A Domain-Adaptive Enhancement-Based Method and System for Predicting Sulfur Dioxide Emissions from Circulating Fluidized Beds

Abstract

This invention discloses a method and system for predicting sulfur dioxide emissions from a circulating fluidized bed boiler based on domain adaptive enhancement. The method includes: collecting parameter information of a circulating fluidized bed boiler and constructing a knowledge graph to characterize the boiler's operating characteristics; performing numerical perception embedding modeling on entities related to the parameter information in the knowledge graph according to the real-time operating conditions of the circulating fluidized bed boiler to obtain an embedded boiler operating knowledge graph; inputting the embedded boiler operating knowledge graph into a graph structure feature extraction model for processing and extracting graph structure features that reflect the differences in different operating conditions; collecting unstructured text information from multivariable sensors during the operation of the circulating fluidized bed boiler, and performing feature extraction and fusion to obtain a fused feature vector; constructing and improving a prediction model for predicting sulfur dioxide emissions from the circulating fluidized bed boiler based on the fused feature vector, and predicting sulfur dioxide emission levels based on the improved model.

Description

Technical Field

[0001] This invention belongs to the field of industrial process modeling and pollutant emission prediction technology, specifically involving a method and system for predicting sulfur dioxide emissions from circulating fluidized beds based on domain adaptive enhancement. Background Technology

[0002] Despite increasing attention on renewable energy, coal-fired power generation remains a crucial component of the global energy mix. Among various coal-fired power generation technologies, circulating fluidized bed (CFB) boilers are widely used due to their strong fuel adaptability and efficient sulfur capture capabilities achieved through in-furnace desulfurization. These advantages enable CFB boilers to meet increasingly stringent emission regulations while burning low-quality, heterogeneous fuels. However, due to the complex multiphase reactions, fuel characteristic differences, and dynamic operating conditions within CFB systems, accurate prediction and control of SO2 emissions still face numerous challenges. In particular, the strong coupling characteristics between combustion, desulfurization, and fluidization processes result in highly nonlinear and time-varying emission behavior.

[0003] Sulfur dioxide emissions from coal combustion cause acid rain, air pollution, and health hazards, posing a serious threat to the ecological environment and public health. Therefore, effective control measures for sulfur dioxide emissions are necessary. In practice, such control relies on timely and reliable information on future emission trends. Based on this, accurate prediction of sulfur dioxide emission levels is crucial for guiding desulfurization strategies and emission control decisions. Furthermore, accurate emission forecasts support operators in making proactive operational adjustments, enabling units to address potential risks before emissions exceed limits. Simultaneously, such forecasts can provide quantitative data for optimizing control strategies under dynamic operating conditions, thereby improving operational efficiency while meeting environmental protection requirements.

[0004] In recent years, various methods have been proposed for predicting sulfur dioxide emissions, including mechanistic modeling-based prediction methods and data-driven machine learning models. Furthermore, some studies have combined mechanistic and data-driven models to construct hybrid modeling frameworks for SO2 emission prediction. However, in actual circulating fluidized bed boiler operation, SO2 emission behavior is influenced by multiple factors, including fuel characteristics, air distribution, unit load changes, and operational strategy adjustments. These factors exhibit significant coupling relationships and strong nonlinearity and time dependence, making the emission dynamics extremely complex. Simultaneously, monitoring data collected from industrial sites often suffers from sensor noise, measurement lag, and inconsistent sampling frequencies among different variables, further increasing the difficulty of modeling and prediction. Against this backdrop, existing methods often struggle to simultaneously achieve accurate characterization of complex dynamic processes, effective integration of domain process knowledge, and robust prediction under varying operating conditions, which limits their application effectiveness in real-world industrial scenarios.

[0005] Large Language Models (LLMs) have made significant progress in several fields in recent years due to their capabilities in natural language understanding and generation. However, their application in industrial process prediction tasks is still in the exploratory stage. Limited by a lack of domain knowledge, difficulty in effectively integrating structured sensor data, and insufficient interpretability, existing LLMs still face many challenges in numerical prediction problems for complex industrial systems.

[0006] Knowledge graphs (KGs) demonstrate unique advantages in characterizing complex industrial mechanisms by explicitly representing domain knowledge and entity relationships in a structured form, and have been proven to effectively enhance the reasoning ability and interpretability of artificial intelligence models. However, existing research on the integration of knowledge graphs and large language models (KG-LLM) mainly focuses on natural language processing tasks, paying insufficient attention to the inherent dynamic, numerical, and time-series characteristics of industrial process data, and lacking a systematic modeling framework suitable for industrial prediction scenarios. Summary of the Invention

[0007] To address the shortcomings of existing methods for predicting sulfur dioxide emissions from circulating fluidized bed (CFB) boilers, such as difficulty in effectively integrating domain-specific process knowledge with industrial operating data, insufficient adaptability to complex operating conditions, and significant impact of operating state drift on prediction accuracy and robustness, this invention proposes an improved KG-LLM framework for CFB sulfur dioxide emission prediction. This framework aims to enhance the accuracy and stability of sulfur dioxide emission prediction under different operating conditions. To achieve the above objectives, this invention provides the following solution: A domain-adaptive enhanced method for predicting sulfur dioxide emissions from circulating fluidized beds includes the following steps: Collect parameter information of circulating fluidized bed boilers, organize the parameter information based on preset entity types and relationship types, and construct a knowledge graph to characterize the boiler's operating characteristics; Based on the real-time operating conditions of the circulating fluidized bed boiler, numerical perception embedding modeling is performed on the entities related to the parameter information in the knowledge graph to obtain the embedded boiler operation knowledge graph. The embedded boiler operation knowledge graph is input into the graph structure feature extraction model for processing. By introducing an attention allocation mechanism related to operating conditions, adaptive weights are assigned to different entities and their relationships, and graph structure features that can reflect the differences in different operating conditions are extracted. During the operation of the circulating fluidized bed boiler, unstructured text information from multivariable sensors is collected, and feature extraction and fusion are performed on the unstructured text information and the graph structure features to obtain a fused feature vector for subsequent sulfur dioxide emission prediction. Based on the fused feature vector, a prediction model for predicting sulfur dioxide emissions from circulating fluidized bed boilers is constructed. Uncertainty assessment and concept drift adaptive mechanisms are introduced into the prediction model for improvement, and sulfur dioxide emission levels are predicted based on the improved model.

[0008] Preferably, the parameter information includes: equipment structure information, operating parameter data, chemical reaction mechanism information, operating status information, and time information; The knowledge graph includes: equipment component entities, operating parameter entities, chemical substance entities, operating status entities, and time entities.

[0009] Preferably, the method for numerically-aware embedding modeling includes: The real-time values ​​of the parameter information are mapped to the numerical offsets of the corresponding entity embedding vectors, and time-related terms are introduced to characterize the dynamic characteristics of the parameter information as it changes over time, thereby realizing the adaptive representation of the boiler operating conditions by the knowledge graph and completing the numerical perception embedding modeling.

[0010] Preferably, when the graph structure feature extraction model updates entity features, contextual information related to the operating conditions is introduced, and attention weights are dynamically assigned to the connection relationships between entities, so that entities and relationships with different degrees of influence on sulfur dioxide emissions under different operating conditions obtain different feature contributions.

[0011] Preferred methods for feature extraction and fusion include: Extract the sensor temporal features of the multivariable sensor and the textual semantic features of the unstructured text information; The graph structure features and the sensor temporal features are fused together, and then the fusion result is further fused with the text semantic features to obtain the fused feature vector.

[0012] The present invention also provides a domain-adaptive enhancement-based circulating fluidized bed sulfur dioxide emission prediction system, wherein the system applies the above-mentioned method and includes: a knowledge graph construction module, an embedding modeling module, a graph structure feature extraction module, a feature fusion module, and a prediction module; The knowledge graph construction module is used to collect parameter information of circulating fluidized bed boilers, organize the parameter information based on preset entity types and relationship types, and construct a knowledge graph to characterize the boiler's operating characteristics. The embedding modeling module performs numerical perception embedding modeling on entities related to the parameter information in the knowledge graph based on the real-time operating conditions of the circulating fluidized bed boiler, and obtains the embedded boiler operation knowledge graph. The graph structure feature extraction module inputs the embedded boiler operation knowledge graph into the graph structure feature extraction model for processing. By introducing an attention allocation mechanism related to operating conditions, it assigns adaptive weights to different entities and their relationships, and extracts graph structure features that can reflect the differences in different operating conditions. During the operation of the circulating fluidized bed boiler, the feature fusion module collects unstructured text information from multivariable sensors, and extracts and fuses features from the unstructured text information and the graph structure features to obtain a fused feature vector for subsequent sulfur dioxide emission prediction. The prediction module constructs a prediction model for predicting sulfur dioxide emissions from circulating fluidized bed boilers based on the fused feature vector. The prediction model is improved by introducing uncertainty assessment and concept drift adaptive mechanisms, and sulfur dioxide emission levels are predicted based on the improved model.

[0013] Preferably, the parameter information includes: equipment structure information, operating parameter data, chemical reaction mechanism information, operating status information, and time information; The knowledge graph includes: equipment component entities, operating parameter entities, chemical substance entities, operating status entities, and time entities.

[0014] Preferably, the numerical perception embedding modeling process includes: The real-time values ​​of the parameter information are mapped to the numerical offsets of the corresponding entity embedding vectors, and time-related terms are introduced to characterize the dynamic characteristics of the parameter information as it changes over time, thereby realizing the adaptive representation of the boiler operating conditions by the knowledge graph and completing the numerical perception embedding modeling.

[0015] Preferably, when the graph structure feature extraction model updates entity features, contextual information related to the operating conditions is introduced, and attention weights are dynamically assigned to the connection relationships between entities, so that entities and relationships with different degrees of influence on sulfur dioxide emissions under different operating conditions obtain different feature contributions.

[0016] The preferred process for feature extraction and fusion includes: Extract the sensor temporal features of the multivariable sensor and the textual semantic features of the unstructured text information; The graph structure features and the sensor temporal features are fused together, and then the fusion result is further fused with the text semantic features to obtain the fused feature vector.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves a unified representation of static process knowledge and dynamic process relationships by constructing a domain-adaptive knowledge graph oriented towards the operating characteristics of circulating fluidized bed boilers, thereby enhancing the model's ability to express complex combustion and desulfurization mechanisms. By introducing an attention allocation mechanism based on operating conditions, the model can dynamically focus on key influencing factors according to changes in operating conditions, improving its adaptability to operating condition coupling and state drift. By fusing structured process data and unstructured text information, the accuracy and robustness of sulfur dioxide emission prediction are improved, providing decision support for desulfurization system operation optimization and pollutant emission control. Attached Figure Description

[0018] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall framework of the prediction model in an embodiment of the present invention. Detailed Implementation

[0020] 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.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1 In this embodiment, as Figure 1 , Figure 2 As shown, the domain-adaptive enhancement-based circulating fluidized bed sulfur dioxide emission prediction method includes the following steps: S1. Collect parameter information of circulating fluidized bed boiler, organize the parameter information based on preset entity types and relationship types, and construct a knowledge graph to characterize the boiler's operating characteristics.

[0023] The parameter information includes: equipment structure information, operating parameter data, chemical reaction mechanism information, operating status information, and time information; the knowledge graph includes: equipment component entities, operating parameter entities, chemical substance entities, operating status entities, and time entities.

[0024] In this embodiment, the entity set of the knowledge graph It includes five categories, including device component entities. , running parameter entity Chemical substances Running status entity and time stamp entities Relationship set Defined by four relation types, namely structural relations This describes the compositional and connectivity relationships between devices, including part-of and connected-to; functional relationships. Used to describe the interaction between equipment, parameters, and chemical processes, including effects, controls, and producers; quantitative relationships. This describes the numerical attributes and trends of operating parameters, including has-value, increases-with, and time relationships. This is used to describe the temporal relationships between entities, including precedes and co-occurs-with.

[0025] Considering the changing operating conditions of the circulating fluidized bed boiler over time, the constructed knowledge graph is a dynamic graph structure whose topological relationships evolve dynamically with operating conditions. Therefore, the structure of the knowledge graph is updated at different times, and a dynamic adjacency matrix is ​​used to describe the relationships between entities. The dynamic adjacency matrix is ​​defined as follows: in, Represents a dynamic adjacency matrix. Representing entities i , Indicates a relationship. Representing entities j , express t The set of entity-relation-entity triples contained in the Moment Knowledge Graph.

[0026] S2. Based on the real-time operating conditions of the circulating fluidized bed boiler, numerical perception embedding modeling is performed on entities related to parameter information in the knowledge graph to obtain the embedded boiler operation knowledge graph.

[0027] The method of numerical perception embedding modeling includes: mapping the real-time numerical values ​​of parameter information to the numerical offset of the corresponding entity embedding vector, and introducing time-related terms to characterize the dynamic characteristics of parameter information changing over time, thereby realizing the adaptive representation of boiler operating conditions changes by the knowledge graph and completing numerical perception embedding modeling.

[0028] In this embodiment, for the running parameter entity Its time t The corresponding value is denoted as To characterize the impact of changes in runtime parameter values ​​on entity features, a parameter entity is constructed over time. t Embedded representation It is composed of basis embeddings, numerical dependency components, and time drift components, and is defined as follows: in, Represents the base embedding features of the parameter entity. This represents the value-dependent components related to the parameter's numerical value. This represents the time drift component used to characterize the impact of changes in operating conditions over time.

[0029] To enhance the model's ability to express the nonlinear variation characteristics of operating parameters, value-dependent components... Modeling is performed using a Gaussian process, which is defined as follows: Among them, kernel function Defined as: in, , and This represents the kernel function parameters, used to control the scale and smoothness of the impact of parameter value changes on entity embedding.

[0030] Furthermore, to reflect the changing importance of entity relationships under different operating conditions, an adaptive adjustment mechanism based on operational context is introduced for relationship embedding.r In time t Embedded representation Defined as: in, Represents the initial embedding of the relation. Indicates the current running context Related adjustments.

[0031] To quantify the similarity between operating conditions at different points in time, time is defined. t 1 and t The context similarity for 2 is as follows: in, M This indicates the number of runtime parameters involved in context modeling. Indicates parameters p The feature scale is described above. Through this method, entity embedding and relation embedding are adaptively adjusted to changes in operating conditions.

[0032] S3. The embedded boiler operation knowledge graph is input into the graph structure feature extraction model for processing. By introducing an attention allocation mechanism related to operating conditions, adaptive weights are assigned to different entities and their relationships, and graph structure features that can reflect the differences in different operating conditions are extracted.

[0033] When updating entity features in the graph structure feature extraction model, contextual information related to operating conditions is introduced, and attention weights are dynamically assigned to the connection relationships between entities, so that entities and relationships with different degrees of impact on sulfur dioxide emissions under different operating conditions can obtain different feature contributions.

[0034] In this embodiment, a graph transformer is used to transform the features of the knowledge graph to generate entity representations with context awareness. By jointly modeling entities and their neighborhood relationships, a comprehensive characterization of key influencing factors under boiler operation is achieved.

[0035] Neighborhood feature aggregation based on attention mechanism: for any entity in the knowledge graph i Its relationship with entities in the neighboring domain j Attention weights between aij Calculate as follows: in, Representing entities i The set of neighboring entities, W Represents the feature transformation weight matrix. This represents a trainable attention parameter vector. and Representing entities respectively i and entity j Embedding features, This indicates transpose.

[0036] Entity feature update rule: After obtaining the attention weights, the embedded features of entity i are updated by weighted aggregation, and the update rule is as follows: in, This represents a non-linear activation function. Through the above method, adaptive fusion of entity features with neighborhood information is achieved.

[0037] Residual Connectivity and Layer Normalization: To enhance the stability of feature propagation and alleviate the gradient vanishing problem in deep structures, a residual connectivity and layer normalization mechanism is introduced during entity feature update, specifically as follows: Dynamic Depth Adjustment Based on Graph Complexity: Considering the differences in knowledge graph size and structural complexity under different operating conditions, this embodiment dynamically adjusts the number of layers in the graph transformer based on the graph structure characteristics. The graph transformer adjusts its depth over time... t number of layers Defined as: in, Indicates the maximum number of floors allowed. and These represent the reference entity size and the reference triple size, respectively. This represents the number of basic layers. Through the above method, an adaptive match between model complexity and graph structure characteristics is achieved.

[0038] S4. During the operation of the circulating fluidized bed boiler, unstructured text information from multivariable sensors is collected, and features of the unstructured text information and graph structure features are extracted and fused to obtain a fused feature vector for subsequent sulfur dioxide emission prediction.

[0039] The methods for feature extraction and fusion include: extracting the sensor temporal features of multivariable sensors and the textual semantic features of unstructured text information; fusing the graph structure features and sensor temporal features; and further fusing the fusion result with the textual semantic features to obtain a fused feature vector.

[0040] In this embodiment, for the multivariable sensor time-series data collected during boiler operation, a time-series feature extraction model is used to embed and model it, in order to characterize the dynamic changes of operating parameters in the time dimension. Sensor data in time... t The embedding is represented as: in, Indicates time t The corresponding sensor data vector, Indicates the length of the backtracking time window.

[0041] Simultaneously, for unstructured text information such as boiler operation logs, operation records, and anomaly descriptions, a domain-adapted large language model is used to encode semantic features of the text to extract the semantic information of operating status, operational behavior, and anomalies. The text feature embedding is represented as follows: in, Indicates time t The runtime log text, This indicates the length of the backtracking window for the text information.

[0042] In obtaining graph structure features Sensor timing characteristics and text semantic features Subsequently, a hierarchical feature fusion approach was adopted to jointly model the multi-source features. First, graph structure features were fused with sensor temporal features to obtain intermediate fused features. It is represented as: in, , This represents the weight coefficient of the corresponding feature. Indicates the interaction moderating coefficient. This represents the Hadamard product.

[0043] Based on this, the textual semantic features are further integrated into the intermediate fusion features to obtain the final fusion feature vector. : in, The weight coefficients represent the graph-temporal fusion features. The weight coefficients representing text features. express and The interaction coefficients. The weights of each feature are adaptively adjusted through an attention mechanism. The weight relationship between graph-temporal fusion features and text features is as follows: in, This represents the attention weight vector.

[0044] S5. Based on the fusion feature vector, a prediction model for predicting sulfur dioxide emissions from circulating fluidized bed boilers is constructed. Uncertainty assessment and concept drift adaptive mechanisms are introduced into the prediction model for improvement. Sulfur dioxide emission levels are then predicted based on the improved model.

[0045] In this embodiment, the prediction model uses fused feature vectors. As input, the prediction network generates predicted sulfur dioxide emissions for future time steps, and the prediction results are expressed as follows: in, Indicates the prediction time interval. This represents the prediction function, used to characterize the mapping relationship between fusion features and sulfur dioxide emissions.

[0046] To improve the reliability of the prediction results, prediction uncertainty is modeled simultaneously with the output prediction values. Specifically, the cognitive uncertainty of the model is estimated using a Monte Carlo sampling method based on random deactivation, and the calculation method is as follows: in, This indicates cognitive uncertainty. K Indicates the number of samples. k Indicates the first k Monte Carlo random sampling This represents the prediction result obtained through multiple random samplings.

[0047] Meanwhile, stochastic uncertainty is modeled by introducing an independent uncertainty estimation network, the expression of which is: in, Indicates random uncertainty. It is a standalone network used to predict log-variance.

[0048] Combining the two types of uncertainty mentioned above, the total uncertainty in the prediction process is obtained as follows: in, This indicates total uncertainty.

[0049] Furthermore, to address the concept drift problem caused by changes in operating conditions during the operation of a circulating fluidized bed boiler, a sliding window monitoring method is used to assess changes in model performance. The drift detection index is defined as follows: in, Indicates drift detection index, UThis indicates the size of the sliding window. When the drift detection metric exceeds a preset threshold... When this occurs, the model's adaptive update mechanism is triggered to adjust the prediction model parameters. The update rule is as follows: in, Indicates model parameters, Represents the loss function. This represents the adaptive learning rate, and its value is: in, This represents the initial learning rate. min This indicates taking the minimum value.

[0050] Example 2 In this embodiment, the domain-adaptive enhanced circulating fluidized bed sulfur dioxide emission prediction system includes: a knowledge graph construction module, an embedding modeling module, a graph structure feature extraction module, a feature fusion module, and a prediction module.

[0051] The knowledge graph construction module is used to collect parameter information from circulating fluidized bed boilers. Based on preset entity types and relationship types, it organizes the parameter information to construct a knowledge graph characterizing the boiler's operating characteristics. The parameter information includes: equipment structure information, operating parameter data, chemical reaction mechanism information, operating status information, and time information. The knowledge graph includes: equipment component entities, operating parameter entities, chemical substance entities, operating status entities, and time entities.

[0052] The embedding modeling module performs numerical perception embedding modeling on entities related to parameter information in the knowledge graph based on the real-time operating conditions of the circulating fluidized bed boiler, resulting in an embedded boiler operation knowledge graph. The numerical perception embedding modeling process includes: mapping real-time parameter information values ​​to the numerical offsets of the corresponding entity embedding vectors, and introducing time-related terms to characterize the dynamic characteristics of parameter information changes over time, thereby achieving adaptive representation of boiler operating condition changes in the knowledge graph and completing the numerical perception embedding modeling.

[0053] The graph structure feature extraction module inputs the embedded boiler operation knowledge graph into the graph structure feature extraction model for processing. By introducing an attention allocation mechanism related to operating conditions, it assigns adaptive weights to different entities and their relationships, and extracts graph structure features that reflect the differences in different operating conditions. When updating entity features in the graph structure feature extraction model, contextual information related to operating conditions is introduced, and attention weights are dynamically allocated to the connections between entities, so that entities and relationships with different degrees of impact on sulfur dioxide emissions under different operating conditions receive different feature contributions.

[0054] During the operation of the circulating fluidized bed boiler, the feature fusion module collects unstructured text information from multivariable sensors and extracts and fuses features from the unstructured text information and graph structure features to obtain a fused feature vector for subsequent sulfur dioxide emission prediction. The feature extraction and fusion process includes: extracting the sensor temporal features from the multivariable sensors and the textual semantic features from the unstructured text information; fusing the graph structure features and the sensor temporal features; and further fusing the fused result with the textual semantic features to obtain the fused feature vector.

[0055] The prediction module constructs a prediction model for sulfur dioxide emissions from circulating fluidized bed boilers based on fused feature vectors. The prediction model is improved by introducing uncertainty assessment and concept drift adaptive mechanisms, and sulfur dioxide emission levels are predicted based on the improved model.

[0056] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A domain-adaptive enhanced method for predicting sulfur dioxide emissions from circulating fluidized beds, characterized in that, Includes the following steps: Collect parameter information of circulating fluidized bed boilers, organize the parameter information based on preset entity types and relationship types, and construct a knowledge graph to characterize the boiler's operating characteristics; Based on the real-time operating conditions of the circulating fluidized bed boiler, numerical perception embedding modeling is performed on the entities related to the parameter information in the knowledge graph to obtain the embedded boiler operation knowledge graph. The embedded boiler operation knowledge graph is input into the graph structure feature extraction model for processing. By introducing an attention allocation mechanism related to operating conditions, adaptive weights are assigned to different entities and their relationships, and graph structure features that can reflect the differences in different operating conditions are extracted. During the operation of the circulating fluidized bed boiler, unstructured text information from multivariable sensors is collected, and feature extraction and fusion are performed on the unstructured text information and the graph structure features to obtain a fused feature vector for subsequent sulfur dioxide emission prediction. Based on the fused feature vector, a prediction model for predicting sulfur dioxide emissions from circulating fluidized bed boilers is constructed. Uncertainty assessment and concept drift adaptive mechanisms are introduced into the prediction model for improvement, and sulfur dioxide emission levels are predicted based on the improved model.

2. The method of claim 1, wherein the method is based on domain adaptation enhancement for predicting sulfur dioxide emissions from a circulating fluidized bed, and The parameter information includes: equipment structure information, operating parameter data, chemical reaction mechanism information, operating status information, and time information; The knowledge graph includes: equipment component entities, operating parameter entities, chemical substance entities, operating status entities, and time entities.

3. The method for predicting sulfur dioxide emissions from a circulating fluidized bed based on domain adaptive enhancement according to claim 1, characterized in that, The numerical-aware embedding modeling method includes: The real-time values ​​of the parameter information are mapped to the numerical offsets of the corresponding entity embedding vectors, and time-related terms are introduced to characterize the dynamic characteristics of the parameter information as it changes over time, thereby realizing the adaptive representation of the boiler operating conditions by the knowledge graph and completing the numerical perception embedding modeling.

4. The method for predicting sulfur dioxide emissions from a circulating fluidized bed based on domain adaptive enhancement according to claim 1, characterized in that, When updating entity features in the graph structure feature extraction model, contextual information related to operating conditions is introduced, and attention weights are dynamically assigned to the connection relationships between entities, so that entities and relationships with different degrees of influence on sulfur dioxide emissions under different operating conditions can obtain different feature contributions.

5. The method for predicting sulfur dioxide emissions from a circulating fluidized bed based on domain adaptive enhancement according to claim 1, characterized in that, Methods for feature extraction and fusion include: Extract the sensor temporal features of the multivariable sensor and the textual semantic features of the unstructured text information; The graph structure features and the sensor temporal features are fused together, and then the fusion result is further fused with the text semantic features to obtain the fused feature vector.

6. A domain-adaptive enhanced circulating fluidized bed sulfur dioxide emission prediction system, wherein the system applies the method described in any one of claims 1-5, characterized in that, include: The module includes a knowledge graph construction module, an embedding modeling module, a graph structure feature extraction module, a feature fusion module, and a prediction module. The knowledge graph construction module is used to collect parameter information of circulating fluidized bed boilers, organize the parameter information based on preset entity types and relationship types, and construct a knowledge graph to characterize the boiler's operating characteristics. The embedding modeling module performs numerical perception embedding modeling on entities related to the parameter information in the knowledge graph based on the real-time operating conditions of the circulating fluidized bed boiler, and obtains the embedded boiler operation knowledge graph. The graph structure feature extraction module inputs the embedded boiler operation knowledge graph into the graph structure feature extraction model for processing. By introducing an attention allocation mechanism related to operating conditions, it assigns adaptive weights to different entities and their relationships, and extracts graph structure features that can reflect the differences in different operating conditions. During the operation of the circulating fluidized bed boiler, the feature fusion module collects unstructured text information from multivariable sensors, and extracts and fuses features from the unstructured text information and the graph structure features to obtain a fused feature vector for subsequent sulfur dioxide emission prediction. The prediction module constructs a prediction model for predicting sulfur dioxide emissions from circulating fluidized bed boilers based on the fused feature vector. The prediction model is improved by introducing uncertainty assessment and concept drift adaptive mechanisms, and sulfur dioxide emission levels are predicted based on the improved model.

7. The domain-adaptive enhanced circulating fluidized bed sulfur dioxide emission prediction system according to claim 6, characterized in that, The parameter information includes: equipment structure information, operating parameter data, chemical reaction mechanism information, operating status information, and time information; The knowledge graph includes: equipment component entities, operating parameter entities, chemical substance entities, operating status entities, and time entities.

8. The domain-adaptive enhanced circulating fluidized bed sulfur dioxide emission prediction system according to claim 6, characterized in that, The process of numerical perception embedding modeling includes: The real-time values ​​of the parameter information are mapped to the numerical offsets of the corresponding entity embedding vectors, and time-related terms are introduced to characterize the dynamic characteristics of the parameter information as it changes over time, thereby realizing the adaptive representation of the boiler operating conditions by the knowledge graph and completing the numerical perception embedding modeling.

9. The domain-adaptive enhanced circulating fluidized bed sulfur dioxide emission prediction system according to claim 6, characterized in that, When updating entity features in the graph structure feature extraction model, contextual information related to operating conditions is introduced, and attention weights are dynamically assigned to the connection relationships between entities, so that entities and relationships with different degrees of impact on sulfur dioxide emissions under different operating conditions obtain different feature contributions.

10. The domain-adaptive enhanced circulating fluidized bed sulfur dioxide emission prediction system according to claim 6, characterized in that, The process of feature extraction and fusion includes: Extract the sensor temporal features of the multivariable sensor and the textual semantic features of the unstructured text information; The graph structure features and the sensor temporal features are fused together, and then the fusion result is further fused with the text semantic features to obtain the fused feature vector.

Images