Thermal power plant unit control system based on AI large model

Abstract

The invention discloses a thermal power plant unit control system based on an AI large model, and relates to the technical field of thermal power intelligent control, and the system comprises a semantic modeling module which collects the operation data of a thermal power unit, carries out the semantic mapping, constructs a semantic map reflecting the operation logic of the unit, and generates a knowledge network of mechanism and data fusion; the intelligent modeling module is used for constructing an AI large model based on a knowledge network, inputting the operation data of the thermal power generating unit into a multi-working-condition embedding layer of the AI large model for pre-training, and generating a feature vector set through a depth feature extraction and self-adaptive parameter adjustment method; and the cooperative control module transmits the feature vector set into a hierarchical linkage control structure and converts the feature vector set into a control instruction capable of coordinately controlling a boiler, a steam turbine and a water supply loop. According to the method, the semantic map and the mechanism parameter matrix are combined, so that deep fusion of the operation characteristics and the physical mechanism of the thermal power generating unit is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for thermal power plants, and in particular to a control system for thermal power plant units based on a large AI model. Background Technology

[0002] As thermal power unit control technology advances towards intelligence, data-driven AI models are increasingly being applied to the fields of operating condition identification and optimization control. Existing technologies extract operating characteristics through methods such as neural networks and time series analysis, and combine them with mechanistic models to achieve multi-variable coordinated regulation, thereby improving the stability and energy efficiency of unit operation. For example, deep learning models are used to predict combustion parameters and load demand, or fuzzy control is used to achieve adaptive adjustment of turbine power, providing a technological foundation for thermal power control under complex operating conditions.

[0003] Existing methods lack sufficient depth in semantic analysis of unit operation logic, resulting in limitations in the integration of data features and physical mechanisms. In particular, when operating conditions are highly variable, traditional models struggle to dynamically quantify the semantic coupling relationships between variables, causing control command generation to rely on static empirical parameters. This lack of adaptive characterization for nonlinear correlations between multiple systems such as combustion, steam, and load affects control accuracy and response speed. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a thermal power plant unit control system based on an AI large model to solve the problem of inaccurate dynamic coupling relationship representation caused by insufficient semantic and mechanism fusion depth in existing thermal power unit control.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a thermal power plant unit control system based on an AI large-scale model, comprising: a semantic modeling module, which collects thermal power unit operating data and performs semantic mapping to construct a semantic graph reflecting the unit's operating logic and generate a knowledge network that integrates mechanisms and data; an intelligent modeling module, which constructs an AI large-scale model based on the knowledge network, inputs the thermal power unit operating data into the multi-condition embedding layer of the AI ​​large-scale model for pre-training, and generates a feature vector set through deep feature extraction and adaptive parameter adjustment methods; a collaborative control module, which inputs the feature vector set into a hierarchical linkage control structure and transforms it into control commands that can coordinate the control of the boiler, turbine, and feedwater circuit; and a feedback correction module, which synchronously monitors changes in operating condition data during the execution of control commands and feeds the monitoring results back to the semantic graph to calibrate the semantic relationship weights in the graph.

[0007] As a preferred embodiment of the thermal power plant unit control system based on the AI ​​large model of the present invention, the thermal power unit operating data includes combustion temperature, fuel flow rate, steam pressure, steam flow rate and unit load power operating data. The specific steps for performing semantic mapping are as follows: The operating data of thermal power units are denoised, normalized, and time-aligned. The preprocessed thermal power unit operating data is input into a semantic encoding function to symbolically represent each operating variable and form an initial semantic vector. Based on the initial semantic vector, the semantic similarity between variables is analyzed according to the time series features to generate a semantic association matrix; The semantic association matrix is ​​input into the embedding transformation function, and a semantic embedding space is generated through nonlinear feature aggregation and parameter optimization. Semantic relevance weights are calculated based on variable dependency paths within the semantic embedding space, and semantic mapping is performed.

[0008] As a preferred embodiment of the AI-based large-scale model-based thermal power plant unit control system of this invention, the specific steps for constructing a semantic graph reflecting the unit's operating logic are as follows: Using semantic mapping as input, the semantic coupling degree between combustion, steam and load is calculated based on the time series features of multidimensional operating variables to form a semantic coupling matrix. Based on the semantic coupling matrix, the semantic importance vector of each running variable is calculated according to the semantic interaction relationship between nodes; A set of directional semantic relations is generated by jointly using the semantic importance vector and the semantic coupling matrix, and weighted directed edge weights are assigned. The weighted directed edge weights are normalized to generate a normalized semantic weight matrix, and a semantic graph reflecting the unit's operating logic is formed based on the normalized semantic weight matrix.

[0009] As a preferred embodiment of the AI-based large-scale model-based control system for thermal power plant units in this invention, the specific process of the generation mechanism and the knowledge network fused with data is as follows: Based on semantic mapping, semantic feature vectors of the operating status of thermal power units are extracted; Thermodynamic, fluid dynamic, and control characteristic parameters are extracted from the operating data of thermal power units to form a mechanism parameter matrix; The semantic feature vector is combined with the normalized semantic weight matrix in the semantic graph to obtain the semantic weighted embedding matrix, and then the time step homogenization process is performed on the mechanism parameter matrix. Based on time homogenization, the normalized semantic weight matrix and mechanism parameter matrix are input into the mechanism consistency discriminant function for nonlinear difference analysis to calculate the degree of coupling between semantic structure and mechanism law; Based on the degree of coupling, semantic features are embedded in the mechanism parameter representation, a weighted fusion function is used to generate a fusion feature vector, the result is normalized to form a fusion knowledge matrix, and a knowledge network of mechanism and data fusion is generated based on time series changes.

[0010] As a preferred embodiment of the AI-based large-scale model-based control system for thermal power plants according to the present invention, the construction of the AI ​​large-scale model involves inputting the thermal power unit operating data into the multi-condition embedding layer of the AI ​​large-scale model for pre-training. The specific process is as follows: Based on the knowledge network, the runtime data, semantic mapping and semantic weighted embedding matrix are time-aligned and windowed to generate working condition labels and sample sequences; Based on the working condition labels and sample sequences, a multi-working condition embedding layer, a temporal coding layer and a cross-variable attention layer are established, and a semantic weight matrix and a mechanism parameter matrix are connected to form the initial model structure; The initial model structure is trained by first warming up under a single working condition and then training under multiple working conditions. Pre-training is performed under the conditions of semantic relationship constraints provided by the semantic graph and mechanism consistency provided by the mechanism parameter matrix. After training, a large AI model is formed.

[0011] As a preferred embodiment of the AI-based large-scale model-based thermal power plant unit control system of this invention, the specific steps for generating the feature vector set are as follows: Semantic representation vectors are generated through joint computation of a multi-condition embedding layer and a cross-variable attention layer. The semantic representation vector and the mechanism parameter matrix are homogenized at time steps to obtain the basic input set. The basic input set is then jointly mapped using a nonlinear coupling function to extract the coupling features between semantic features and mechanism parameters. Gradient matching and feature sparsification are performed on the coupled features to filter out the main feature components with temporal correlation. The main feature components are subjected to multi-stage clustering and adaptive feature fusion to obtain a fused main feature representation set, which is then normalized and scaled to generate a feature vector set.

[0012] As a preferred embodiment of the AI-based large-scale model-based thermal power plant unit control system of the present invention, the hierarchical linkage control structure refers to the set of multi-level execution devices based on the fusion feature vector set output by the AI ​​large-scale model during the operation of the thermal power plant unit, the control constraints are determined according to the multivariate coupling model and fusion knowledge matrix in the semantic graph, and the system is divided into boiler layer, turbine layer, feedwater layer and auxiliary machine layer.

[0013] As a preferred embodiment of the thermal power plant unit control system based on the AI ​​large model of the present invention, the control command refers to the fusion feature vector set output by the AI ​​large model, which is generated by calculating the control constraints based on the multivariate coupling model and fusion knowledge matrix in the semantic graph and then through the hierarchical linkage control structure.

[0014] As a preferred embodiment of the AI-based large-scale model-based thermal power plant unit control system of this invention, the specific steps for synchronously monitoring changes in operating condition data are as follows: Based on the control constraints determined by the semantic graph and the fused knowledge matrix, control commands are allocated and applied to various operating parameters of the boiler, steam turbine, generator and auxiliary equipment to form initial control parameters; During the execution of initial control parameters, operating condition data is collected in real time through distributed sensor nodes and synchronized with time to generate a multi-dimensional synchronized operating condition dataset. The multidimensional synchronous working condition dataset is input into the multivariate coupled model and the AI ​​large model to obtain the estimated sequence, and the multidimensional synchronous working condition dataset is used as the field synchronization sequence. The estimated sequence and the field synchronization sequence are aligned one by one to obtain the feature bias and coupling error. Based on the feature bias and coupling error, the dynamic correlation matrix between variables is updated using the AI ​​large model to re-obtain the adjustment magnitude and direction of the control parameters. Control commands are regenerated based on control parameters and fed back to the hierarchical linkage control structure.

[0015] As a preferred embodiment of the AI-based large-scale model-based thermal power plant unit control system of the present invention, the monitoring results include multi-dimensional operational characteristic data such as combustion temperature, fuel flow rate, steam pressure, main engine speed, power generation, cooling water temperature, flue gas oxygen content, and vibration signals.

[0016] The beneficial effects of this invention are as follows: By combining semantic graphs and mechanistic parameter matrices, a deep integration of the operating characteristics and physical mechanisms of thermal power units is achieved. Through dynamic adjustment of the unit's control characteristics using a weighted fusion function, the AI ​​large-scale model can adaptively capture and optimize the complex nonlinear relationships between combustion, steam, and load under different operating conditions. Through real-time feedback correction and dynamic adjustment of control commands, the thermal power unit can respond rapidly to changing operating environments, thereby significantly improving the unit's operating efficiency and energy efficiency optimization level. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0018] Figure 1 This is a schematic diagram of a thermal power plant unit control system based on an AI-powered large model.

[0019] Figure 2 Flowchart for building the semantic graph of running data.

[0020] Figure 3 A flowchart for generating a knowledge network that integrates mechanisms and data.

[0021] Figure 4 This is a flowchart of closed-loop adaptive intelligent control. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a thermal power plant unit control system based on an AI large model, including the following modules: The semantic modeling module collects operating data from thermal power units and performs semantic mapping to construct a semantic graph that reflects the operating logic of the units, generating a knowledge network that integrates mechanisms and data.

[0026] The thermal power unit operation data is denoised, normalized, and time-aligned; the preprocessed thermal power unit operation data is input into a semantic encoding function to symbolically represent each operation variable and form an initial semantic vector.

[0027] Specifically, the collected thermal power unit operation data is denoised to remove abnormal fluctuations and noise; the operation data is normalized to ensure that the numerical range of each variable is consistent; the operation data is time-aligned according to time series characteristics to ensure that all operation data points are aligned on the same time axis; the processed thermal power unit operation data is input into a semantic encoding function to symbolically represent each operation variable and transform it into an initial semantic vector.

[0028] Based on the initial semantic vector, the semantic similarity between variables is analyzed according to the time series features to generate a semantic association matrix; the semantic association matrix is ​​input into the embedding transformation function, and a semantic embedding space is generated through nonlinear feature aggregation and parameter optimization.

[0029] Specifically, based on the initial semantic vector, the semantic similarity between each running variable is analyzed according to the time series features to obtain the semantic association matrix.

[0030] It should be noted that the parsing process analyzes the time-series data between variables, uses statistical methods to quantify the dynamic changes between variables for evaluation, quantifies their interrelationships, and generates a semantic association matrix reflecting the degree of correlation between each variable. The generated semantic association matrix is ​​then input through a nonlinear mapping mechanism to be converted into an embedding representation in a high-dimensional semantic space, optimizing the representational power of features and enhancing the model's ability to learn complex relationships. Nonlinear feature aggregation and parameter optimization are then used to generate a semantic embedding space.

[0031] Semantic relevance weights are calculated based on variable dependency paths within the semantic embedding space, and semantic mapping is performed.

[0032] Specifically, within the semantic embedding space, variable dependency paths of target semantic nodes are extracted from the multi-layer semantic relation matrix. The path strength is calculated based on the semantic relation values ​​on the path, and paths with inconsistent directions are filtered out after combining path length decay and occurrence frequency enhancement. The path strengths pointing to the same target semantic node are normalized and summarized to obtain semantic relevance weights. After mapping the semantic relevance weights to the range of 0 to 1, they are used as weighting coefficients to perform weighted projection on candidate semantic vectors in the unified semantic feature space.

[0033] Using semantic mapping as input, the semantic coupling degree between combustion, steam and load is calculated based on the time series features of multidimensional operating variables to form a semantic coupling matrix; based on the semantic coupling matrix, the semantic importance vector of each operating variable is calculated according to the semantic interaction relationship between nodes.

[0034] Specifically, using semantic mapping as input and combining the time-series features of multi-dimensional operating variables, the semantic coupling degree between combustion, steam, and load is obtained by calculating the time-series similarity and dependency between variables, generating a semantic coupling matrix. From the semantic coupling matrix, the semantic coupling degree values ​​between each operating variable and other variables are extracted to form a set of association weights. By normalizing the set of association weights, the centrality of each variable in the semantic coupling structure is calculated. The centrality values ​​of each variable at different semantic levels are then weighted and aggregated to generate a semantic importance vector that includes the influence of combustion, steam, and load levels.

[0035] It should be noted that the semantic hierarchy represents the depth and intensity of the dependency relationship between different variables. There are combustion hierarchy, steam hierarchy and load hierarchy. According to the function, dependency relationship and control objective of the variables, the combustion hierarchy, steam hierarchy and load hierarchy are assigned to different levels.

[0036] The semantic coupling degree between each variable and other variables is extracted from the semantic coupling matrix to form a weight set consisting of multiple coupling degree values. The total number of variables represents all running variables participating in the analysis. The semantic coupling degree between each variable and other variables is used to reflect the correlation strength. The weight set is used to describe the coupling distribution of the running variables in the overall semantic structure, where each variable corresponds to an independent weight set.

[0037] The coupling weights of each variable are normalized to eliminate dimensional differences. The calculation formula is as follows: ; The centrality of a variable in a semantically coupled structure is calculated using the following formula: ; In the formula, This represents the normalized semantic coupling weights. This represents the elements of the semantic adjacency matrix. If variables have semantic relationships, then... =1, otherwise =0, The semantic centrality of a variable represents its structural influence within a semantic relational network. Indicates the index number of the target variable being calculated. Indicates the relationship with the first The index numbers of other variables that are semantically related to each variable. Indicates the total number of variables used in the operation; The centrality values ​​are weighted and aggregated according to semantic hierarchy, and the calculation formula is expressed as follows: ; The final formula for calculating the semantic importance vector is expressed as follows: ; In the formula, Indicates the first The comprehensive semantic coupling strength or comprehensive index of individual variables. Indicates the first The flow characteristic coupling value of each variable, It describes the dynamic characteristics of a fluid or energy transfer process, referring to the changes in the state, temperature, pressure, and flow rate of the fluid or energy during a stage. Indicates the first The structural feature coupling value of each variable , Indicates the first The logical characteristic coupling value of each variable, , This represents the transpose of a vector.

[0038] A directional semantic relation set is generated by jointly using the semantic importance vector and the semantic coupling matrix, and weighted directed edge weights are assigned. The weighted directed edge weights are then normalized to generate a normalized semantic weight matrix. Based on the normalized semantic weight matrix, a semantic graph reflecting the unit's operating logic is formed.

[0039] Specifically, based on the semantic importance vector and the semantic coupling matrix, a set of directional semantic relations is jointly generated; by analyzing the similarity and dependency between variables in the semantic importance vector and the semantic coupling matrix, correlation analysis, covariance calculation or mutual information measurement are usually used to quantify the correlation between the semantic importance vector and the variables in the semantic coupling matrix, identify the directional relationship between each variable, and assign weighted directed edge weights to each relationship; It should be noted that for each pair of variables, a weighted product is calculated using the semantic importance vector and the coupling values ​​in the semantic coupling matrix. The strength of the association between the variables is quantified using dot product and cosine similarity methods. An example of the calculation formula is as follows: ; In the formula, This represents the weighted directed edge weight of the variable. This represents the weighting function used to calculate the semantic importance vector. Elements in the semantic coupling matrix The correlation, A vector representing the semantic importance of variables. This represents the semantic coupling value between variables in the semantic coupling matrix.

[0040] The weighted directed edge weights are allocated based on the product of the coupling degree between variables and the semantic importance.

[0041] The weighted directed edge weights are normalized to generate a normalized semantic weight matrix; based on the normalized semantic weight matrix, a semantic graph reflecting the unit's operating logic is formed.

[0042] Based on semantic mapping, semantic feature vectors of the operating status of thermal power units are extracted; thermodynamic, fluid dynamic and control feature parameters are extracted based on the operating data of thermal power units to form a mechanism parameter matrix.

[0043] Specifically, based on semantic mapping, semantic feature vectors of the operating status of thermal power units are extracted; by symbolizing the operating data of thermal power units, the actual operating data is transformed into semantic feature vectors reflecting the operating status of the units; based on the operating data of thermal power units, thermodynamic parameters, fluid dynamic parameters and control feature parameters are extracted and integrated to form a mechanism parameter matrix.

[0044] The semantic feature vector is combined with the normalized semantic weight matrix in the semantic graph to obtain the semantic weighted embedding matrix, and then the time step homogenization process is performed on the mechanism parameter matrix.

[0045] Specifically, the semantic feature vectors arranged in time series are stacked into a semantic feature matrix. The semantic feature matrix and the normalized semantic weight matrix of the semantic graph are indexed and aligned according to the semantic node order, and matrix multiplication is performed to obtain the semantic weighted embedding matrix. Based on the time index of the mechanism parameter matrix, the semantic weighted embedding matrix and the mechanism parameter matrix are homogenized in time step.

[0046] Based on time homogenization, the normalized semantic weight matrix and the mechanism parameter matrix are input into the mechanism consistency discriminant function for nonlinear difference analysis to calculate the degree of coupling between semantic structure and mechanism law.

[0047] Specifically, based on time homogenization, and by combining physical laws and operational principles, nonlinear analysis methods are used to evaluate the difference between the output of the large AI model and the mechanism parameter matrix. This yields a function for judging the consistency between semantic structure and mechanism. The normalized semantic weight matrix and the mechanism parameter matrix are input into the mechanism consistency discriminant function to perform difference analysis on the normalized semantic weight matrix and the mechanism parameter matrix. Using the mechanism consistency discriminant function, nonlinear difference analysis is performed. By using nonlinear regression to fit the nonlinear relationship between the normalized semantic weight matrix and the mechanism parameter matrix, the regression residual or fitting coefficient is calculated to quantify the coupling degree between semantic structure and mechanism, and to verify the mutual consistency between semantics and mechanism.

[0048] It should be noted that by comparing the performance of each variable at different time steps, the dependency between semantic features and mechanistic features is evaluated. By calculating the correlation at different time steps, correlation analysis is used to quantify the degree of mutual dependency between semantic features and mechanistic features, and further quantify the degree of coupling between semantic features and mechanistic features.

[0049] Based on the degree of coupling, semantic features are embedded in the mechanism parameter representation, a weighted fusion function is used to generate a fusion feature vector, the result is normalized to form a fusion knowledge matrix, and a knowledge network of mechanism and data fusion is generated based on time series changes.

[0050] Specifically, based on the degree of coupling, semantic features are combined with the mechanism parameter matrix. The semantic features are embedded into the mechanism parameter representation through a weighted fusion function to generate a fusion feature vector. The fusion feature vector is then normalized to form a fusion knowledge matrix. Based on the time series changes, the fusion knowledge matrix is ​​used to generate a knowledge network that integrates mechanism and data.

[0051] A superior approach, compared to the conventional method of statically correcting mechanistic parameters based solely on empirical parameters, dynamically adjusts the weighted fusion of semantic features and mechanistic parameters according to the degree of coupling, generates a fusion knowledge matrix by combining time series changes, and constructs a knowledge network that integrates mechanism and data based on the fusion knowledge matrix. By optimizing semantic features and mechanistic parameters through a weighted fusion function, it can not only automatically adjust the coupling relationship between features, but also dynamically optimize the fusion feature vector according to actual operating conditions, thereby enabling the knowledge network to more accurately reflect the nonlinear correlation between variables. This improves the intelligence level of thermal power unit control, enabling the control process to more accurately cope with complex and dynamically changing operating environments, and thus improves the optimization accuracy and response capability of unit operation.

[0052] The intelligent modeling module constructs a large AI model based on a knowledge network. It inputs the operating data of thermal power units into the multi-condition embedding layer of the large AI model for pre-training and generates a feature vector set through deep feature extraction and adaptive parameter adjustment methods.

[0053] Based on the knowledge network, the runtime data, semantic mapping and semantic weighted embedding matrix are time-aligned and windowed to generate operating condition labels and sample sequences.

[0054] Specifically, under the unified time index provided by the knowledge network, the running data, semantic mapping, and semantic weighted embedding matrix are aligned by timestamp and resampled to the same time length. Linear interpolation or forward padding is performed on missing points, and the numerical values ​​are normalized to the interval. The data is then divided into a continuous window sequence according to a fixed window length and sliding step size. At the same time, the rules and threshold relationships in the knowledge network are called to map the state variables in each window to working condition labels. The output consists of a sample sequence composed of aligned running data fragments, semantic mapping fragments, and semantic weighted embedding fragments, as well as the corresponding working condition labels.

[0055] It should be noted that the relationship between rules and thresholds refers to the standards or conditions defined by the knowledge network through logical reasoning methods, used to determine the process status and determine the operating condition label based on the temperature and pressure of the operating data.

[0056] Based on the working condition labels and sample sequences, a multi-working condition embedding layer, a temporal coding layer and a cross-variable attention layer are established, and a semantic weight matrix and a mechanism parameter matrix are connected to form the initial model structure.

[0057] Specifically, based on the operating condition labels and sample sequences, a multi-operating condition embedding layer is first constructed to map the running data segments, semantic mapping segments, and semantically weighted embedding segments to a unified feature space. By employing recurrent neural networks, long short-term memory networks, or gated recurrent unit structures, the dependencies between time steps in the time series data are modeled to capture the temporal dynamic characteristics of the data, thus establishing a time-series coding layer. By adopting a self-attention mechanism, the correlation and importance between different variables are obtained, and the weights of each variable are adjusted according to their mutual influence to enhance the dynamic coupling between variables, thus constructing a cross-variable attention layer. Finally, the semantic weight matrix and the mechanism parameter matrix are integrated into the network to form the initial model structure.

[0058] The initial model structure is trained by first warming up under a single working condition and then training under multiple working conditions. Pre-training is performed under the conditions of semantic relationship constraints provided by the semantic graph and mechanism consistency provided by the mechanism parameter matrix. After training, a large AI model is formed.

[0059] Specifically, the initial model structure is preheated by inputting a subset of samples corresponding to the same working condition label; based on the sample sequences corresponding to multiple working condition labels, the initial model is jointly trained in a mixed working condition training manner, and semantic relationship constraints in the semantic graph and physical consistency conditions in the mechanism parameter matrix are added during the training process; after training is completed, the optimized initial model structure and initial model parameters are solidified to form a large AI model.

[0060] Semantic representation vectors are generated through joint computation of a multi-condition embedding layer and a cross-variable attention layer.

[0061] Specifically, the sample sequences with working condition labels are sequentially input into the multi-working condition embedding layer and the cross-variable attention layer. In the multi-working condition embedding layer, the samples are encoded with features under the working condition conditions. In the cross-variable attention layer, the correlation between the various operating variables is weighted and the features are aggregated. The output of the cross-variable attention layer is then summarized at the model output to generate a semantic representation vector.

[0062] The semantic representation vector and the mechanism parameter matrix are homogenized at time steps to obtain the basic input set. The basic input set is then jointly mapped using a nonlinear coupling function to extract the coupling features between semantic features and mechanism parameters.

[0063] Specifically, the semantic representation vector and mechanism parameter matrix output by the AI ​​large model are homogenized according to time steps to align all variables on the same time scale. The aligned basic input set is then input into a nonlinear coupling function. By jointly mapping the complex nonlinear relationship between the semantic representation vector and the mechanism parameter matrix, coupling features reflecting the coupling law between semantic features and mechanism parameters are extracted.

[0064] It should be noted that a nonlinear coupling function is a mathematical or computational function, typically employing methods such as neural networks, support vector machines, or multinomial regression. It learns the complex nonlinear relationship between the input semantic representation vector and the mechanistic parameter matrix, outputting features that reflect the coupling patterns between these two matrices. This type of function can perform nonlinear transformations on the input data to capture the deep-seated dependencies between semantic features and mechanistic parameters.

[0065] Gradient matching and feature sparsification are performed on the coupled features to select the main feature components with temporal correlation. The main feature components are then subjected to multi-stage clustering and adaptive feature fusion to obtain a fused main feature representation set. Normalization and scaling are then performed to generate a feature vector set.

[0066] Specifically, gradient matching is performed on the coupling features. By calculating the point-by-point difference between the gradient change of each coupling feature component in the time series and the gradient sequence of the reference coupling behavior, the point-by-point difference between the gradient sequences of the coupling feature components and the reference coupling behavior is calculated. This quantifies the consistency of the changes in the gradient sequences of the coupling feature components and the reference coupling behavior in the time series, measures the degree of consistency between the features and the temporal correlation, and identifies feature components that are highly consistent with the temporal correlation. At the same time, sparsification is applied to the coupling features to remove redundant features and screen out the main feature components of the temporal correlation. The main feature components are input into a multi-stage clustering process, grouped according to feature similarity, and an adaptive feature fusion method is applied within each cluster to integrate information, resulting in a fused main feature representation set. Through normalization and scaling, the fusion result is transformed into a unified dimension and range, generating a feature vector set for subsequent analysis and control.

[0067] It should be noted that redundant features refer to features that are highly correlated with other features within the feature set and provide repetitive information, but do not add effective independent information to the large AI model.

[0068] In contrast to using single-stage clustering analysis or basic feature selection methods to process coupled features and generate feature vector sets, gradient matching is used to identify feature components that are highly correlated with time series, and sparsification is combined to remove low-correlation or redundant information. At the same time, multi-stage clustering and adaptive feature fusion methods are used to integrate information to obtain a fused main feature representation set. Based on normalization and scale mapping to unify the dimensions, the generated feature vector set can accurately represent the nonlinear coupling and dynamic dependence between semantic features and mechanistic parameters, thereby improving the accuracy and robustness of AI large models in analyzing the operating status of thermal power units and generating control commands.

[0069] The collaborative control module inputs the feature vector set into the hierarchical linkage control structure and generates control commands for combustion regulation, turbine power and feedwater circuit coordination based on the multivariate coupling model in the semantic graph.

[0070] The hierarchical linkage control structure refers to a multi-level set of actuators in the operation of thermal power plant units, based on the fusion feature vector set output by the AI ​​large model, the control constraints determined by the multivariate coupling model and fusion knowledge matrix in the semantic graph, and the boiler layer, turbine layer, feedwater layer and auxiliary machine layer.

[0071] Specifically, the hierarchical linkage control structure determines the control constraints of each operation link based on the fusion feature vector set output by the AI ​​big model, combined with the multivariate coupling model and fusion knowledge matrix in the semantic graph; according to different levels such as boiler, steam turbine, water supply and auxiliary equipment, corresponding control tasks are assigned and priorities are set, and the control instructions of each level are effectively executed under coordination and constraints.

[0072] Control commands are generated by a hierarchical linkage control structure after the control constraints are determined based on the fusion feature vector set output by the AI ​​large model and the multivariate coupling model and fusion knowledge matrix in the semantic graph.

[0073] Specifically, based on the fusion feature vector set output by the AI ​​large model, control constraints are derived through the multivariate coupling model and fusion knowledge matrix in the semantic graph; the hierarchical linkage control structure calculates the specific execution instructions for each control layer according to these constraints and allocates them according to the priority of the boiler, steam turbine, feedwater and auxiliary machine levels; and generates control instructions.

[0074] Based on the control constraints determined by the semantic graph and the fused knowledge matrix, control commands are allocated and applied to various operating parameters of the boiler, steam turbine, generator and auxiliary equipment to form initial control parameters.

[0075] Specifically, based on the control constraints determined by the semantic graph and the fused knowledge matrix, the control commands are decomposed and mapped to specific operating parameters of the boiler, turbine, generator, and auxiliary equipment, including fuel flow rate, air volume ratio, main steam pressure setting, turbine control valve opening, generator active power output, and auxiliary equipment start-stop status. After completing the one-to-one mapping of the control commands to each operating parameter, the target setpoints of each operating parameter are obtained, and these target setpoints are combined to form the initial control parameters.

[0076] During the execution of initial control parameters, operating condition data is collected in real time through distributed sensor nodes and synchronized with time to generate a multi-dimensional synchronized operating condition dataset.

[0077] Specifically, during the execution of initial control parameters, distributed sensing nodes continuously collect operating condition data, and the collected operating data is timestamped and corrected; the synchronized operating condition data is then organized in chronological order to generate a multidimensional synchronized operating condition dataset.

[0078] The multidimensional synchronous working condition dataset is input into the multivariate coupling model and the AI ​​large model to obtain the estimated sequence, and the multidimensional synchronous working condition dataset is used as the field synchronization sequence.

[0079] Specifically, the multidimensional synchronous operating condition dataset is input into the multivariate coupling model and the AI ​​large model, and the estimated sequence of each operating variable and its coupling relationship is calculated respectively. The prediction results are generated based on the input data. The original multidimensional synchronous operating condition dataset is used as the field synchronization sequence.

[0080] The estimated sequence and the field synchronization sequence are aligned one by one to obtain the characteristic bias and coupling error.

[0081] According to the timestamp and the name of the running variable, the values ​​of each running variable in the estimated sequence are aligned one by one with the values ​​of the corresponding time step and the corresponding running variable in the field synchronization sequence; the numerical difference of each pair of aligned running variable values ​​is evaluated to obtain the characteristic deviation that reflects the degree of deviation of a single running variable, and the coupling error that reflects the degree of deviation of the correlation between variables is calculated based on the joint change relationship between multiple variables.

[0082] Based on feature bias and coupling error, the dynamic correlation matrix between variables is updated using a large AI model to re-obtain the adjustment magnitude and direction of the control parameters.

[0083] Specifically, by analyzing feature bias and coupling error, the dynamic correlation matrix between variables is updated using a large AI model, the weights in the dynamic correlation matrix are adjusted, and based on the updated dynamic correlation matrix, the adjustment magnitude and direction of the control parameters are re-obtained.

[0084] Control commands are regenerated based on control parameters and fed back to the hierarchical linkage control structure.

[0085] Specifically, based on the relationship between control parameters and control constraints, the specific adjustment quantities of each operating object, such as boiler fuel flow and air volume settings, turbine valve opening, generator active power output, and auxiliary machine start-stop status, are obtained. Then, the adjustment quantities of each operating object are combined to generate a new set of control commands, and the set of control commands is fed back to the hierarchical linkage control structure as the input for the next adjustment cycle.

[0086] In summary, this invention achieves a deep integration of the operating characteristics and physical mechanisms of thermal power units by combining semantic graphs and mechanistic parameter matrices. By dynamically adjusting the unit's control characteristics through a weighted fusion function, the AI ​​large-scale model can adaptively capture and optimize the complex nonlinear relationships between combustion, steam, and load under different operating conditions. Through real-time feedback correction and dynamic adjustment of control commands, the thermal power unit can respond rapidly to changing operating environments, thereby significantly improving the unit's operating efficiency and energy efficiency optimization level.

[0087] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A thermal power plant unit control system based on an AI large-scale model, characterized in that: include, The semantic modeling module collects operating data of thermal power units and performs semantic mapping to construct a semantic graph that reflects the operating logic of the units, generating a knowledge network that integrates mechanisms and data. The intelligent modeling module constructs a large AI model based on a knowledge network. It inputs the operating data of thermal power units into the multi-condition embedding layer of the large AI model for pre-training and generates a feature vector set through deep feature extraction and adaptive parameter adjustment methods. The collaborative control module inputs the feature vector set into the hierarchical linkage control structure and transforms it into control commands that can coordinate the control of the boiler, steam turbine and feedwater circuit. The feedback correction module synchronously monitors changes in operating condition data during the execution of control commands and feeds the monitoring results back to the semantic graph to calibrate the semantic relationship weights in the graph.

2. The thermal power plant unit control system based on an AI large model as described in claim 1, characterized in that: The operating data of the thermal power unit includes combustion temperature, fuel flow rate, steam pressure, steam flow rate, and unit load power operating data; The specific steps for performing semantic mapping are as follows: The operating data of thermal power units are denoised, normalized, and time-aligned. The preprocessed thermal power unit operating data is input into a semantic encoding function to symbolically represent each operating variable and form an initial semantic vector. Based on the initial semantic vector, the semantic similarity between variables is analyzed according to the time series features to generate a semantic association matrix; The semantic association matrix is ​​input into the embedding transformation function, and a semantic embedding space is generated through nonlinear feature aggregation and parameter optimization. Semantic relevance weights are calculated based on variable dependency paths within the semantic embedding space, and semantic mapping is performed.

3. The thermal power plant unit control system based on an AI large model as described in claim 2, characterized in that: The specific steps for constructing a semantic graph reflecting the unit's operational logic are as follows: Using semantic mapping as input, the semantic coupling degree between combustion, steam and load is calculated based on the time series features of multidimensional operating variables to form a semantic coupling matrix. Based on the semantic coupling matrix, the semantic importance vector of each running variable is calculated according to the semantic interaction relationship between nodes; A set of directional semantic relations is generated by jointly using the semantic importance vector and the semantic coupling matrix, and weighted directed edge weights are assigned. The weighted directed edge weights are normalized to generate a normalized semantic weight matrix, and a semantic graph reflecting the unit's operating logic is formed based on the normalized semantic weight matrix.

4. The thermal power plant unit control system based on an AI large model as described in claim 3, characterized in that: The knowledge network that integrates the generation mechanism and data is described in the following specific process. Based on semantic mapping, semantic feature vectors of the operating status of thermal power units are extracted; Thermodynamic, fluid dynamic, and control characteristic parameters are extracted from the operating data of thermal power units to form a mechanism parameter matrix; The semantic feature vector is combined with the normalized semantic weight matrix in the semantic graph to obtain the semantic weighted embedding matrix, and then the time step homogenization process is performed on the mechanism parameter matrix. Based on time homogenization, the normalized semantic weight matrix and mechanism parameter matrix are input into the mechanism consistency discriminant function for nonlinear difference analysis to calculate the degree of coupling between semantic structure and mechanism law; Based on the degree of coupling, semantic features are embedded in the mechanism parameter representation, a weighted fusion function is used to generate a fusion feature vector, the result is normalized to form a fusion knowledge matrix, and a knowledge network of mechanism and data fusion is generated based on time series changes.

5. The thermal power plant unit control system based on an AI large model as described in claim 4, characterized in that: The construction of the AI ​​large model involves inputting the operating data of the thermal power unit into the multi-condition embedding layer of the AI ​​large model for pre-training. The specific process is as follows. Based on the knowledge network, the runtime data, semantic mapping and semantic weighted embedding matrix are time-aligned and windowed to generate working condition labels and sample sequences; Based on the working condition labels and sample sequences, a multi-working condition embedding layer, a temporal coding layer and a cross-variable attention layer are established, and a semantic weight matrix and a mechanism parameter matrix are connected to form the initial model structure; The initial model structure is trained by first warming up under a single working condition and then training under multiple working conditions. Pre-training is performed under the conditions of semantic relationship constraints provided by the semantic graph and mechanism consistency provided by the mechanism parameter matrix. After training, a large AI model is formed.

6. The thermal power plant unit control system based on an AI large model as described in claim 5, characterized in that: The specific steps for generating the feature vector set are as follows: Semantic representation vectors are generated through joint computation of a multi-condition embedding layer and a cross-variable attention layer. The semantic representation vector and the mechanism parameter matrix are homogenized at time steps to obtain the basic input set. The basic input set is then jointly mapped using a nonlinear coupling function to extract the coupling features between semantic features and mechanism parameters. Gradient matching and feature sparsification are performed on the coupled features to filter out the main feature components with temporal correlation. The main feature components are subjected to multi-stage clustering and adaptive feature fusion to obtain a fused main feature representation set, which is then normalized and scaled to generate a feature vector set.

7. The thermal power plant unit control system based on an AI large model as described in claim 6, characterized in that: The hierarchical linkage control structure refers to a multi-level set of actuators based on the fusion feature vector set output by the AI ​​large model during the operation of thermal power plant units, the multivariate coupling model in the semantic graph and the fusion knowledge matrix to determine control constraints, and the boiler layer, turbine layer, feedwater layer and auxiliary machine layer.

8. The thermal power plant unit control system based on an AI large model as described in claim 7, characterized in that: The control command refers to the fusion feature vector set output by the AI ​​large model, which is generated by calculating the control constraints based on the multivariate coupling model and fusion knowledge matrix in the semantic graph and then through a hierarchical linkage control structure.

9. The thermal power plant unit control system based on an AI large model as described in claim 8, characterized in that: The specific steps for synchronously monitoring changes in operating condition data are as follows: Based on the control constraints determined by the semantic graph and the fused knowledge matrix, control commands are allocated and applied to various operating parameters of the boiler, steam turbine, generator and auxiliary equipment to form initial control parameters; During the execution of initial control parameters, operating condition data is collected in real time through distributed sensor nodes and synchronized with time to generate a multi-dimensional synchronized operating condition dataset. The multidimensional synchronous working condition dataset is input into the multivariate coupled model and the AI ​​large model to obtain the estimated sequence, and the multidimensional synchronous working condition dataset is used as the field synchronization sequence. The estimated sequence and the field synchronization sequence are aligned one by one to obtain the characteristic bias and coupling error; Based on feature bias and coupling error, the dynamic correlation matrix between variables is updated using a large AI model to re-obtain the adjustment magnitude and direction of the control parameters; Control commands are regenerated based on control parameters and fed back to the hierarchical linkage control structure.

10. The thermal power plant unit control system based on an AI large model as described in claim 9, characterized in that: The monitoring results include multi-dimensional operational characteristic data such as combustion temperature, fuel flow rate, steam pressure, main engine speed, power generation, cooling water temperature, flue gas oxygen content, and vibration signals.

Images