effective proportion of active sites.

[0015] Preferably, the GNN network architecture design includes:

[0016] The input layer receives graph structure data and passes it to the feature interaction layer;

[0017] The feature interaction layer, based on node interaction rules, combines the close association distance, feature collaboration strength coefficient, and feature action rate coefficient to complete the initial feature fusion of in-degree nodes, out-degree nodes, and core nodes.

[0018] The graph convolutional layer is set to 4 layers. Each layer prioritizes aggregating features in the in-degree nodes that are strongly correlated with activity decay, and through the interaction of feature differences between neighboring nodes, it explores the hidden activity decay diffusion patterns among multimodal data.

[0019] The fully connected layer converts the feature vector output by the graph convolutional layer into the dimension corresponding to the active state;

[0020] The output layer outputs the predicted value of catalyst activity decay.

[0021] Preferably, it further includes: correcting the activity prediction value through an error correction mechanism, specifically including:

[0022] Construct a residual model adapted to the characteristics of activity decay bias, and learn the correlation between multimodal feature combination and prediction bias;

[0023] The original predicted value is adjusted by using the deviation compensation value output by the residual model and combining it with the characteristic difference direction of the current working condition.

[0024] A 2% prediction error threshold is set, and the correction results are monitored in real time. When the error exceeds the limit continuously, the microstructure data acquisition cycle is shortened and the error is controlled to be stable below the prediction error threshold. The final output includes the current activity value, decay rate, and remaining lifetime. When the activity drops to the preset activity threshold, a lifetime warning is triggered.

[0025] Preferably, shortening the microstructure data acquisition cycle and controlling the error to remain stable below the prediction error threshold includes:

[0026] The number of consecutive exceedances is obtained, and combined with the current catalyst activity decay stage, the characteristic type corresponding to the current error source is identified and the shortening period is determined.

[0027] For the microstructure data collected after shortening the cycle, features related to the current error source are extracted in a targeted manner to form a high-frequency microstructure feature subset. The correlation distance between graph structure nodes and core nodes, as well as the feature coordination strength coefficient between in-degree nodes, are adjusted to achieve stable control.

[0028] Preferably, a feature collaboration strength coefficient is assigned to the connection between in-degree nodes, and a feature action rate coefficient is assigned to the connection between out-degree nodes and core nodes, including:

[0029] Based on the close correlation distance, the difference between the features of the in-degree node and the features of the core node is assigned a basic fusion weight according to the rule that the smaller the distance, the greater the weight.

[0030] The feature collaboration strength coefficient between the corresponding in-degree node and other in-degree nodes is superimposed to correct the basic fusion weights;

[0031] Based on each out-degree node, the weight of the feature transferred from the core node to the out-degree node is set according to the rule that the larger the rate coefficient, the higher the weight of the transfer. At the same time, the weight of the transfer is calibrated again by taking into account the close relationship between the out-degree node and the core node.

[0032] Real-time monitoring of the fluctuation amplitude of multimodal features under the current working conditions. If the fluctuation amplitude of the corresponding in-degree node feature exceeds the preset threshold, the close association distance weight of the corresponding node will be temporarily increased. If the deviation between the change rate of the corresponding out-degree node feature and the activity decay rate of the core node exceeds the preset range, the feature action rate coefficient of the corresponding node will be temporarily adjusted.

[0033] The processed in-degree node fusion features and processed out-degree node fusion features are weighted and aggregated with the initial activity features of the core nodes to obtain the output result after preliminary fusion of multimodal features.

[0034] Preferably, the in-degree node features of the first layer of the convolutional layer related to catalyst poisoning, the in-degree node features of the second layer of the convolutional layer related to structural aging, the in-degree node features of the third layer of the convolutional layer related to reaction mass transfer, and the synergistic combination of the features of the first, second and third layers of the convolutional layer in the fourth layer of the convolutional layer.

[0035] The first layer of the convolutional layer explores the inhibition of activity by the adsorption and diffusion of harmful components on the catalyst surface. The second layer of the convolutional layer explores the erosion of active sites by the spatial diffusion of microstructural degradation. The third layer of the convolutional layer explores the constraint of the reaction temperature field distribution diffusion on the catalytic reaction rate. The fourth layer of the convolutional layer integrates the rules of the first, second, and third layers to explore the global diffusion law of activity decay under the coupling of multiple factors.

[0036] Secondly, a system for predicting the activity of ultra-low temperature denitrification catalysts based on multimodal data includes:

[0037] The preprocessing module is used to collect multimodal raw data under ultra-low temperature denitrification scenario and preprocess the multimodal raw data to obtain a standardized feature set. The multimodal raw data includes flue gas composition data, reaction temperature data and catalyst microstructure data.

[0038] A model building module is used to build a catalyst activity decay model based on a graph neural network, wherein the model building module includes:

[0039] The graph structure data mapping unit is used to map the standardized feature set into graph structure data that conforms to the dynamic diffusion law of activity decay. The real-time activity state of the catalyst is used as the core node, and the features corresponding to the flue gas composition, reaction temperature, and catalyst microstructure are respectively used as the associated nodes in the graph. Feature nodes that have a direct effect on the activity state of the core node are defined as in-degree nodes, and feature nodes affected by the activity state of the core node are defined as out-degree nodes. Each associated node and the core node is assigned a close association distance, and the close association distance is determined by the degree of direct correlation between the corresponding feature and the activity decay.

[0040] The node interaction rule design unit is used to combine the differences between the features of in-degree nodes and core nodes, and the differences between the features of out-degree nodes and core nodes, and adjust the weight of the feature differences based on the closeness of association. At the same time, it assigns a feature collaboration strength coefficient to the connection between in-degree nodes and a feature action rate coefficient to the connection between out-degree nodes and core nodes.

[0041] The GNN network architecture design unit includes: input layer, feature interaction layer, graph convolutional layer, fully connected layer and output layer;

[0042] The model training and parameter optimization unit is used to divide the training set, validation set, and test set based on historical activity decay data. The training set is used to learn the mapping relationship between graph structure nodes and activity states, the validation set is used to adjust the number of graph convolutional layers and feature interaction weights, and the test set is used to evaluate the model's generalization performance.

[0043] The activity prediction module is used to input the preprocessed data into the trained catalyst activity decay model to obtain the current activity state of the catalyst and the predicted value of the activity change trend within a preset period of time.

[0044] Thirdly, the present invention also provides an electronic device, comprising: a processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein when the processor executes the computer instructions, the electronic device performs the method as described in the first aspect above and any possible implementation thereof.

[0045] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor of an electronic device, cause the processor to perform a method as described in the first aspect above and any possible implementation thereof.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] By collecting multimodal data on flue gas, temperature, and microstructure, the key factors affecting catalyst activity are comprehensively covered, avoiding feature loss from single-modal data. The activity decay model built based on GNN accurately captures the dynamic diffusion law of activity decay through graph structure mapping and hierarchical convolution. Compared with traditional machine learning models, the prediction accuracy is improved by 15%-20%. The model training process adopts hierarchical sampling and cross-validation, which has excellent generalization performance and can adapt to different ultra-low temperature denitrification conditions, providing accurate data support for catalyst replacement and maintenance.

[0048] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the background art, the accompanying drawings used in the embodiments of the present invention or the background art will be described below.

[0050] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the specification, serve to illustrate the technical solutions of this disclosure.

[0051] Figure 1 A flowchart illustrating a method for predicting the activity of ultra-low temperature denitration catalysts based on multimodal data, provided in a certain embodiment of the present invention;

[0052] Figure 2 A structural diagram of an ultra-low temperature denitrification catalyst activity prediction system based on multi-modal data provided in a certain embodiment of the present invention. Detailed Implementation

[0053] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. 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 apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0055] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0056] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0057] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without certain specific details. In some instances, methods, means, elements, and circuits well known to those skilled in the art have not been described in detail in order to highlight the spirit of the invention.

[0058] Firstly, a method for predicting the activity of ultra-low temperature denitrification catalysts based on multimodal data, such as... Figure 1 As shown, it includes:

[0059] Step 1: Collect multimodal raw data under ultra-low temperature denitrification scenario, and preprocess the multimodal raw data to obtain a standardized feature set. The multimodal raw data includes flue gas composition data, reaction temperature data, and catalyst microstructure data.

[0060] In this embodiment, the ultra-low temperature denitrification scenario refers to the denitrification reaction scenario in which the flue gas temperature is maintained at 80-150℃. It is commonly seen in the end-of-pipe purification stage of sintering machine flue gas in the steel industry and waste incineration flue gas. In this scenario, the catalyst activity is easily affected by factors such as the adsorption of harmful components at low temperature and the slow reaction rate.

[0061] In this embodiment, the flue gas composition data refers to the concentration data of components in the flue gas that directly affect the catalyst activity, including harmful components such as… Essential components include .

[0062] In this embodiment, the reaction temperature data refers to the real-time temperature and temperature fluctuation of different areas of the denitrification catalyst bed. Three sets of thermocouples are arranged in the catalyst bed along the flue gas flow direction, namely one set each at the inlet, middle and outlet.

[0063] In this embodiment, catalyst microstructure data refers to structural parameters that reflect the number and effectiveness of catalyst active sites, such as total active site parameters including specific surface area and active component loading; and active site effective proportion parameters including active site dispersion and pore size distribution.

[0064] In this embodiment, preprocessing and standardizing the feature set refers to performing outlier removal, missing value imputation, and feature standardization on the multimodal raw data to eliminate differences in data units and noise interference, resulting in a standardized feature set that can be directly input into the model. Outlier removal involves: using... The guidelines will Data exceeding the mean ± 3 × standard deviation in concentration were removed; missing value imputation: linear interpolation was used for missing values ​​in temperature data. For example, if the temperature at 10:00 was missing, it was interpolated from 88℃ at 9:59 and 89℃ at 10:01 to obtain 88.5℃; standardization: Min-Max normalization was used to map the feature values ​​to the interval between 0 and 1.

[0065] Step 2: Construct a catalyst activity decay model based on a graph neural network. The model construction process includes:

[0066] In this embodiment, Graph Neural Network (GNN) is a deep learning model specifically designed for processing graph-structured data. It can mine the topological relationships of data through the connections between nodes and is suitable for capturing the dynamic diffusion law of multimodal features and activity decay, such as the transmission relationship of harmful components → microstructure deterioration → activity decrease.

[0067] Step 2.1: Graph Structure Data Mapping: The standardized feature set is mapped to graph structure data that conforms to the dynamic diffusion law of activity decay. The real-time activity state of the catalyst is used as the core node, and the features corresponding to the flue gas composition, reaction temperature, and catalyst microstructure are respectively used as associated nodes in the graph. Feature nodes that have a direct effect on the activity state of the core node are defined as in-degree nodes, and feature nodes affected by the activity state of the core node are defined as out-degree nodes. A close association distance is assigned to the connection between each associated node and the core node, and the close association distance is determined by the degree of direct correlation between the corresponding feature and the activity decay.

[0068] In this embodiment, the graph structure data consists of nodes and edges. Nodes represent multimodal features or activity states, and edges represent the relationships between nodes. In this invention, the graph structure must conform to the dynamic diffusion law of activity decay, i.e., the transmission direction from features affecting activity → activity → features affected by activity. For example, the graph structure contains one core node (real-time catalyst activity, such as a denitrification efficiency of 85%) and nine associated nodes (four components of the flue gas: ...). ; 3 reaction temperatures: inlet, middle, and outlet temperatures; 2 microstructures: specific surface area and dispersion); edges represent the relationships between nodes (e.g., Edge representation of node → core node (Inhibition of activity). Based on the NetworkX library, a graph structure is constructed. The node feature matrix has a dimension of 10×12 (10 nodes, each node corresponds to the association feature in the 12 standardized features), and the adjacency matrix has a dimension of 10×10 (1 indicates that there is a connection between nodes, 0 indicates that there is no connection, and the weight of the edge is the close association distance).

[0069] In this embodiment, the core node is a node in the graph structure that focuses on the real-time activity state of the catalyst. It is the prediction target of the activity decay model, and its characteristic value is the denitrification efficiency of the catalyst under the current operating conditions. The characteristic value of the core node is calculated by the NOx concentration at the inlet and outlet of the denitrification system (denitrification efficiency = (inlet NOx concentration - outlet NOx concentration) / inlet NOx concentration × 100%), and the data comes from the online NOx analyzers at the inlet and outlet.

[0070] In this embodiment, associated nodes are nodes in the graph structure that are linked to the core nodes. They correspond to various features such as flue gas composition, reaction temperature, and catalyst microstructure, and are fundamental data nodes that influence or are affected by the core nodes (activity). That is, features of each dimension are extracted from the standardized feature set, with each feature corresponding to one associated node. The node ID is mapped one-to-one with the feature name (e.g., ID1 corresponds to...). Concentration, ID2 corresponding Concentration), stored in the node feature dictionary.

[0071] In this embodiment, in-degree nodes are nodes among the associated nodes that directly affect the core node (active state), meaning that changes in the characteristics of this type of node will lead to changes in the characteristics of the core node (e.g., ...). Increased concentration → decreased activity) is the causal node for activity decay. In-degree nodes are screened using the Pearson correlation coefficient. The correlation coefficient between the characteristics of each associated node and the core node is calculated, and nodes with an absolute value > 0.5 are determined to be in-degree nodes (e.g., ...). The correlation coefficient with activity was -0.72, indicating it was an in-degree node.

[0072] In this embodiment, the exit node is a node among the associated nodes that is affected by the core node (active state). That is, changes in the characteristics of the core node will lead to changes in the characteristics of this type of node (such as decreased activity → catalyst surface). A decrease in adsorption capacity is a "result node" of activity decay. Out-of-degree nodes are screened using causal relationship tests, such as Granger causality tests. If the change in the core node's characteristics is a Granger cause of the change in the related node's characteristics (p-value < 0.05), then it is determined to be an out-of-degree node. For example, if the change in activity is... The Granger criterion for the change in adsorption amount, p=0.02, indicates an out-degree node.

[0073] In this embodiment, the close association distance is a quantified parameter assigned to the connection (edge) between the associated node and the core node, representing the degree of direct correlation between the corresponding feature and activity decay. The smaller the distance, the stronger the correlation (the greater the weight). Based on historical activity decay data, a random forest model is used to calculate the feature importance of each feature to activity decay. After normalizing the importance (0-1), 1 - the normalized importance is taken as the close association distance, such as... The feature importance is 0.8, and the close association distance is 1 - 0.8 = 0.2.

[0074] Step 2.2: Node interaction rule design: Combine the differences between the features of in-degree nodes and core nodes, and the differences between the features of out-degree nodes and core nodes, and adjust the weight of the feature differences based on the closeness of association. At the same time, assign a feature collaboration strength coefficient to the connection between in-degree nodes, and assign a feature action rate coefficient to the connection between out-degree nodes and core nodes.

[0075] In this embodiment, the feature difference is the difference between the features of the in-degree node and the core node, and the difference between the features of the out-degree node and the core node, reflecting the degree of feature deviation between nodes. The feature difference is the absolute value of the difference between the feature value of the core node and the feature value of the associated node.

[0076] In this embodiment, the feature synergy strength coefficient is a quantified parameter assigned to the connection (edge) between in-degree nodes, representing the degree of synergy between the features of different in-degree nodes on activity decay. Based on the correlation between feature combinations and activity decay rates in historical data, a gradient boosting tree model is used to fit the synergy strength coefficient. When the feature combination of two in-degree nodes leads to an increase in the activity decay rate of more than 10%, the synergy coefficient is ≥0.7; when the increase is 5%-10%, the coefficient is 0.4-0.6; and when the increase is <5%, the coefficient is <0.4.

[0077] In this embodiment, the feature action rate coefficient is a quantization parameter assigned to the connection (edge) between the out-degree node and the core node. It represents the action rate of the core node (activity) on the out-degree node (feature). The larger the rate coefficient, the faster the core node affects the out-degree node (the higher the transfer weight). The result of 1 - normalized time is taken as the action rate coefficient.

[0078] In this embodiment, node interaction rules combine feature differences, close association distance, collaboration strength coefficient, and action rate coefficient to adjust the weights of feature interactions, achieving preliminary feature fusion of in-degree nodes, out-degree nodes, and core nodes. For example, the weight calculation for in-degree nodes is as follows: 1) Node affinity distance 0.2 → base weight = 1 - 0.2 = 0.8; 2) Calculation of out-degree node weight: (This is based on the synergy coefficient with the HCl node, which is 0.9, and the adjusted weight = 0.8 × 0.9 = 0.72) The rate coefficient of the adsorption node is 0.8 → basic transfer weight = 0.8; combined with the close correlation distance of 0.3 → secondary calibration weight = 0.8 × (1 - 0.3) = 0.56; feature fusion: the weighted sum of the in-degree node correction weight (0.72), the out-degree node calibration weight (0.56), and the core node feature (0.85) is used to obtain the preliminary fusion feature value.

[0079] Step 2.3: GNN network architecture design: including: input layer, feature interaction layer, graph convolutional layer, fully connected layer and output layer;

[0080] Step 2.4: Model Training and Parameter Optimization: Based on historical activity decay data, the model is divided into training, validation, and test sets. The training set is used to learn the mapping relationship between graph structure nodes and activity states. The validation set is used to adjust the number of graph convolutional layers and feature interaction weights. The test set is used to evaluate the model's generalization performance.

[0081] In this embodiment, historical activity decay data refers to the multimodal data and corresponding catalyst activity data set of the ultra-low temperature denitrification system over a period of time, such as within one year. This data is used for model training, validation, and testing, and must cover different operating conditions (such as different flue gas compositions and temperature fluctuations). The historical activity decay data is divided into three subsets proportionally for model parameter learning, hyperparameter adjustment, and generalization performance evaluation, respectively, to ensure model reliability. A 7:2:1 ratio is used for the division.

[0082] The model is trained using gradient descent, hyperparameters are adjusted using the validation set, and generalization performance is evaluated using the test set until the model converges (the loss function value stabilizes). Specifically: Loss function: Mean squared error (MSE) loss function is used, calculating the squared difference between the predicted and actual activity values; Optimizer: Adam optimizer is used, with an initial learning rate of 0.001, which decays to 0.8 times every 5 epochs (training rounds); Hyperparameter adjustment: The number of convolutional layers in the graph is tested on the validation set (3, 4, 5 layers). It is found that 4 layers result in the lowest MSE (0.18), thus determining 4 convolutional layers as the optimal number; The distribution ratio of feature interaction weights is adjusted (40% in-degree, 30% out-degree, 30% core weights result in the lowest MSE on the validation set); Generalization performance evaluation: The test set MSE is 0.22, and the coefficient of determination is 0.92. The closer the coefficient of determination is to 1, the better the generalization performance and the more convergent the model.

[0083] Step 3: Input the preprocessed data into the trained catalyst activity decay model to obtain the current activity state of the catalyst and the predicted value of the activity change trend in the future preset period.

[0084] In this embodiment, the preprocessed data refers to the multimodal data after outlier removal, missing value filling, and standardization.

[0085] The core output of the model is the predicted value of the current activity state and the trend of activity change within a preset future period. This includes the catalyst activity value at the current moment (such as denitrification efficiency) and the activity decay curve (daily activity value) within a preset future period (such as 10 days or 30 days).

[0086] Preferably, the flue gas composition data includes the concentration of harmful components that inhibit catalyst activity and the concentration of essential components that support the denitrification reaction; the reaction temperature data includes the real-time temperature of different regions of the catalyst bed and the degree of temperature fluctuation per unit time; and the catalyst microstructure data includes structural parameters reflecting the total amount of active sites and distribution parameters reflecting the effective proportion of active sites.

[0087] The beneficial effects of the above technical solution are as follows: by collecting multimodal data on flue gas, temperature, and microstructure, it comprehensively covers the key factors affecting catalyst activity, avoids feature loss in single-modal data, and the activity decay model built based on GNN accurately captures the dynamic diffusion law of activity decay through graph structure mapping and hierarchical convolution. Compared with traditional machine learning models, the prediction accuracy is improved by 15%-20%. The model training process adopts hierarchical sampling and cross-validation, which has excellent generalization performance and can adapt to different ultra-low temperature denitrification conditions, providing accurate data support for catalyst replacement and maintenance.

[0088] This invention provides a method for predicting the activity of ultra-low temperature denitrification catalysts based on multimodal data. The GNN network architecture design includes:

[0089] The input layer receives graph structure data and passes it to the feature interaction layer;

[0090] The feature interaction layer, based on node interaction rules, combines the close association distance, feature collaboration strength coefficient, and feature action rate coefficient to complete the initial feature fusion of in-degree nodes, out-degree nodes, and core nodes.

[0091] The graph convolutional layer is set to 4 layers. Each layer prioritizes aggregating features in the in-degree nodes that are strongly correlated with activity decay, and through the interaction of feature differences between neighboring nodes, it explores the hidden activity decay diffusion patterns among multimodal data.

[0092] The fully connected layer converts the feature vector output by the graph convolutional layer into the dimension corresponding to the active state;

[0093] The output layer outputs the predicted value of catalyst activity decay.

[0094] The core function of the input layer is to read the graph structure data, i.e., the node feature matrix and the adjacency matrix, convert it into a PyTorch tensor format adapted to the GNN model, and then pass it in batches to the feature interaction layer to ensure that the data format is consistent with the model requirements. For example, the graph structure data contains 10 nodes (1 core + 9 associated nodes), the node feature matrix is ​​a 10×12 NumPy array (e.g., [[0.85,0,...], [0.47,0,...], etc.], the first feature of the core node is the activity value, and the corresponding features of the associated nodes are the modal data), and the adjacency matrix is ​​a 10×10 NumPy array (e.g., the core node and the associated node are the activity value, the associated nodes are the modal data, etc.). The edge weight of the node is 0.2, and the edge weight with the outlet temperature node is 0.6; the input layer converts it into a PyTorch tensor (data type float32, device is cuda:0), with a batch size of 32 (i.e., 32 graph structure data are transmitted each time), and then passes it to the feature interaction layer. Implementation method: Create graph data objects based on the PyG Data class, where x=torch.tensor(node ​​feature matrix, dtype=torch.float32, device='cuda'), edge_index=torch.tensor(edge ​​index of adjacency matrix, dtype=torch.long, device='cuda'), and its edge index format is [2,E], where E is the number of edges, such as the core node ID0 and... The edge of node ID1 has indices of [[0,1],[1,0]]; data is loaded using PyG's DataLoader class, with batch_size=32, shuffle=True (data is shuffled during training), and num_workers=4 (multi-threaded loading), and the data iterator is passed to the feature interaction layer.

[0095] In this embodiment, the core function of the feature interaction layer is to incorporate the close association distance (the degree of association between a node and the core), the feature collaboration strength coefficient (the collaborative effect between in-degree nodes), and the feature action rate coefficient (the rate at which the core affects out-degree nodes) into feature calculation according to the designed node interaction rules, thereby eliminating node feature redundancy and obtaining a preliminary fused feature vector. For example, in-degree node fusion: The close association distance of node (ID1) is 0.2 → basic weight 0.8, and the synergy coefficient with HCl node (ID4) is 0.9 → adjusted weight 0.8 × 0.9 = 0.72; The node (ID2) has a close association distance of 0.3 → a base weight of 0.7, and... The collaboration coefficient of node (ID3) is 0.8 → adjusted weight 0.7 × 0.8 = 0.56; the in-degree feature fusion value = 0.72 × 0.47. Features) + 0.56 × 0.5 Features) = 0.34 + 0.28 = 0.62; Out-degree node fusion: The rate coefficient of adsorption quantity node (ID5) is 0.8 → basic transfer weight 0.8, and the close association distance is 0.3 → calibration weight 0.8×0.7=0.56; the out-degree feature fusion value = 0.56×0.6 (adsorption quantity feature) = 0.34; 3) Overall fusion: the preliminary fusion feature vector = in-degree fusion value (0.62)×0.4 + out-degree fusion value (0.34)×0.3 + core node feature (0.85)×0.3 = 0.25 + 0.10 + 0.26 = 0.61 (weight allocation is based on the proportion of node type), and finally outputs a 1×12 preliminary fusion feature vector (each feature dimension is calculated according to this rule).

[0096] The core function of the fully connected layer is to compress the high-dimensional feature vector (e.g., 128-dimensional) output by the graph convolutional layer to a low-dimensional dimension (e.g., 2-dimensional: current activity value, daily decay rate) related to the catalyst's active state through linear transformation, thereby achieving the mapping from high-dimensional features to target parameters. For example, the fourth layer of the convolutional layer outputs a 128-dimensional feature vector (e.g., [0.12, 0.35, 0.21, ..., 0.48]). The fully connected layer contains two linear layers: 1) The first linear layer (in_features=128, out_features=64): the 128-dimensional vector is converted into a 64-dimensional vector (e.g., [0.25, 0.41, ..., 0.33]) through the weight matrix (128×64), and then output after Dropout (0.3) and Tanh activation; 2) The second linear layer (in_features=64, out_features=2): the 64-dimensional vector is converted into a 2-dimensional vector, e.g., [0.85, 0.005], corresponding to the normalized current activity value and daily decay rate.

[0097] The output layer receives a 2D vector from the fully connected layer, maps it to the interval [0 to 1] using the Sigmoid activation function, and then converts it into the actual activity decay prediction value through normalization and inversion, providing users with intuitive prediction results. For example, the fully connected layer outputs a 2D vector [0.85, 0.005]; Sigmoid activation: since the fully connected layer output is already in [-1, 1] (Tanh activation), the Sigmoid value after activation is [0.80, 0.50]; Normalization inversion: the normalization range of the current activity is [0, 1] (corresponding to the actual 0%-100%), so the actual current activity = 0.85 × 100% = 85%; the normalization range of the daily decay rate is [0, 1], which corresponds to the actual 0%-0.01% / day, so the actual daily decay rate = 0.005 × 0.01% / day; output result: the current catalyst activity is 85%, the future daily decay is 0.00005%, and the expected activity after 30 days is 85% - 30 × 0.00005% = 84.9985%.

[0098] The beneficial effects of the above technical solution are as follows: by refining the functions of each layer of the GNN network architecture, the input layer ensures data format adaptation, the feature interaction layer eliminates redundant features, the four-layer graph convolutional layer mines the decay law in a single-factor → multi-factor logical manner, the fully connected layer achieves accurate dimensionality reduction of high-dimensional features, and the output layer provides intuitive prediction results. Compared with the traditional GNN architecture, the feature mining depth of this architecture is increased by 40%-50%, the dimensionality matching degree of the prediction results is higher, and the MSE of the test set is reduced by 0.05-0.08, further improving the accuracy and reliability of liveness prediction.

[0099] This invention provides a method for predicting the activity of an ultra-low temperature denitrification catalyst based on multimodal data, further comprising: correcting the predicted activity value through an error correction mechanism, specifically including:

[0100] Construct a residual model adapted to the characteristics of activity decay bias, and learn the correlation between multimodal feature combination and prediction bias;

[0101] The original predicted value is adjusted by using the deviation compensation value output by the residual model and combining it with the characteristic difference direction of the current working condition;

[0102] A 2% prediction error threshold is set, and the correction results are monitored in real time. When the error exceeds the limit continuously, the microstructure data acquisition cycle is shortened and the error is controlled to be stable below the prediction error threshold. The final output includes the current activity value, decay rate, and remaining lifetime. When the activity drops to the preset activity threshold, a lifetime warning is triggered.

[0103] In this embodiment, the residual model is an auxiliary model used to learn the correlation between multimodal feature combinations and prediction bias. The bias feature refers to the difference between the model's original predicted value and the actual prediction value. The input of the residual model is the multimodal feature combination, and the output is the bias compensation value, used to correct the original predicted value. Specifically, the residual model architecture uses a 3-layer MLP (input 12 → hidden layer 32 → hidden layer 16 → output 1), with ReLU as the activation function and MSE (squared difference between prediction bias and actual bias) as the loss function.

[0104] Training data construction: Extract 50,000 sets of multimodal features, original predicted values ​​and actual active values ​​from historical data, calculate residual = actual active value - original predicted value, and construct feature-residual training set (7:2:1 split training / validation / test set).

[0105] Model training: Use the Adam optimizer (learning rate 0.0005), train for 50 epochs, stop training when the validation set MSE ≤ 0.01, and save the model weights (residual_model.pth).

[0106] Model Deployment: Deployed in parallel with the GNN model, receiving multimodal features in real time and outputting bias compensation values.

[0107] The specific steps for learning the correlation between multimodal feature combinations and prediction bias are as follows: Permutation Importance is used to analyze the correlation strength between features and bias; the MSE change of the residual model after each feature is randomly shuffled is calculated; the greater the change, the stronger the correlation. A Partial Dependence Plot (PDP) is then drawn to visually represent a single feature (such as...) The relationship between concentration and deviation (e.g.) When the concentration is >150 ppm, the PDP curve shows an upward trend, indicating that the deviation increases with... Increased concentration leads to increased intensity; the residual model is updated regularly (every 30 days) with new historical data to ensure the timeliness of the correlation. Incremental training is used during the update (the weights of the first two layers are frozen, and only the output layer is trained) to reduce training time.

[0108] In this embodiment, the 2% prediction error threshold is interpreted as an industrial-grade error standard for judging whether the corrected prediction value is reliable. It is set based on the actual needs of the ultra-low temperature denitrification scenario. When the absolute error between the corrected value and the actual activity value exceeds 2%, it is judged as a prediction anomaly.

[0109] Determination of shortened microstructure data acquisition cycle when errors continuously exceed limits:

[0110] Continuous over-limit determination: The program counts the number of consecutive over-limit errors and sets a continuous over-limit threshold (e.g., 3 times). Once the threshold is reached, a periodic adjustment signal is triggered.

[0111] Collection cycle adjustment: Establish a table corresponding to the activity decay stage - number of consecutive exceedances - shortened cycle, as shown in Table 1. Based on the current activity decay stage (initial stage: 80%-100%, middle stage: 60%-80%, final stage: <60%) and the number of exceedances, automatically look up the corresponding shortened cycle in the table.

[0112] High-frequency acquisition and execution: Through a timing control module (such as a PLC timer), the microstructure sampling device (such as an automatic sampler) is triggered according to the shortened cycle. After sampling, the sample is automatically sent to the relevant instrument for testing, and the test data is transmitted to the preprocessing module within 1 hour.

[0113] Table 1 Correspondence between the activity decay stage and the shortened acquisition cycle

[0114]

[0115] In this embodiment, the remaining lifetime = (current activity value - preset activity threshold) / daily decay rate. If the current activity is less than or equal to the threshold, the system will output that the device needs to be replaced immediately.

[0116] In this embodiment, the warning triggering logic is to compare the current activity value with a preset threshold in real time. If the current activity is less than or equal to the threshold, a warning is triggered immediately. If the current activity is greater than the threshold but the predicted activity is less than or equal to the threshold within the next 10 days, a warning is triggered 5 days in advance.

[0117] The beneficial effects of the above technical solution are as follows: by accurately learning the correlation between multimodal features and prediction bias through residual model, the prediction error is reduced by 40%-50%; the 2% error threshold monitoring ensures that the prediction results meet the industrial practical standards and avoids invalid data from misleading decisions; when the error exceeds the limit continuously, the microstructure acquisition cycle is shortened, and the error control response speed is improved by 40%-50%, while the targeted acquisition of features reduces the testing cost; the final output of the complete results of current activity-decay rate-remaining lifetime and lifetime warning provides a precise time node for catalyst replacement. Compared with the prediction scheme without error correction, the remaining lifetime prediction accuracy is improved by 35%-45%, effectively avoiding the problems of excessive or untimely catalyst replacement.

[0118] This invention provides a method for predicting the activity of ultra-low temperature denitration catalysts based on multimodal data, which shortens the microstructure data acquisition cycle and controls the error to remain stable below the prediction error threshold, including:

[0119] The number of consecutive exceedances is obtained, and combined with the current catalyst activity decay stage, the characteristic type corresponding to the current error source is identified and the shortening period is determined.

[0120] For the microstructure data collected after shortening the cycle, features related to the current error source are extracted in a targeted manner to form a high-frequency microstructure feature subset. The correlation distance between graph structure nodes and core nodes, as well as the feature coordination strength coefficient between in-degree nodes, are adjusted to achieve stable control.

[0121] In this embodiment, the number of consecutive exceedances is the number of times the prediction error exceeds 2% consecutively (e.g., 3 times, 5 times), reflecting the severity of the error; the activity decay stage is divided into the initial stage (80%-100%, slow decay), the middle stage (60%-80%, fast decay), and the final stage (<60%, extremely fast decay) according to the current activity. Both stages jointly determine the priority of the error source and the shortened collection cycle.

[0122] In this embodiment, targeted extraction involves collecting only microstructural features related to the type of error source. For example, if the error source is dispersion, only dispersion is tested, skipping irrelevant features such as crystal structure to avoid invalid testing. The high-frequency microstructural feature subset is a set of features extracted in a targeted manner after shortening the collection cycle, such as specific surface area and dispersion collected once every 3 days, used to quickly update the model input and improve the sensitivity to error sources.

[0123] In this embodiment, the adjusted distance = original distance × (1 - feature change amplitude / 100%), and the adjusted coefficient = original coefficient × 1.2.

[0124] The beneficial effects of the above technical solution are as follows: by combining the number of consecutive exceedances with the activity decay stage, the error source feature type can be accurately located, avoiding blindly shortening the acquisition cycle and reducing unnecessary testing costs; the high-frequency microstructure feature subset can be extracted in a targeted manner, reducing the testing time by 50%-60% and improving the data transmission efficiency by 40%-50%; by adjusting the correlation close distance and the coordination strength coefficient, the sensitivity of the model to the error source features can be increased by 35%-45%, and the error control stability can be improved by 50%. Compared with the fixed cycle and fixed parameters, it can more quickly and cost-effectively control the prediction error within 2%, adapting to the dynamic changes of different decay stages of the catalyst.

[0125] This invention provides a method for predicting the activity of ultra-low temperature denitrification catalysts based on multimodal data, assigning characteristic cooperative strength coefficients to the connections between in-degree nodes and characteristic action rate coefficients to the connections between out-degree nodes and core nodes, including:

[0126] Based on the close correlation distance, the difference between the features of the in-degree node and the features of the core node is assigned a basic fusion weight according to the rule that the smaller the distance, the greater the weight.

[0127] The feature collaboration strength coefficient between the corresponding in-degree node and other in-degree nodes is superimposed to correct the basic fusion weights;

[0128] Based on each out-degree node, the weight of the feature transferred from the core node to the out-degree node is set according to the rule that the larger the rate coefficient, the higher the weight of the transfer. At the same time, the weight of the transfer is calibrated again by taking into account the close relationship between the out-degree node and the core node.

[0129] Real-time monitoring of the fluctuation amplitude of multimodal features under the current working conditions. If the fluctuation amplitude of the corresponding in-degree node feature exceeds the preset threshold, the close association distance weight of the corresponding node will be temporarily increased. If the deviation between the change rate of the corresponding out-degree node feature and the activity decay rate of the core node exceeds the preset range, the feature action rate coefficient of the corresponding node will be temporarily adjusted.

[0130] The processed in-degree node fusion features and processed out-degree node fusion features are weighted and aggregated with the initial activity features of the core nodes to obtain the output result after preliminary fusion of multimodal features.

[0131] In this embodiment, the basic fusion weight is the initial weight for fusing the features of in-degree nodes and the features of core nodes. It follows the rule that the smaller the distance between closely related features, the greater the weight, ensuring that features with strong correlations account for a higher proportion in the fusion. Basic weight = 1 - distance between closely related features.

[0132] In this embodiment, the feature collaboration strength coefficient reflects the collaborative effect between in-degree nodes, such as When coexisting with HCl, the inhibitory effect on activity is stronger than when it exists alone, and the synergistic coefficient is high. Adding this coefficient can correct the basic weight, making the fusion weight more closely reflect the actual synergistic effect and avoiding underestimating / overestimating the impact of the synergistic effect on activity. Corrected weight = basic weight × synergistic coefficient.

[0133] In this embodiment, the weight transfer is the transfer of features from the core node (active) to the out-degree node (e.g., ...). The weight of the adsorption amount is determined by the rule that the larger the characteristic action rate coefficient, the higher the weight of the transfer, ensuring that the out-degree nodes with fast response account for a higher proportion in the fusion. The base value of the transfer weight = action rate coefficient, where action rate coefficient = 1 - (stabilization time / maximum stabilization time).

[0134] In this embodiment, the secondary calibration adjusts the base value of the transfer weight using the close proximity between the out-degree node and the core node (reflecting the degree of correlation between degree and activity). The rule is that the smaller the close proximity, the higher the weight after calibration, ensuring that the transfer weight takes into account both response speed and correlation strength. Secondary calibration weight = base value × (1 - close proximity).

[0135] In this embodiment, the fluctuation amplitude is the change in the in-degree node characteristics within a unit of time (e.g., 1 hour). The concentration increased from 120 ppm to 140 ppm, with a fluctuation range of 20 ppm. The preset threshold is the maximum allowable fluctuation range for this feature. (±15ppm / h, temperature ±3℃ / h) Monitoring fluctuations can promptly detect abnormal operating conditions and avoid prediction bias caused by the use of outdated weights / coefficients in the model.

[0136] In this embodiment, when the fluctuation of the in-degree node's feature exceeds the limit, the weight of the close association distance between that feature and the core node is temporarily increased (e.g., from 0.72 to 0.85), making the model pay more attention to the fluctuating feature; when the deviation between the out-degree node and the core node exceeds the limit, the feature action rate coefficient is temporarily adjusted (e.g., from 0.6 to 0.75) to make the propagation weights fit the current deviation and avoid expanding the prediction deviation. For example, Concentration fluctuation range of 25ppm (exceeding the threshold of 15ppm), temporarily... The weight of the close association with the core node is increased from 0.72 (corrected weight) to 0.72 × 1.2 = 0.86; The generation rate and activity decay rate deviated by 16% (10% beyond the range), so temporarily... The characteristic action rate coefficient of the generation rate was increased from 0.6 to 0.6 × 1.25 = 0.75; after adjustment, the model... Increased sensitivity to fluctuations The propagation weights of the generation rate are more accurate, and the prediction error is reduced from 2.2% to 1.9% (<2%).

[0137] In this embodiment, the processed in-degree node fusion features (weighted summation after correction), the out-degree node fusion features (weighted summation after calibration), and the initial activity features of the core node are weighted and summed according to a preset ratio (e.g., 40% in-degree, 30% out-degree, and 30% core) to obtain the preliminary fused feature vector.

[0138] The beneficial effects of the above technical solution are: by correcting the fusion weights of in-degree nodes through basic weights and synergy coefficients, the accuracy of representing the synergistic effect between in-degree nodes is improved by 35%-45%, avoiding underestimation. The impact of strong synergistic features such as HCl is considered; by optimizing the weight transfer of out-degree nodes through weight transfer and secondary calibration, the response speed and correlation strength are balanced, and the feature fusion accuracy of out-degree nodes is improved by 30%-40%; the real-time monitoring and temporary adjustment mechanism improves the model's adaptability to operating condition fluctuations by 50%, avoiding the expansion of prediction bias caused by fluctuations; the preliminary fusion features obtained by the final weighted aggregation have a feature signal-to-noise ratio improved by 60%-70% compared with the fusion scheme without parameter optimization, providing more accurate basic data for subsequent graph convolutional layers to mine attenuation patterns, and indirectly improving the overall prediction accuracy of the model by 15%-20%.

[0139] This invention provides a method for predicting the activity of ultra-low temperature denitrification catalysts based on multimodal data. Figure 1 shows the in-degree node features related to catalyst poisoning in the first layer of the convolutional layer, the in-degree node features related to structural aging in the second layer of the convolutional layer, the in-degree node features related to reaction mass transfer in the third layer of the convolutional layer, and the synergistic combination of the features of the first, second, and third layers of the convolutional layer in the fourth layer of the convolutional layer.

[0140] The first layer of the convolutional layer explores the inhibition of activity by the adsorption and diffusion of harmful components on the catalyst surface. The second layer of the convolutional layer explores the erosion of active sites by the spatial diffusion of microstructural degradation. The third layer of the convolutional layer explores the constraint of the reaction temperature field distribution diffusion on the catalytic reaction rate. The fourth layer of the convolutional layer integrates the rules of the first, second, and third layers to explore the global diffusion law of activity decay under the coupling of multiple factors.

[0141] In this embodiment, the four-layer graph convolutional layer aggregates features hierarchically from single-factor to multi-factor. Layers 1-3 focus on three single factors: catalyst poisoning, structural aging, and reaction mass transfer, respectively. Layer 4 integrates these three factors, exploring their diffusion patterns (such as adsorption diffusion, spatial diffusion, and temperature field diffusion) and coupling patterns. For example, the first layer (poisoning-related) preferentially aggregates features from in-degree nodes. (ID1) and HCl (ID4) features, input dimensions 12→24, after aggregation by GCNConv, the poisoning feature dimensions in the output feature vector, such as SO2 adsorption amount and HCl reaction amount, account for up to 60% of the weight, and patterns are discovered: For every 50 ppm increase in concentration, the amount of sulfate generated on the catalyst surface increases by 20%, the blockage rate of active sites increases by 15%, and the activity decreases by 4%. The second convolutional layer (related to structural aging) preferentially aggregates specific surface area (ID6) and dispersion (ID7) features in the in-degree nodes. The input dimension increases from 24 to 48, and the weight of the "structural feature dimension" in the output feature vector reaches 55%. The pattern is discovered: for every 50 ppm increase in specific surface area... The number of active sites decreased by 12%, the micropore blockage rate increased by 8%, and the activity decreased by 6%; the third layer of graph convolution (reaction-mass transfer related): the mid-temperature (ID3) and flue gas velocity (supplementary feature, ID8) features in the preferential polymerization in-degree node, the input dimension 48→64, and the weight of the mass transfer feature dimension in the output feature vector reached 50%, the discovery rule: for every 10℃ decrease in bed temperature, the denitrification reaction rate constant decreased by 15%, the residence time of reactants on the catalyst surface increased by 10%, but the activity still decreased by 5% (due to the decrease in reaction rate as the main factor); the fourth layer of graph convolution (multi-factor coupling): the output features of the first 3 layers of polymerization (dimension 64→128), the discovery rule: 150ppm (poisoning) + specific surface area 75 Under the coupled conditions of (aging) + temperature 80℃ (mass transfer), the activity decay rate is 1.2% / day, which is a single factor ( At 150 ppm, 0.5% / day, specific surface area 75 The percentage of precipitation is 1.5 times the sum of 0.4% / day at 80℃ and 0.3% / day at 80℃, demonstrating a coupling amplification effect. Specifically, each graph convolutional layer uses PyG's GCNConv module, setting different input and output dimensions (12→24→48→64→128); priority aggregation is achieved through feature masks: a specific feature mask matrix is ​​defined for each layer (e.g., in the mask matrix of the first layer, ...). (The dimension corresponding to HCl is 1, and the other dimensions are 0.5). The mask matrix is ​​multiplied with the output features of GCNConv to strengthen the weight of the target features. Pattern mining verification: The contribution of the output features of each layer to the activity decay is analyzed by SHAP (SHapley Additive ex Planations) values ​​to ensure that the first layer... 1) The SHAP value of HCl accounts for ≥50%, the SHAP value of specific surface area and dispersion in the second layer accounts for ≥45%, the SHAP value of temperature and flow rate in the third layer accounts for ≥40%, and the sum of the SHAP values ​​of the first three features in the fourth layer accounts for ≥70%; 4) BatchNorm2d (momentum 0.9, eps=1e-5) and Dropout (probability 0.3) are added after each layer to prevent overfitting, and LeakyReLU (negative slope 0.2) is used as the activation function.

[0142] The beneficial effects of the above technical solution are as follows: By using four layers of graph convolutional layers to aggregate similar features, interference from different decay mechanisms is avoided, improving the targeting of single-mechanism pattern mining by 50%-60%; the first to third layers accurately mine single patterns of adsorption diffusion, spatial diffusion, and temperature field diffusion, providing a foundation for global patterns; the fourth layer mines global patterns of multi-mechanism coupling, effectively capturing coupling amplification effects, and improving the prediction accuracy of multi-factor coupled conditions by 40%-50% compared to GNN models without layered aggregation; at the same time, the quantification formulas of the patterns in each layer are deeply integrated with the model, enabling the model not only to output predicted values ​​but also to explain why decay occurs, improving the interpretability of the model and providing industrial users with more comprehensive decision-making basis.

[0143] Secondly, a system for predicting the activity of ultra-low temperature denitrification catalysts based on multimodal data, such as... Figure 2 As shown, it includes:

[0144] The preprocessing module is used to collect multimodal raw data under ultra-low temperature denitrification scenario and preprocess the multimodal raw data to obtain a standardized feature set. The multimodal raw data includes flue gas composition data, reaction temperature data and catalyst microstructure data.

[0145] A model building module is used to build a catalyst activity decay model based on a graph neural network, wherein the model building module includes:

[0146] The graph structure data mapping unit is used to map the standardized feature set into graph structure data that conforms to the dynamic diffusion law of activity decay. The real-time activity state of the catalyst is used as the core node, and the features corresponding to the flue gas composition, reaction temperature, and catalyst microstructure are respectively used as the associated nodes in the graph. Feature nodes that have a direct effect on the activity state of the core node are defined as in-degree nodes, and feature nodes affected by the activity state of the core node are defined as out-degree nodes. Each associated node and the core node is assigned a close association distance, and the close association distance is determined by the degree of direct correlation between the corresponding feature and the activity decay.

[0147] The node interaction rule design unit is used to combine the differences between the features of in-degree nodes and core nodes, and the differences between the features of out-degree nodes and core nodes, and adjust the weight of the feature differences based on the closeness of association. At the same time, it assigns a feature collaboration strength coefficient to the connection between in-degree nodes and a feature action rate coefficient to the connection between out-degree nodes and core nodes.

[0148] The GNN network architecture design unit includes: input layer, feature interaction layer, graph convolutional layer, fully connected layer and output layer;

[0149] The model training and parameter optimization unit is used to divide the training set, validation set, and test set based on historical activity decay data. The training set is used to learn the mapping relationship between graph structure nodes and activity states, the validation set is used to adjust the number of graph convolutional layers and feature interaction weights, and the test set is used to evaluate the model's generalization performance.

[0150] The activity prediction module is used to input the preprocessed data into the trained catalyst activity decay model to obtain the current activity state of the catalyst and the predicted value of the activity change trend within a preset period of time.

[0151] The beneficial effects of the above technical solution are as follows: by collecting multimodal data on flue gas, temperature, and microstructure, it comprehensively covers the key factors affecting catalyst activity, avoids feature loss in single-modal data, and the activity decay model built based on GNN accurately captures the dynamic diffusion law of activity decay through graph structure mapping and hierarchical convolution. Compared with traditional machine learning models, the prediction accuracy is improved by 15%-20%. The model training process adopts hierarchical sampling and cross-validation, which has excellent generalization performance and can adapt to different ultra-low temperature denitrification conditions, providing accurate data support for catalyst replacement and maintenance.

[0152] It is understood that the system provided in this embodiment has functions or includes modules that can be used to execute the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0153] The present invention also provides an electronic device, comprising: a processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein when the processor executes the computer instructions, the electronic device performs a method as described in any of the above possible implementations.

[0154] The present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor of an electronic device, cause the processor to perform a method as described in any of the above possible implementations.

[0155] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0156] Those skilled in the art will readily understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. Those skilled in the art will also readily understand that the various embodiments of the present invention have different focuses, and for the sake of convenience and brevity, the same or similar parts may not be repeated in different embodiments. Therefore, parts not described or not described in detail in one embodiment can be referred to in other embodiments.

[0157] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0158] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0159] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0160] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., digital versatile discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0161] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for predicting the activity of ultra-low temperature denitrification catalysts based on multimodal data, characterized in that, The method includes: Step 1: Collect multimodal raw data under ultra-low temperature denitrification scenario, and preprocess the multimodal raw data to obtain a standardized feature set. The multimodal raw data includes flue gas composition data, reaction temperature data, and catalyst microstructure data. Step 2: Construct a catalyst activity decay model based on a graph neural network. The model construction process includes: Step 2.1: Graph Structure Data Mapping: The standardized feature set is mapped to graph structure data that conforms to the dynamic diffusion law of activity decay. The real-time activity state of the catalyst is used as the core node, and the features corresponding to the flue gas composition, reaction temperature, and catalyst microstructure are respectively used as associated nodes in the graph. Feature nodes that have a direct effect on the activity state of the core node are defined as in-degree nodes, and feature nodes affected by the activity state of the core node are defined as out-degree nodes. A close association distance is assigned to the connection between each associated node and the core node, and the close association distance is determined by the degree of direct correlation between the corresponding feature and the activity decay. Step 2.2: Node interaction rule design: Combine the differences between the features of in-degree nodes and core nodes, and the differences between the features of out-degree nodes and core nodes, and adjust the weight of the feature differences based on the closeness of association. At the same time, assign a feature collaboration strength coefficient to the connection between in-degree nodes, and assign a feature action rate coefficient to the connection between out-degree nodes and core nodes. Step 2.3: GNN network architecture design: including: input layer, feature interaction layer, graph convolutional layer, fully connected layer and output layer; Step 2.4: Model Training and Parameter Optimization: Based on historical activity decay data, the model is divided into training, validation, and test sets. The training set is used to learn the mapping relationship between graph structure nodes and activity states. The validation set is used to adjust the number of graph convolutional layers and feature interaction weights. The test set is used to evaluate the model's generalization performance. Step 3: Input the preprocessed data into the trained catalyst activity decay model to obtain the current activity state of the catalyst and the predicted value of the activity change trend in the future preset period.

2. The method for predicting the activity of ultra-low temperature denitrification catalysts based on multi-modal data according to claim 1, characterized in that, The flue gas composition data includes the concentrations of harmful components that inhibit catalyst activity and the concentrations of essential components that support the denitrification reaction; the reaction temperature data includes the real-time temperature of different regions of the catalyst bed and the degree of temperature fluctuation per unit time; the catalyst microstructure data includes structural parameters reflecting the total amount of active sites and distribution parameters reflecting the effective proportion of active sites.

3. The method for predicting the activity of ultra-low temperature denitrification catalysts based on multi-modal data according to claim 1, characterized in that, The GNN network architecture design includes: The input layer receives graph structure data and passes it to the feature interaction layer; The feature interaction layer, based on node interaction rules, combines the close association distance, feature collaboration strength coefficient, and feature action rate coefficient to complete the initial feature fusion of in-degree nodes, out-degree nodes, and core nodes. The graph convolutional layer is set to 4 layers. Each layer prioritizes aggregating features in the in-degree nodes that are strongly correlated with activity decay, and through the interaction of feature differences between neighboring nodes, it explores the hidden activity decay diffusion patterns among multimodal data. The fully connected layer converts the feature vector output by the graph convolutional layer into the dimension corresponding to the active state; The output layer outputs the predicted value of catalyst activity decay.

4. The method for predicting the activity of ultra-low temperature denitrification catalysts based on multi-modal data according to claim 1, characterized in that, Also includes: The predicted activity value is corrected through an error correction mechanism, specifically including: Construct a residual model adapted to the characteristics of activity decay bias, and learn the correlation between multimodal feature combination and prediction bias; The original predicted value is adjusted by using the deviation compensation value output by the residual model and combining it with the characteristic difference direction of the current working condition. A 2% prediction error threshold is set, and the correction results are monitored in real time. When the error exceeds the limit continuously, the microstructure data acquisition cycle is shortened and the error is controlled to be stable below the prediction error threshold. The final output includes the current activity value, decay rate, and remaining lifetime. When the activity drops to the preset activity threshold, a lifetime warning is triggered.

5. The method for predicting the activity of ultra-low temperature denitrification catalysts based on multi-modal data according to claim 4, characterized in that, Shortening the microstructure data acquisition cycle and controlling the error to remain stable below the prediction error threshold includes: The number of consecutive exceedances is obtained, and combined with the current catalyst activity decay stage, the characteristic type corresponding to the current error source is identified and the shortening period is determined. For the microstructure data collected after shortening the cycle, features related to the current error source are extracted in a targeted manner to form a high-frequency microstructure feature subset. The correlation distance between graph structure nodes and core nodes, as well as the feature coordination strength coefficient between in-degree nodes, are adjusted to achieve stable control.

6. The method for predicting the activity of ultra-low temperature denitrification catalysts based on multi-modal data according to claim 1, characterized in that, Assign a feature collaboration strength coefficient to the connections between in-degree nodes, and a feature action rate coefficient to the connections between out-degree nodes and core nodes, including: Based on the close correlation distance, the difference between the features of the in-degree node and the features of the core node is assigned a basic fusion weight according to the rule that the smaller the distance, the greater the weight. The feature collaboration strength coefficient between the corresponding in-degree node and other in-degree nodes is superimposed to correct the basic fusion weights; Based on each out-degree node, the weight of the feature transferred from the core node to the out-degree node is set according to the rule that the larger the rate coefficient, the higher the weight of the transfer. At the same time, the weight of the transfer is calibrated again by taking into account the close relationship between the out-degree node and the core node. Real-time monitoring of the fluctuation amplitude of multimodal features under the current working conditions. If the fluctuation amplitude of the corresponding in-degree node feature exceeds the preset threshold, the close association distance weight of the corresponding node will be temporarily increased. If the deviation between the change rate of the corresponding out-degree node feature and the activity decay rate of the core node exceeds the preset range, the feature action rate coefficient of the corresponding node will be temporarily adjusted. The processed in-degree node fusion features and processed out-degree node fusion features are weighted and aggregated with the initial activity features of the core nodes to obtain the output result after preliminary fusion of multimodal features.

7. The method for predicting the activity of ultra-low temperature denitrification catalysts based on multi-modal data according to claim 3, characterized in that, Figure 1 shows the in-degree node characteristics related to catalyst poisoning in the first layer of the convolutional layer, Figure 2 shows the in-degree node characteristics related to structural aging in the second layer of the convolutional layer, Figure 3 shows the in-degree node characteristics related to reaction mass transfer in the third layer of the convolutional layer, and Figure 4 shows the synergistic combination of the features of the first, second, and third layers of the convolutional layer. The first layer of the convolutional layer explores the inhibition of activity by the adsorption and diffusion of harmful components on the catalyst surface. The second layer of the convolutional layer explores the erosion of active sites by the spatial diffusion of microstructural degradation. The third layer of the convolutional layer explores the constraint of the reaction temperature field distribution diffusion on the catalytic reaction rate. The fourth layer of the convolutional layer integrates the rules of the first, second, and third layers to explore the global diffusion law of activity decay under the coupling of multiple factors.

8. A system for predicting the activity of ultra-low temperature denitrification catalysts based on multimodal data, characterized in that, include: The preprocessing module is used to collect multimodal raw data under ultra-low temperature denitrification scenario and preprocess the multimodal raw data to obtain a standardized feature set. The multimodal raw data includes flue gas composition data, reaction temperature data and catalyst microstructure data. A model building module is used to build a catalyst activity decay model based on a graph neural network, wherein the model building module includes: The graph structure data mapping unit is used to map the standardized feature set into graph structure data that conforms to the dynamic diffusion law of activity decay. The real-time activity state of the catalyst is used as the core node, and the features corresponding to the flue gas composition, reaction temperature, and catalyst microstructure are respectively used as the associated nodes in the graph. Feature nodes that have a direct effect on the activity state of the core node are defined as in-degree nodes, and feature nodes affected by the activity state of the core node are defined as out-degree nodes. Each associated node and the core node is assigned a close association distance, and the close association distance is determined by the degree of direct correlation between the corresponding feature and the activity decay. The node interaction rule design unit is used to combine the differences between the features of in-degree nodes and core nodes, and the differences between the features of out-degree nodes and core nodes, and adjust the weight of the feature differences based on the closeness of association. At the same time, it assigns a feature collaboration strength coefficient to the connection between in-degree nodes and a feature action rate coefficient to the connection between out-degree nodes and core nodes. The GNN network architecture design unit includes: input layer, feature interaction layer, graph convolutional layer, fully connected layer and output layer; The model training and parameter optimization unit is used to divide the training set, validation set, and test set based on historical activity decay data. The training set is used to learn the mapping relationship between graph structure nodes and activity states, the validation set is used to adjust the number of graph convolutional layers and feature interaction weights, and the test set is used to evaluate the model's generalization performance. The activity prediction module is used to input the preprocessed data into the trained catalyst activity decay model to obtain the current activity state of the catalyst and the predicted value of the activity change trend within a preset period of time.

9. An electronic device, characterized in that, include: A processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein when the processor executes the computer instructions, the electronic device performs the method for predicting the activity of an ultra-low temperature denitrification catalyst based on multimodal data as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which includes program instructions that, when executed by a processor of an electronic device, cause the processor to perform the method for predicting the activity of an ultra-low temperature denitrification catalyst based on multimodal data as described in any one of claims 1 to 7.

Images