A deep peak regulation control method and system for a thermal power generating unit based on digital twinning

Abstract

The application discloses a kind of based on digital twinning's thermal power unit deep peak shaving control method and system, belong to electric power peak shaving control field.Method includes: construct with entity node and state node as vertex, with dynamic association rule that adapts to condition from with edge unit operation state chart, by its instantiation into digital twin model, with physical thermal power unit parallel simulation calculation real-time residual vector, when residual characteristic representation model structure mismatch, by establishing temporary data causal diagram, and it is with unit operation state chart Subgraph difference, update digital twin model structure, based on the digital twin model after updating rolling optimization is carried out, solves optimal control instruction and executes;The application reduces the response delay of peak shaving control instruction, and the hysteresis of different step of virtual and real decision, enhances the safety margin and operation economy of thermal power unit, realizes to thermal power unit fast peak shaving.

Description

Technical Field

[0001] This invention belongs to the field of power peak shaving control technology, and relates to a deep peak shaving control method and system for thermal power units based on digital twins. Background Technology

[0002] With the increasing proportion of renewable energy power generation, thermal power units need to undertake more frequent and larger-scale deep peak shaving tasks to smooth grid fluctuations. Digital twin technology can build virtual models to optimize the operating performance of units under wide load and rapid load change conditions. However, the peak shaving process of thermal power involves strong coupling of multiple physical fields such as combustion, heat transfer, and flow, and is affected by multiple factors such as equipment status and coal quality fluctuations. It is a complex process with high dimensions, nonlinearity, and strong dynamics, which places extremely high demands on the construction and application of digital twin systems.

[0003] Currently, a hybrid data model approach is typically used, such as combining Kalman filtering and deep learning, to calibrate the parameters of thermal power unit mechanism models online in order to improve the short-term accuracy of the models under specific operating conditions. However, existing technologies cannot meet the needs of thermal power plants for real-time performance, spatial correlation, and integrated predictive intervention. They lack the utilization of the spatial correlation of the entire process and cross-equipment of thermal power units, and control commands are slow and risky when dealing with rapidly fluctuating peak-shaving commands. Summary of the Invention

[0004] The purpose of this invention is to provide a deep peak-shaving control method and system for thermal power units based on digital twins, so as to solve the problems of the disconnect between digital twin models and dynamic operating conditions, and the slow update of peak shaving of thermal power units with high risk in the prior art.

[0005] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution.

[0006] A deep peak-shaving control method for thermal power units based on digital twins, comprising:

[0007] Collect operating data of thermal power units, extract equipment operating parameters, set dynamic correlation rules that change with operating conditions, and construct a unit operating status diagram;

[0008] Based on the unit's operating status diagram, an instantiation process is performed to establish a digital twin model. In each control cycle of the thermal power unit, parallel simulations are conducted on the thermal power unit and the digital twin model to calculate the real-time residual between the simulation output and the actual physical output.

[0009] Based on the real-time residual, the evolution command is set, the operating data of the thermal power unit is collected when the evolution command is executed, the data cause-effect graph is generated from the operating data, and it is matched and calculated with the unit operating status diagram to obtain the set of undefined causal relationship edges.

[0010] The digital twin model is updated based on the edge set of causal relationships, and the unit's peak-shaving control is carried out in conjunction with the control loop of the thermal power unit.

[0011] Furthermore, operating data from thermal power units is collected, equipment operating parameters are extracted, and dynamic correlation rules that change with operating conditions are set to construct a unit operating status diagram, including:

[0012] Real-time data streams from the physical equipment of thermal power units are collected according to a preset data sampling frequency, and spatiotemporal alignment is performed to obtain an operational data sequence.

[0013] Based on the process design principles of thermal power units, the operating data sequence is used to construct entity nodes representing physical equipment and state nodes representing equipment operating parameters, and to assign static attributes and dynamic real-time attributes.

[0014] Clustering is used to identify the current operating conditions and operating data sequences of thermal power units under the corresponding historical and real-time data slices. The dynamic correlation strength between state nodes is calculated and quantified as an edge with a correlation function attribute. The correlation function is used to describe the dynamic mathematical relationship between nodes.

[0015] The unit operation status diagram is constructed based on entity nodes, status nodes, and edges with correlation function attributes. The edges with correlation function attributes are defined as dynamic correlation rules that change with operating conditions.

[0016] Furthermore, by clustering to identify the current operating conditions and operating data sequences of the thermal power unit under corresponding historical and real-time data slices, the dynamic correlation strength between state nodes is calculated and quantified as edges with correlation function properties. The correlation function describes the dynamic mathematical relationship between nodes, including:

[0017] The operating conditions of the running data sequence are identified by online clustering to obtain the current operating condition cluster identifier. Data slices that match the current operating condition cluster identifier are then extracted from historical and real-time data streams.

[0018] Based on the prior knowledge of the physical topology of the thermal power unit, state node pairs with physical coupling relationships are selected to form a set of node pairs to be evaluated.

[0019] For each pair of nodes in the set to be evaluated, the correlation strength is calculated based on the corresponding time-series data in the data slice using correlation metrics.

[0020] The association strength is filtered according to the preset association strength threshold, and the node pairs with the association strength greater than the association strength threshold are judged to have dynamic association;

[0021] For node pairs that are determined to have dynamic relationships, their corresponding time series data are used to generate a correlation function describing the quantitative relationship between the two through parametric fitting, and this function is used as an attribute of the edge connecting the node pairs.

[0022] Furthermore, based on the unit's operating status diagram, a digital twin model is established through instantiation. In each control cycle of the thermal power unit, parallel simulations are performed on the thermal power unit and the digital twin model. The real-time residual between the simulation output and the actual physical output is calculated, including:

[0023] Based on the current operating conditions, activate the matching dynamic association rules from the unit operating status diagram, record copies of relevant nodes and edge structures, and generate instances of activated subgraphs;

[0024] Based on the association model encapsulated by each dynamic association rule in the activation subgraph, a digital twin model matching the current control cycle is instantiated and generated.

[0025] Based on the real-time commands of the thermal power unit control system, parallel state calculations with a preset step size are performed on the physical thermal power unit and the digital twin model to obtain the actual physical output of the thermal power unit and the simulation output of the digital twin model.

[0026] Vector difference and statistical analysis are performed on the actual physical output of the thermal power unit and the simulation output of the digital twin model to obtain the real-time residual vector representing the multidimensional state differences.

[0027] Furthermore, vector difference and statistical analysis are performed on the actual physical output of the thermal power unit and the simulation output of the digital twin model to obtain a real-time residual vector representing multidimensional state differences, including:

[0028] Based on the sensor measured values ​​of the thermal power unit and the synchronous calculation output values ​​of the digital twin model, the real-time residual vector of the current control cycle is calculated through multi-channel vector difference.

[0029] Based on the time series formed by the real-time residual vector in the current and past consecutive control cycles, the mean shift, variance and autocorrelation characteristics of each residual component are calculated through sliding window statistics and trend analysis.

[0030] For the mean shift, variance and autocorrelation characteristics of each residual component, parameter calibration is triggered by a preset decision threshold, and the calibration result is used as the real-time residual representing the multidimensional state difference.

[0031] Furthermore, based on the real-time residual settings, evolution commands are set, and the operating data of the thermal power unit during the execution of the evolution commands is collected. A data causal graph is generated from the operating data and matched with the unit's operating state diagram to obtain a set of undefined causal relationship edges, including:

[0032] Evolution instructions are set based on real-time residuals, whereby the evolution instructions encapsulate the associated abnormal residual characteristics, specific state nodes, and the corresponding dynamic association rules in the unit operation status diagram.

[0033] Collect operating data of thermal power units when evolution commands are executed, and construct a variable set for causal relationship learning by windowing the operating data;

[0034] By analyzing the conditional independence relationships between variables in a variable set, causal inferences are made, and a data causal graph is constructed.

[0035] Based on the definitions of the corresponding nodes and edges in the data causal graph and the unit operation status graph, the set of causal relationship edges that exist in the data causal graph but are not defined in the dynamic association rules of the unit operation status graph is identified through subgraph isomorphic matching and difference calculation.

[0036] The physical rationality of the causal relationship edge set is judged to obtain the filtered causal relationship edge set and its quantified relationship.

[0037] Furthermore, based on the definitions of corresponding nodes and edges in the data causal graph and the unit operation status graph, through subgraph isomorphic matching and difference calculation, the set of causal relationship edges that exist in the data causal graph but are not defined in the dynamic association rules of the unit operation status graph is identified, including:

[0038] By aligning and mapping nodes semantically, a correspondence between nodes is established between the data cause-effect graph and the unit operation status graph.

[0039] Traverse the edges in the data cause-effect graph according to the node correspondence, and search in the unit operation status graph for edges with the same source node and target node, and whose dynamic association rules cover the current operating condition.

[0040] If the corresponding edge does not exist, the edge in the data cause-effect graph is marked as a candidate causal relationship edge. If the corresponding edge exists, the causal strength and direction of the edge in the data cause-effect graph are compared with the corresponding edge in the unit operation status graph.

[0041] The significance of causal strength of candidate causal relationship edges in the data causal graph is statistically analyzed, and a comprehensive score is obtained by combining the proximity of the source node and the target node in the physical thermal power unit structure to establish a set of candidate causal relationship edges.

[0042] The candidate causal relationship edge set is integrated with the candidate conflicting causal relationship edge set to obtain the causal relationship edge set.

[0043] Furthermore, the saliency of causal strength of candidate causal edges in the data causal graph is statistically analyzed. A comprehensive score is then calculated based on the proximity of the source and target nodes within the physical structure of the thermal power unit, establishing a set of candidate causal edges, including:

[0044] Calculate the topological score and statistical score of the candidate causal relationship edges based on the statistical significance of the centrality and causal strength of the connecting nodes in the data causal graph.

[0045] The physical proximity score is calculated based on the distance between the source node and the target node of the candidate causal relationship edge in the physical topology of the thermal power unit.

[0046] A comprehensive score is calculated based on topological score, statistical score, and physical proximity score, and a set of candidate causal edge pairs is obtained by combining the score with a preset screening threshold.

[0047] Furthermore, the digital twin model is updated based on the causal relationship edge set, and peak-shaving control of the unit is performed in conjunction with the control loop of the thermal power unit, including:

[0048] Based on the new causal relationship edges and their quantification relationships, the state equation and input-output mapping relationship of the digital twin model are reconstructed to generate a digital twin model with updated structure.

[0049] The updated digital twin model is deployed as a new digital twin model, dynamically instantiated, and reconnected to the parallel simulation process;

[0050] Based on the real-time peak-shaving instructions issued by the power generation control group of the power grid to which the thermal power unit is connected, a control instruction sequence is generated through the simulation results of the new digital twin model;

[0051] The control command sequence is optimized in the rolling time domain, and peak-shaving control of thermal power units is performed based on the optimization results.

[0052] A deep peak-shaving control system for thermal power units based on digital twins includes a state sensing module, a simulation module, a model evolution module, and a control module.

[0053] The status perception module is used to collect operating data of thermal power units, extract equipment operating parameters, set dynamic association rules that change with operating conditions, and construct a unit operating status diagram.

[0054] The simulation module is used to instantiate the unit's operating status diagram, establish a digital twin model, and perform parallel simulation of the thermal power unit and the digital twin model in each control cycle of the thermal power unit, and calculate the real-time residual between the simulation output and the actual physical output.

[0055] The model evolution module is used to set evolution instructions based on real-time residuals, collect the operating data of the thermal power unit when the evolution instructions are executed, learn and generate a data causal graph from the operating data, match and calculate it with the unit operating status diagram, and obtain a set of undefined causal relationship edges.

[0056] The control module is used to update the digital twin model based on the causal relationship edge set and to perform peak-shaving control of the thermal power unit in conjunction with the control loop of the thermal power unit.

[0057] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: A unit operation status diagram is constructed, and each thermal power unit is instantiated into a digital twin model during its control cycle. This model is then simulated in parallel with the physical thermal power unit, receiving the same commands. This determines whether the digital twin model is mismatched. A temporary data causal graph is learned based on the thermal power unit's operating data, and subgraph difference is performed between this graph and the unit operation status diagram to update the digital twin model structure. The updated digital twin model is then used for look-ahead rolling optimization to solve for peak-shaving control commands. This invention solves the problems in existing technologies where the digital twin model is disconnected from dynamic operating conditions, resulting in delayed and high-risk peak-shaving updates for thermal power units. It reduces the response delay of peak-shaving control commands and the lag in asynchronous virtual and real-world decision-making, enhancing the safety margin and operational economy of thermal power units, and enabling rapid peak-shaving for thermal power units.

[0058] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of this disclosure. Attached Figure Description

[0059] Figure 1 A flowchart of a deep peak-shaving control method for thermal power units based on digital twins provided by the present invention;

[0060] Figure 2 The correlation strength comparison bar chart provided for this invention;

[0061] Figure 3 The unit operation status diagram provided by this invention;

[0062] Figure 4 This invention provides a structural diagram of a deep peak-shaving control system for thermal power units based on digital twins. Detailed Implementation

[0063] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0064] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0065] Example 1

[0066] Please see Figures 1-3 This invention provides an embodiment of a deep peak-shaving control method for thermal power units based on digital twins, comprising the following specific steps:

[0067] Step S1: Collect operating data of thermal power units, extract equipment operating parameters, set dynamic association rules that change with operating conditions, and construct a unit operating status diagram.

[0068] The specific steps of step S1 are as follows:

[0069] Step S101: Collect real-time data streams from the physical equipment of the thermal power unit according to the preset data sampling frequency, and perform spatiotemporal alignment to obtain the operating data sequence.

[0070] In this embodiment, real-time data streams from the peak-shaving control system, auxiliary equipment control system, and environmental monitoring system of the thermal power unit are synchronously collected according to a preset data sampling frequency. The real-time data streams include, but are not limited to, boiler main steam pressure and temperature, turbine power generation, coal feed rate of each coal mill, total air volume, oxygen content and nitrogen oxide concentration in flue gas, and current and frequency of key auxiliary equipment. For data from different subsystems with different initial sampling periods and transmission delays, a network time protocol is used to unify the timestamps, and all data are linearly interpolated and resampled at 1-second intervals to eliminate timing misalignment caused by communication and acquisition asynchrony, resulting in a time-stricken and multi-dimensionally synchronized operating data sequence. By eliminating time deviations, a data foundation is provided for dynamic association rules, improving the reliability of operating condition characteristics. The data sampling frequency is, for example, 1 second.

[0071] Step S102: Based on the process design principles of thermal power units, construct entity nodes representing physical equipment and state nodes representing equipment operating parameters for the operating data sequence, and assign static attributes and dynamic real-time attributes.

[0072] In this embodiment, the operating data sequence is analyzed and classified based on the process design principles and system topology of the thermal power unit. Physical equipment in the thermal power unit, including but not limited to coal mills, forced draft fans, induced draft fans, feedwater pumps, turbine high-pressure cylinders, and condensers, is constructed as entity nodes. Each entity node is assigned static attributes and dynamic real-time attributes. Static attributes are entered based on equipment nameplates and design drawings, including but not limited to rated power, design efficiency, and system to which the equipment belongs. Dynamic real-time attributes are obtained from the operating data sequence and continuously updated, including but not limited to real-time operating status and health assessment index of the equipment. Key parameters representing system status in the operating data sequence, such as main steam pressure, main steam temperature, power generation, furnace negative pressure, total coal feed, total air volume, furnace flue gas oxygen content, nitrogen oxide concentration, and condenser vacuum, are constructed as state nodes. Each state node is assigned static attributes and dynamic real-time attributes. Static attributes include but are not limited to parameter units, normal value ranges, and the physical system to which the parameter belongs. Dynamic real-time attributes are the current measured value and short-term historical sequence of the parameter in the operating data sequence.

[0073] Step S103: Identify the current operating conditions and operating data sequences of the thermal power unit under the corresponding historical and real-time data slices by clustering, calculate the dynamic correlation strength between state nodes, and quantify it into edges with correlation function attributes. The correlation function is used to describe the dynamic mathematical relationship between nodes.

[0074] The specific steps of step S103 are as follows:

[0075] Step S1031: Identify the operating conditions of the running data sequence through online clustering to obtain the current operating condition cluster identifier, and extract data slices that match the current operating condition cluster identifier from historical and real-time data streams.

[0076] In this embodiment, the operating data sequence of the most recent 72 hours collected by the peak-shaving control system of the thermal power unit is used as input. Four dimensions are selected as clustering features: power generation load, main steam pressure, total coal feed, and flue gas oxygen content. Four clusters are set, corresponding to 450MW rated operating condition, 350MW conventional peak-shaving operating condition, 250MW deep peak-shaving operating condition, and 180MW ultra-low load stable combustion operating condition, respectively. The online clustering algorithm reads 3600 sets of real-time data from the most recent hour as initial cluster centers. Each time a new set of real-time data is received, the Euclidean distance between the data point and the four cluster centers is recalculated. The data point is assigned to the nearest cluster, and the center coordinates of that cluster are dynamically updated. Assuming the current real-time data is: power generation load 253MW, main steam pressure 16.8MPa, total coal feed 82t / h, and flue gas oxygen content 3.8%, the calculated distance to the center of the deep peak-shaving operating condition cluster is 2.1, and the distance to the center of the rated operating condition cluster is 17.6. This data is identified as a deep peak-shaving operating condition cluster. Using this cluster identifier as an index, data slices are created by combining all steady-state operating data segments belonging to the deep peak-shaving operating condition cluster from the past 30 days, along with data segments identified as belonging to the deep peak-shaving operating condition cluster within 300 seconds before and after the current moment, cached from the real-time data stream. The total amount of data input for subsequent correlation analysis is compressed from approximately 259,000 sets in the original 72 hours to 41,000 sets, a reduction of approximately 84%. Abnormal fluctuation data during non-steady-state processes such as start-up and shutdown, load shedding, and burner switching are removed, ensuring that the time-series data within the data slices are all within the same physical operating boundary. The steady-state operating data segment is represented by removing data rows from periods of significant disturbance such as start-up and shutdown, load shedding, burner switching, coal mill switching, and soot blowing operations when extracting historical data, retaining continuous data segments where the operating parameters of the thermal power unit change smoothly and are within the same operating condition boundary. Online clustering algorithms, such as streaming K-Means, are used.

[0077] Step S1032: Based on the prior knowledge preset on the physical topology of the thermal power unit, select state node pairs that have physical coupling relationships to form a set of node pairs to be evaluated.

[0078] In this embodiment, design data for establishing the thermal power unit is acquired, including but not limited to boiler thermal calculation sheets, turbine heat balance diagrams, flue gas duct instrumentation diagrams, and control logic function diagrams. Based on the boiler thermal calculation sheets, the main boiler-side equipment and their measurement point layout logic are extracted from the heat absorption distribution ratio of each heating surface and the flow paths of flue gas and working fluid. Based on the turbine heat balance diagram, the energy transfer paths of each turbine-side cylinder, heater, and related pressure and temperature measurement points are extracted from the locations of each stage of extraction steam inlets, heater drain directions, and feedwater enthalpy rise distribution. Based on the flue gas system duct and instrumentation diagram, the material flow relationships between flue gas system equipment and corresponding flow and pressure measurement points are extracted from the entire flue gas duct from the blower outlet to the chimney, the locations of each baffle measurement point, and the flue gas recirculation pipeline. Based on the control logic function diagram, the control signals and actuators are extracted from the controlled variables, setpoints, feedforward signals, and signal flow directions of the actuators in the regulating loop. The control relationship between the actuator and process parameters is defined. Equipment, measuring points, pipelines, and signal lines in the four types of design data are mapped as nodes, and working fluid flow, energy transfer, torque balance, and signal drive are mapped as edges, forming a physical topology association list. This serves as prior knowledge of the physical topology of the thermal power unit. A physical topology association matrix is ​​constructed between entity nodes (including but not limited to coal mills, forced draft fans, induced draft fans, feedwater pumps, and high-pressure cylinders of turbines) and state nodes (including power generation load, main steam pressure, main steam temperature, reheat steam pressure, reheat steam temperature, furnace negative pressure, and vibration amplitude of each bearing). From this physical topology association matrix, state node pairs with direct working fluid flow, energy transfer, or torque balance relationships are extracted, specifically including but not limited to total coal feed and power generation load, total coal feed and main steam pressure, total air volume and furnace flue gas oxygen content, and feedwater flow and main steam temperature, forming a set of node pairs to be evaluated.

[0079] Step S1033: For each pair of nodes in the set of nodes to be evaluated, calculate their correlation strength based on the corresponding time-series data in the data slice using correlation metrics.

[0080] In this embodiment, taking the total coal supply and power generation load node pair as an example, a sequence of total coal supply values ​​and a sequence of power generation load values ​​per second within a continuous 1800-second period are extracted from the deep peak-shaving operating condition data slice. During this period, the total coal supply is stable at 78 t / h for the first 600 seconds, then gradually increases linearly from 78 t / h to 86 t / h for the next 300 seconds, remains at 86 t / h for the next 600 seconds, and gradually decreases to 80 t / h for the last 300 seconds. Correspondingly, the power generation load is stable at 249 MW for the first 600 seconds, then gradually increases to 273 MW for the next 300 seconds, remains at 273 MW for the next 600 seconds, and gradually decreases to 254 MW for the last 300 seconds. The direction of change of the two sets of time series data at the same point in time is calculated, that is, the sign of the increment of each second compared to the previous second. It is assumed that the proportion of the period when the total coal supply increases positively is 93%, and the proportion of the period when the total coal supply decreases negatively is 93%. The percentage of periods with decreasing air volume was 91%, and the overall positive synchronous change rate was 92%. Based on this, the correlation strength value was calculated to be 0.92. Using the same calculation method to process the node pair of total air volume and furnace flue gas oxygen content, in the same data slice, the total air volume was stable at 165 kg / s for the first 800 seconds, then gradually increased to 185 kg / s for the next 400 seconds, and then gradually decreased to 150 kg / s for the next 600 seconds. Correspondingly, the furnace flue gas oxygen content was stable at 3.5% for the first 800 seconds, then gradually increased to 4.6% for the next 400 seconds, and then gradually decreased to 3.2% for the next 600 seconds. The synchronous change direction of the two sets of time series data was calculated. Assuming that the results show that the percentage of periods with a positive increase in furnace flue gas oxygen content when the total air volume increases is 95%, and the percentage of periods with a negative decrease in furnace flue gas oxygen content when the total air volume decreases is 93%, the positive synchronous change rate is as high as 94%, and the correlation strength value was calculated to be 0.94.

[0081] exist Figure 2 In the diagram, the horizontal axis is labeled with four sets of parameter node pairs, including total coal feed rate - main steam pressure, feedwater flow rate - main steam temperature, turbine control valve opening - power generation, and total air volume - furnace flue gas oxygen content. The vertical axis represents the correlation strength value in the 0-1 range. Each set of node pairs corresponds to a solid gray column, with the specific correlation strength value labeled at the top, which are 0.89, 0.73, 0.92, and 0.27 respectively.

[0082] Step S1034: Filter the association strength according to the preset association strength threshold, and determine the node pairs with association strength greater than the association strength threshold as having dynamic association.

[0083] In this embodiment, assuming a correlation strength threshold of 0.60, there are 62 node pairs to be evaluated. The correlation strength of each node pair is compared with the correlation strength threshold. Taking the total coal feed rate and main steam pressure node pair as an example, its calculated correlation strength value is 0.89, which is greater than 0.60, and it is determined to have a dynamic correlation. Taking the total air volume and flue gas oxygen content node pair as an example, its calculated correlation strength value is 0.27, which is less than 0.60, and it is determined not to have a dynamic correlation. Taking the feedwater flow rate and main steam temperature node pair as an example, its correlation strength is 0.73, which is greater than 0.60, and it is determined to have a dynamic correlation. Association; After threshold screening, it is assumed that among the 62 node pairs to be evaluated, 49 pairs of node pairs have an association strength exceeding 0.60 and are judged to have dynamic association. The remaining 13 pairs of node pairs are excluded due to insufficient association strength. The identification of dynamic association is standardized by quantitative numerical judgment, which reduces the consumption of computing resources and time overhead, and eliminates pseudo-associations that have certain correlations in statistics but are weak in physical mechanism or induced by third-party variables. Among them, the association strength threshold is obtained by statistically analyzing the association strength values ​​corresponding to node pairs with strong physical coupling relationship confirmed by thermal tests based on historical operating data of thermal power units.

[0084] Based on the historical operating data of thermal power units over the past three years, the correlation strength values ​​of nodes that have been confirmed by on-site thermal tests and experts to have a strong physical coupling relationship are statistically analyzed, and the quartiles are taken as the criteria for determining the correlation strength threshold.

[0085] Step S1035: For node pairs determined to have dynamic association, use their corresponding time series data to generate an association function describing the quantitative relationship between the two through parametric fitting, and use it as the attribute of the edge connecting the node pairs.

[0086] In this embodiment, for each node pair that has passed the threshold screening, a 3600-second time series of values ​​for that node pair during the most recent steady-state operation is extracted from the corresponding operating condition data slice. This ensures that no disturbance events such as start-up / shutdown of the mill, soot blowing, or sudden changes in load rate occur during this period, reflecting the most fundamental relationship between the nodes. A successive iterative approximation parameter estimation method is adopted, using one state node as the input and the other state node as the output. By adjusting the undetermined coefficients in the function expression, the sum of the squared errors between the predicted output value calculated based on the current coefficients and the measured output value in the data slice is minimized. For relationships exhibiting monotonic linear characteristics, a ratio is used. An expression with coefficients and a constant intercept is used to determine relationships exhibiting clear piecewise characteristics. The relationships are divided into multiple intervals based on the range of input values. A linear expression is fitted within each interval. For relationships exhibiting saturation or inflection point characteristics, a piecewise linear expression including inflection point parameters is used. After solving for the coefficients, the expression, its coefficients, the sum of squared residuals, data timestamps, the number of sample points involved in the fitting, and the applicable operating condition identifier are encapsulated to form the correlation function attribute of the node pair under the current operating condition. This correlation function attribute is written into the unit operation status diagram and stored using directed edges connecting the corresponding source and target nodes. Each edge carries a unique dynamic correlation rule identifier and the correlation function attribute package.

[0087] For example, taking the generation process of the correlation function between the total coal feed rate and the main steam pressure node pair as an example, a sequence of total coal feed rate values ​​and a sequence of main steam pressure values ​​per second within a continuous 3600-second period are extracted from the deep peak-shaving operating condition data slice. During this period, the total coal feed rate values ​​are 78t, 78t, 79t, 81t, 83t, 85t, 86t, 86t, 85t, and 84t per second, respectively, and the corresponding main steam pressure values ​​are 16.2MPa, 16.2MPa, 16.3MPa, and 16.5MPa, respectively. a. 16.7MPa, 17.0MPa, 17.2MPa, 17.2MPa, 17.1MPa, and 17.0MPa. Set an initial proportional coefficient guess of 0.18MPa / t and an initial intercept guess of 2.0MPa. Calculate the predicted main steam pressure at each moment, obtaining the predicted sequence. Difference the predicted values ​​with the measured sequence point by point. Sum the squares of all differences to obtain the first round of error summation. Adjust the proportional coefficient in steps of 0.01 and the intercept in steps of 0.1. Repeat the above calculation process. When the proportional coefficient... When the calculated value was reduced to 0.19 MPa / t and the intercept was reduced to 1.9 MPa, the sum of squared errors reached the minimum value among all trial combinations. 0.19 MPa / t was taken as the final proportionality coefficient for this node pair under the current operating conditions, and 1.9 MPa as the final intercept coefficient. These two parameters together constitute the core parameters of the correlation function. The physical meaning of the correlation function is that under the 250 MW deep peak-shaving and stable combustion condition, for every ton increase in total coal feed, the main steam pressure rises by an average of 0.19 MPa, and when the theoretical coal feed is zero, the main steam pressure tends to... The pressure is close to 1.9 MPa. This proportionality coefficient, intercept, sum of squared residuals, data acquisition start time, number of samples involved in the fitting (3600 points), and applicable operating condition identifier are encapsulated and written as attributes of a directed edge into the unit operation status diagram. This edge points from the total coal feed status node to the main steam pressure status node. When the digital twin model is instantiated under this operating condition, the control system calculates the expected response value of the main steam pressure based on the real-time total coal feed by calling this correlation function, which reduces the computational overhead compared to the method of re-mining correlation rules in real time.

[0088] exist Figure 3 In the diagram, three dashed boxes divide the system into three core areas: the boiler system, the turbine system, and the auxiliary equipment system. The boiler system area has rectangular physical nodes at its edge, labeled with the names of the coal mill, the turbine high-pressure cylinder, and the feedwater pump, respectively. The turbine system area has circular status nodes labeled with the main steam pressure parameters. The auxiliary equipment system area has the power generation parameters labeled. The nodes in each area are connected by arrows to reflect dynamic association rules, and cross-system associations are achieved through dashed arrows.

[0089] Step S104: Construct a unit operation status diagram based on entity nodes, status nodes, and edges with correlation function attributes. The edges with correlation function attributes are defined as dynamic correlation rules that change with operating conditions.

[0090] In this embodiment, directed edges with correlation function attributes are encapsulated into rules. Each directed edge is defined as an independent dynamic correlation rule, whose fields include, but are not limited to, rule identifier, source node name, target node name, correlation function expression, rule confidence level, applicable operating condition cluster identifier, and rule generation timestamp. The rule confidence level is calculated based on the fitting error and the number of sample points, and is divided into high confidence, medium confidence, and unverified, which is used to manage the rule credibility in a hierarchical manner. The dynamic correlation rules are written into the unit operation status diagram. Each dynamic correlation rule exists as a directed edge, with the direction of the edge pointing from the source node to the target node, indicating the causal or strong correlation influence direction between two state nodes under a specific operating condition. Entity nodes and state nodes are imported into the graph structure. In addition to static parameters and dynamic real-time values, the node attributes include the system partition to which the node belongs, the node importance level, and the node's associated edge count information. The nodes and edges are organized according to the attribute graph data model. The attribute sets of nodes and edges are stored in key-value pairs. All nodes and edges together constitute a complete unit operation status diagram.

[0091] Step S2: Instantiate the unit's operating status diagram and establish a digital twin model. In each control cycle of the thermal power unit, perform parallel simulations on the thermal power unit and the digital twin model, and calculate the real-time residual between the simulation output and the actual physical output.

[0092] The specific steps of step S2 are as follows:

[0093] Step S201: Activate the matching dynamic association rules from the unit operation status diagram according to the current operating conditions, record copies of relevant nodes and edge structures, and generate an instance of the activated subgraph.

[0094] In this embodiment, the current operating condition cluster identifier is used as the input, and the operating condition cluster identifier is used as the index key to perform dynamic association rule matching and retrieval on the unit operation status diagram. The retrieval scope covers the dynamic association rules stored in the form of directed edges in the unit operation status diagram. Each dynamic association rule contains a list of applicable operating condition cluster identifiers in its attribute field. The intersection operation is performed between this list and the current operating condition cluster identifier. If a match exists, the rule is determined to be active under the current operating condition. If no match exists, the rule is determined to be dormant under the current operating condition. After matching and retrieval, assuming there are a total of 49 directed edges, 32 of which have applicable operating condition cluster identifiers that include the deep peak shaving 250MW operating condition, are marked as active. Using the source node and target node connected by the 32 active directed edges as the boundary, the corresponding node and all its attributes are extracted from the unit operation status diagram. The corresponding directed edges and their carried association function attribute packages, confidence levels, and timestamps are used to form an active subgraph instance that contains valid information about the current operating condition and is physically isolated from the global state diagram.

[0095] Step S202: Based on the association model encapsulated by each dynamic association rule in the activation subgraph, instantiate and generate a digital twin model that matches the current control cycle.

[0096] In this embodiment, an active subgraph instance is used as input. All 32 directed edges in the active subgraph are traversed. The associated function attribute package is extracted from the attribute fields of each edge. Each associated function attribute package contains subfields such as function expression type identifier, coefficient set, applicable working condition identifier, and fitting error. The function expression is categorized according to its type: for linear expressions, the proportional coefficient and intercept coefficient are extracted; for piecewise linear expressions, the endpoints of each segment and their corresponding proportional coefficients and intercept coefficients are extracted; for polyline expressions, the inflection point coordinates and slopes on both sides are extracted. The coefficient set is parsed into executable numerical parameters based on the state nodes in the active subgraph. The dynamic real-time attribute values ​​are obtained, and the real-time measurement values ​​of each state node at the current moment are used as the initial state vector of the digital twin model. According to the directed edge topological connection relationship of each dynamic association rule in the active subgraph, the computational dependency order between state nodes is established. Using 32 directed edges as computation links, a directed acyclic graph computational structure with the input state node as the root and the output state node as the leaf is formed. This computational structure is loaded into a contiguous storage area in memory, and the storage location of the output variable is reserved. The instantiation of the digital twin model in memory is completed. It should be noted that the digital twin model is dynamically generated according to the current active subgraph, and its structure is adaptively reconstructed as the working conditions change.

[0097] For example, taking the instantiation process of three core dynamic correlation rules in the active subgraph under the deep peak shaving 250MW operating condition as an example, the first dynamic correlation rule is the total coal feed to main steam pressure. Its correlation function attribute package stores a proportionality coefficient of 0.19 MPa / t and an intercept coefficient of 1.9 MPa. The current real-time value of the total coal feed read from the peak shaving control system is 83 t / h. Multiplying this by the proportionality coefficient 0.19 yields 15.77 MPa, and adding this to the intercept coefficient 1.9 MPa, the simulated output value of the main steam pressure is calculated to be 17.67 MPa. The second dynamic correlation rule is the feedwater flow to main steam temperature. Its correlation function attribute package stores a proportionality coefficient of -0.03℃ / (t / h) and an intercept coefficient of 602℃. The current real-time value of the feedwater flow read is 1150 t / h. Multiplying this by the proportionality coefficient -0.03 yields -34.5℃, and adding this to the intercept coefficient 602℃, the simulated output value of the main steam temperature is calculated to be 567. The third dynamic association rule is from the turbine valve opening to the power generation. The proportional coefficient stored in the association function attribute package is 2.1MW / %, and the intercept coefficient is -25MW. The real-time value of the turbine valve opening read at the current moment is 42%. Multiplying it by the proportional coefficient 2.1 gives 88.2MW. Adding it to the intercept coefficient -25MW, the simulated power generation output value is calculated to be 63.2MW. The deviations of the above three simulated output values ​​from the measured values ​​of the physical unit at the same time, which are 17.65MPa, 567.2℃, and 63.0MW, are 0.02MPa, 0.3℃, and 0.2MW, respectively. All of them are within the allowable deviation range of the control system. Through the instantiation process, each dynamic association rule is transformed into a numerical calculation unit with a clear input-output mapping relationship. After all calculation units are connected in series according to the topology of the active subgraph, the expected response trajectory of the thermal power unit after receiving the same control command can be deduced based on the real-time measured values ​​of each state node at the current moment.

[0098] Step S203: Based on the real-time instructions of the thermal power unit control system, perform parallel state calculations with a preset step size on the physical thermal power unit and the digital twin model to obtain the actual physical output of the thermal power unit and the simulation output of the digital twin model.

[0099] In this embodiment, the real-time load command issued by the power grid dispatch automatic generation control system is used as input. Assuming the real-time load command value is 250MW and the rate of change is 3MW / min, the real-time load command is distributed to two actuators. This includes writing the command into the setpoint register of the thermal power unit coordinated control system via the network communication interface of the peak-shaving control system; driving the actuators such as the coal feeder and turbine control valves through the underlying PID control loop to move the thermal power unit closer to the target load; and writing the real-time command into the input port of the digital twin model as the external driving signal for the model within this control cycle. The preset step size for parallel state calculation is set to 300ms, and three consecutive calculation steps are executed within each control cycle, covering the period from the current moment to 0.9 MW / min. The dynamic response process of the unit is as follows: The digital twin model, based on the real-time values ​​of 32 state nodes such as the current total coal feed rate of 83t / h, water flow rate of 1150t / h, and turbine valve opening of 42%, combined with the real-time load command of 250MW and three core correlation functions, extrapolates the predicted values ​​of key output quantities such as main steam pressure, main steam temperature, and power generation 300ms later; The physical thermal power unit performs closed-loop regulation by the peak-shaving control system within the same 300ms, and its sensors transmit the actual physical output values ​​such as main steam pressure, main steam temperature, and power generation at the end of the calculation; The simulation output of the digital twin model and the output in the sensor input register of the physical thermal power unit are collected to obtain two sets of output results with timestamp alignment.

[0100] Step S204: Perform vector difference and statistical analysis on the actual physical output of the thermal power unit and the simulation output of the digital twin model to obtain the real-time residual vector representing the multidimensional state differences.

[0101] The specific steps of step S204 are as follows:

[0102] Step S2041: Based on the sensor measured values ​​of the thermal power unit and the synchronous calculation output values ​​of the digital twin model, calculate the real-time residual vector of the current control cycle through multi-channel vector difference.

[0103] In this embodiment, the measured values ​​of sensors at key status nodes of the physical unit and the output of the corresponding status nodes in the digital twin model are read. The two sets of values ​​are arranged into two multi-dimensional vectors, assuming each vector has a dimension of 32, and are denoted as the physical measured vector and the model simulation vector, respectively. By performing element-wise difference operations, that is, subtracting the corresponding component in the model simulation vector from each component in the physical measured vector, the real-time residual vector of the current control cycle is obtained. Each component in the real-time residual vector corresponds to the virtual-real deviation of a status node at the current moment, including but not limited to the residuals of main steam pressure, main steam temperature, power generation, furnace negative pressure, furnace flue gas oxygen content, and nitrogen oxide concentration. By compressing it into a structured vector, the problem of insensitivity to multi-variable coupling mismatch in traditional single-point threshold monitoring is avoided.

[0104] For example, taking the residual calculation process of three core state nodes under the current deep peak shaving 250MW operating condition as an example, at the end of the 300ms calculation step, the measured value of the main steam pressure sensor read from the peak shaving control system is 17.65MPa, and the synchronous calculation output value of the main steam pressure read from the digital twin model is 17.67MPa. Subtracting the two, the real-time residual of the main steam pressure is -0.02MPa; the measured value of the main steam temperature sensor read from the peak shaving control system is 589.2℃; the value read from the digital twin model is... The main steam temperature synchronous calculation output value is 589.6℃. Subtracting the two values ​​gives a real-time residual of -0.4℃ for the main steam temperature. The measured value of the power generation sensor read from the peak control system is 63.0MW. The power generation output read from the digital twin model is 63.2MW. Subtracting the two values ​​gives a real-time residual of -0.2MW for the power generation. The real-time residuals of the remaining 29 state nodes are calculated in the same way. The 32 residual values ​​are combined into a 32-dimensional vector according to the node order, which is the real-time residual vector of the current control cycle.

[0105] Step S2042: Based on the time series formed by the real-time residual vector in the current and past consecutive control cycles, calculate the mean shift, variance and autocorrelation characteristics of each residual component through sliding window statistics and trend analysis.

[0106] In this embodiment, the sliding window length is set to 30 control cycles, corresponding to an actual time of 30 seconds. Based on the current moment and the past 29 control cycles, 30 real-time residual vectors are obtained, each vector being 32-dimensional, forming a 30-dimensional window. A residual matrix of 32 is generated. Statistical analysis is performed on this residual matrix column by column. For example, the average value of the main steam pressure residual column is calculated to obtain the mean shift of the residual component within the current sliding window. The dispersion of each component of the main steam pressure residual column relative to the mean is calculated to obtain the variance of the residual component. The correlation between each component of the main steam pressure residual column and its own time-delayed sequence is calculated to obtain the autocorrelation coefficient of the residual component. The calculation process is repeated for the 32 residual components to obtain 32 sets of mean shifts, variances, and autocorrelation coefficients. The sliding window mechanism enables continuous tracking of the dynamic evolution of errors in the digital twin model. The mean shift characterizes whether the model has persistent bias, the variance characterizes the stability of the model's predictions, and the autocorrelation coefficient characterizes whether the error has a trend of accumulation. When the ratio between the total coal feed and the main steam pressure shifts due to a slow decrease in the calorific value of the fuel, the mean shift of the main steam pressure residual will gradually increase from near zero, the variance will remain stable, and the autocorrelation coefficient will show a positive correlation.

[0107] Step S2043: For the mean shift, variance and autocorrelation characteristics of each residual component, trigger parameter calibration through a preset decision threshold, and use the calibration result as a real-time residual vector representing the multidimensional state difference.

[0108] In this embodiment, a decision threshold is set for each residual component. The decision threshold is set based on the measurement noise amplitude and allowable fluctuation range of the process under peak-shaving conditions of the thermal power unit. Specifically, a mean offset threshold, a variance threshold, and an autocorrelation coefficient threshold are set. For example, the mean offset threshold is taken as 1.5 times the maximum allowable error corresponding to the accuracy level of the measuring instrument at that state node; the variance threshold is taken as twice the historical statistical value of the residual variance during steady-state operation; and the autocorrelation coefficient threshold is taken as 0.6 to distinguish between random fluctuations and trend drift. Taking the main steam pressure residual as an example, the preset mean offset threshold is ±0.05 MPa, and the variance threshold is 0.01 MPa. 2 The autocorrelation coefficient threshold is 0.60. The mean shift, variance, and autocorrelation coefficient of the main steam pressure residuals are compared one by one with the decision threshold. If the absolute value of the mean shift is greater than 0.05 MPa and the variance is less than 0.01 MPa... 2If the autocorrelation coefficient is greater than 0.60, the residual component is determined to exhibit steady-state bias characteristics, triggering the parameter calibration process for the dynamic association rule related to that component. The parameter calibration process takes thirty sets of data within a sliding window as input, and uses the principle of minimizing the sum of squared errors to update the proportional coefficient and intercept coefficient of the association function online. The updated coefficients are written into the attribute field of the corresponding dynamic association rule in the unit operation status diagram, and the simulation output value at the current moment is recalculated using the calibrated association function to obtain the calibrated real-time residual component. If any of the indicators, such as mean offset, variance, or autocorrelation coefficient, does not meet the decision threshold, parameter calibration is not initiated. After all residual components have been judged and calibrated, the calibrated or original residual components are reassembled into a new real-time residual vector.

[0109] Step S3: Set evolution instructions based on real-time residuals, collect the operating data of the thermal power unit when the evolution instructions are executed, learn and generate a data causal graph from the operating data, match and calculate it with the unit operating status diagram, and obtain a set of undefined causal relationship edges.

[0110] The specific steps of step S3 are as follows:

[0111] Step S301: Set evolution instructions based on real-time residuals, wherein the evolution instructions encapsulate the abnormal residual characteristics associated with them, specific state nodes, and the corresponding dynamic association rules in the unit operation status diagram.

[0112] In this embodiment, the real-time residual vector is used as input, and each residual component is compared with a preset anomaly judgment threshold. The anomaly judgment threshold is set based on the accuracy of the measuring instruments at each state node and the allowable fluctuation range of the process. For example, the main steam pressure residual anomaly judgment threshold is... 0.10 MPa, the threshold for judging abnormal residual temperature of main steam is... At 1.5℃, the threshold for judging abnormal residual power generation is... For 1.0MW, if the absolute value of any residual component exceeds the corresponding anomaly judgment threshold for three consecutive control cycles, the state node is marked as an abnormal state node, and its residual value, residual change rate, and over-limit duration at the current moment are recorded as abnormal residual features. With the abnormal state node as the endpoint, the directed edges with the abnormal state node as the target node are back-searched in the unit operation status diagram. The directed edges with applicable operating condition identifiers that include the current operating condition and have a confidence level of medium or high are selected. The dynamic association rule identifier, source node name, and association function coefficient are extracted. The abnormal residual features, abnormal state node identifiers, associated dynamic association rule identifiers, and rule details are encapsulated into a structured evolution instruction.

[0113] Step S302: Collect the operating data of the thermal power unit when the evolution command is executed, and construct a variable set for causal relationship learning by windowing the operating data.

[0114] In this embodiment, the abnormal state node identifier and its associated dynamic association rules carried in the first evolution command are parsed. A breadth-first traversal is performed on the unit operation status diagram with the abnormal state node as the core, extracting all state nodes within a distance of no more than two hops as candidate related variables. The source nodes involved in the dynamic association rules in the evolution command are included in the variable set. For example, for the current deep peak shaving 250MW operating condition, if the abnormal state node is the main steam pressure, 15 candidate nodes are obtained through traversal. Adding the total coal feed of the source nodes in the dynamic association rules, the variable set is finally determined to contain 16 state nodes. Using the time of evolution command generation as the reference time zero, the time-series data of the 16 state nodes within a 400-second time window (300s before and 100s after the reference time zero) are collected per second, forming a [data set / data set]. The data matrix is ​​subjected to steady-state screening, removing data rows from periods of large disturbances such as coal mill switching and soot blowing operations of thermal power units, and retaining data rows of continuous steady-state operation, resulting in an adjusted data matrix as a variable set.

[0115] Step S303: Infer causal direction by analyzing the conditional independence relationships between variables in the variable set, and establish a data causal graph.

[0116] In this embodiment, each column of the variable set is defined as a random variable, corresponding to the observation sequence of the state node in the unit operation status diagram. Each row represents the synchronous observation value of all random variables at a sampling time. The zero-order partial correlation coefficient and its significance probability between any two random variables are calculated. The significance probability is calculated using a conditional independence test. Random variable pairs with significance probabilities less than an empirical threshold are selected as candidate edges with potential correlation. A conditional independence test method with increasing condition set size is used, and a third random variable is introduced as a condition variable. The conditional independence coefficient is calculated based on the original two random variables given this condition variable. The partial correlation coefficients and significance probabilities of each random variable are analyzed. If the significance probability is greater than an empirical threshold, the edge is considered a spurious association and is removed. If the significance probability is less than the empirical threshold, it is retained. For the retained edges, the causal direction is determined based on the ratio of the residual variance between adjacent random variables and conditional variables or independent component analysis based on non-Gaussianity. After testing and orienting all variable pairs, the retained directed edges and their corresponding random variable nodes are organized into a directed acyclic graph as the data causal graph. The empirical threshold is set to the significance level commonly used in statistics, usually 0.05.

[0117] Step S304: Based on the definitions of the corresponding nodes and edges in the data cause-effect graph and the unit operation status graph, identify the set of causal relationship edges that exist in the data cause-effect graph but are not defined in the dynamic association rules of the unit operation status graph through subgraph isomorphic matching and difference calculation.

[0118] The specific steps of step S304 are as follows:

[0119] Step S3041: Establish the correspondence between nodes in the data cause-effect graph and the unit operation status graph through node semantic alignment and mapping.

[0120] In this embodiment, a naming mapping dictionary is set up to batch convert all node names in the data cause-effect graph into standard node names in the unit operation status graph. For nodes not covered in the naming mapping dictionary, a fuzzy matching algorithm based on string similarity is used to calculate the edit distance between the node names in the data cause-effect graph and each name in the set of node names in the unit operation status graph. The node with the smallest edit distance that is less than a set distance threshold is selected as the matching result, establishing a mapping relationship between the nodes in the data cause-effect graph and the unit operation status graph. It should be noted that both the data cause-effect graph and the unit operation status graph contain status nodes representing the operating data of the thermal power unit. The node names in the data cause-effect graph are described using the original measurement points in the peak-shaving control system, such as main steam pressure, total coal feed, and turbine valve opening. The node names are appended with system affixes and serial numbers, such as main steam pressure (boiler side), total coal feed (coal mill A), and turbine valve opening (high-pressure valve). The naming mapping dictionary uses the original measuring point identifier of the peak-shaving control system as the primary key and the standard node name of the unit operation status diagram as the value. It includes, for example, the correspondence between more than 3,000 measuring points and 1,200 status nodes in a thermal power plant. Assuming there is a supercritical thermal power unit, measuring points with the same physical meaning are named 10HFC10CT001 and 20HFC10CT001 in Unit 1 and Unit 2, respectively. The two are only different in the first two digits, and the edit distance is 2. Due to the phased modification of equipment, the main steam temperature measuring point is named 10HAC30CT101 and 10HAC30CT101R before and after the modification, respectively. The latter adds the suffix R, and the edit distance is 2.

[0121] Step S3042: Traverse the edges in the data cause-effect graph according to the node correspondence relationship, and search in the unit operation status graph for edges with the same source node and target node, and whose dynamic association rules cover the current operating condition.

[0122] In this embodiment, based on the correspondence between nodes and the edge set of the data causal graph, for each edge in the edge set, the source node and target node of the edge are converted into standard node identifiers in the unit operation status diagram through the node mapping relationship. Using the converted source node identifier and target node identifier as the composite key, an exact match query is performed in the dynamic association rule index table of the unit operation status diagram. The dynamic association rule index table uses the concatenated string of the source node identifier and target node identifier as the primary key, storing dynamic association rule records for all versions and all operating conditions under that node. Each record includes a list of applicable operating condition cluster identifiers, a set of association function coefficients, a confidence level, and rule generation. The timestamp is used to filter records in the applicable operating condition cluster identifier list that contain the operating condition cluster identifier in the query results. If there is at least one dynamic association rule that meets the operating condition matching conditions, it is determined that the edge has a corresponding edge in the unit operation status diagram. If there is no dynamic association rule that meets the operating condition matching conditions, it is determined that the edge does not have a corresponding edge in the unit operation status diagram. The operating condition filtering mechanism ensures the dynamic adaptability of the comparison benchmark and avoids misjudging historical rules that are effective in other operating conditions but not applicable to the current operating condition as existing knowledge. The data causal graph edge set contains all directed causal relationship edges obtained from the current analysis. Each edge consists of a source node, a target node, and causal strength.

[0123] Step S3043: If the corresponding edge does not exist, mark the edge in the data cause-effect graph as a candidate causal relationship edge. If the corresponding edge exists, compare the causal strength and direction of the edge in the data cause-effect graph with the corresponding edge in the unit operation status graph.

[0124] In this embodiment, edges in the data causal graph that are determined to have no corresponding edge are marked as candidate causal relationship edges, and the current operating condition cluster identifier, the causal strength value, significance probability, and data time window information of the edge in the data causal graph are appended and written to the candidate causal relationship edge temporary storage area. For edges in the data causal graph that are determined to have a corresponding edge, the dynamic association rules matched in the unit operation status graph are extracted, and the association function coefficient set and function expression type of the dynamic association rules are obtained. From these, the causal direction and causal strength are parsed out. The causal direction is reflected by the direction of the directed edges, and the direction of the edges in the data causal graph is compared with the direction of the directed edges in the unit operation status graph. If the two points are consistent, they are considered to be in the same direction; if they are in opposite directions, they are marked as conflicting edges. The causal strength is compared based on the specific form of the correlation function. For linear expressions, the signs and absolute values ​​of the regression coefficients in the data causal graph and the proportional coefficients in the unit operation status graph are compared. If the signs are opposite or the absolute value deviation exceeds, for example, 50%, they are marked as edges with significant differences in strength. For nonlinear expressions, the signs and magnitudes of the derivatives of the output with respect to the input at key quantiles are compared. Edges with conflicting directions or significant differences in strength are also marked as candidate causal relationship edges, along with the difference type, difference quantification index, and the respective parameter values ​​in the data causal graph and the unit operation status graph.

[0125] Step S3044: Perform significance statistics on the causal strength of candidate causal relationship edges in the data causal graph, and combine the proximity of the source node and the target node in the physical thermal power unit structure to make a comprehensive score, and establish a set of candidate causal relationship edges.

[0126] In this embodiment, based on the causal strength value and significance probability corresponding to the candidate causal relationship edges and their accompanying data causal graph, the significance probability is extracted from the attributes of each candidate causal relationship edge. Edges with significance probability values ​​less than an empirical threshold are judged as statistically significant and assigned a statistical score of 1.0; edges with significance probability values ​​greater than or equal to the empirical threshold and less than 0.10 are judged as marginally significant and assigned a statistical score of 0.6; edges with significance probability values ​​greater than or equal to 0.10 are judged as insignificant and assigned a statistical score of 0.2. The physical proximity score of each candidate causal relationship edge is calculated. The associated information of the physical equipment corresponding to the source node and the target node is read from the unit operation status graph, such as the pipeline connection and equipment layout topology data of the boiler system, turbine system and flue gas system in the physical thermal power unit. A physical connection distance matrix between equipment is established, and the physical connection distance matrix between equipment is established for those directly connected by pipelines and flues. For nodes connected by wires, assuming a hop count of 1, the physical proximity score is 1.0. For nodes indirectly connected via an intermediate device, assuming a hop count of 2, the physical proximity score is 0.7. For nodes connected via two or more intermediate devices, assuming a hop count of 3 or more, the physical proximity score is 0.4. For nodes belonging to different systems without a clear medium transmission path, the physical proximity score is 0.1. The statistical score and the physical proximity score are weighted and summed according to weight coefficients to calculate the comprehensive score of each candidate causal relationship edge. A screening threshold is set, and candidate causal relationship edges with a comprehensive score greater than or equal to the screening threshold are retained. False discovery edges with too low scores, statistical unreliability, or lack of physical connection basis are removed to establish a set of candidate causal relationship edges. The weight coefficients and screening thresholds are obtained through thermal testing, and the screening threshold is a positive number less than 1.

[0127] Step S3045: Integrate the candidate causal relationship edge set with the candidate conflicting causal relationship edge set to obtain the causal relationship edge set.

[0128] In this embodiment, based on the candidate causal relationship edge set, in the unit operation status diagram, candidate conflicting causal relationship edges that have corresponding edges but conflict in direction or have intensity differences exceeding a threshold are parsed one by one. The attached difference type label and the causal intensity parameters in both diagrams are read. For edge conflicting in direction, the causal direction determined by statistical testing in the data causal diagram is used as the standard, generating an edge to be replaced. Its source node and target node are set according to the new direction, and the initial value of the correlation function coefficient is taken as the regression coefficient in the data causal diagram. A discard instruction for the original edge is generated. When the intensity difference edge has the same direction, the node direction of the original edge is retained, and only the correlation function coefficient is updated to the latest fitted coefficient value in the data causal diagram and marked as an edge to be updated. For missing causal edges, they are entered into the integration process as edges to be added. Edges to be added, edges to be updated, and edges to be discarded are categorized according to operation type. The system generates a structured set of causal edges. Each record includes an opcode, source node identifier, target node identifier, set of correlation function coefficients, applicable operating condition cluster identifier, initial confidence value, and data source timestamp. By automating conflict resolution strategies and incremental update logic, it achieves consistency and timeliness maintenance of knowledge in the unit operation state diagram, ensuring that the same rule under the same operating condition will not have multiple contradictory versions. The candidate causal edge set includes missing causal edges where no corresponding edge exists in the unit operation state diagram. For conflicting edges where a corresponding edge already exists in the unit operation state diagram but the data causal diagram shows the opposite direction, it is determined that the original dynamic correlation rule has become invalid or is incorrect. Therefore, a discarded edge is generated, indicating that the original dynamic correlation rule should be deleted from the unit operation state diagram to prevent erroneous knowledge from continuing to pollute the digital twin model.

[0129] Step S305: Perform a physical rationality judgment on the causal relationship edge set to obtain the filtered causal relationship edge set and its quantified relationship.

[0130] In this embodiment, based on the causal relationship edge set, the equipment type labels and dimensional information of the source and target nodes in the physical topology of the thermal power unit are loaded for each edge to be added and updated. The physical quantity units corresponding to the source and target nodes are read from the static attribute library of the unit operation status diagram. A physical rationality rule library is set, including dimensional consistency constraints, energy conservation boundary constraints, and equipment output limit constraints. The dimensional consistency constraint requires that the physical quantity units of the source and target nodes can be matched after multiplication and division operations. Edges with mismatched dimensions are judged as unreasonable and removed. The energy conservation boundary constraint is for edges involving power, heat, and flow conversion, and is set according to the energy balance range under the design conditions of the thermal power unit. The upper and lower limits of the coefficients are set. If the coefficient of an edge in the data causal graph exceeds this range, it is judged as unreasonable. For edges that pass the dimensional consistency constraint and energy conservation boundary constraint, the causal direction rationality check is further performed. For example, according to the principle of unidirectional energy transfer of thermal power units, the amount of fuel can affect the steam pressure, but the steam pressure cannot affect the amount of fuel in the reverse direction. The feedwater flow rate can affect the steam temperature, but the steam temperature cannot affect the feedwater flow rate in the reverse direction. If the direction of the edge to be added or updated is opposite to the physical causal direction, it is removed or corrected to the correct direction. For discarded edges, they are directly put into the execution queue. Edges that fail to pass any constraint are removed from the causal relationship edge set, forming a filtered causal relationship edge set and its quantification relationship.

[0131] Step S4: Update the digital twin model based on the causal relationship edge set, and combine it with the control loop of the thermal power unit to carry out peak shaving control of the unit.

[0132] The specific steps of step S4 are as follows:

[0133] Step S401: Based on the new causal relationship edges and their quantization relationships, reconstruct the state equation and input-output mapping relationship of the digital twin model to generate a digital twin model with updated structure.

[0134] In this embodiment, the causal relationship edge set includes three types of operation records: edges to be added, edges to be updated, and edges to be discarded. The edges to be discarded are traversed, each record carrying a source node identifier, a target node identifier, and an applicable operating condition cluster identifier. The storage location of the corresponding dynamic association rule is located from the unit operation status diagram. The state of the dynamic association rule is marked as discarded, and it is removed from the generation source of the active subgraph of the current operating condition. For each edge to be updated, its source node identifier, target node identifier, new association function coefficient set, and applicable operating condition cluster identifier are extracted. Using the primary key of the node pair formed by the source node and target node of the edge to be updated as the search condition, existing dynamic association rules with the same node pair and whose applicable operating condition cluster identifier includes the current operating condition are retrieved from the unit operation status diagram. Once a match is found, the corresponding fields in the association function attribute package of the original dynamic association rule are successively overwritten using the proportional coefficients, intercept coefficients, or segmented interval endpoint parameters and corresponding slopes from the new association function coefficient set. For each new edge to be added, a new directed edge is created in the unit operation status diagram based on its source node identifier and target node identifier. The primary key of its node pair is formed by concatenating these two identifiers. The applicable operating condition cluster identifier, association function coefficient set, function expression type, initial confidence value, rule generation timestamp, and data source evolution instruction of the new directed edge are sequentially written into the edge attributes, and a unique and non-repeating dynamic association rule identifier is assigned. Based on the updated dynamic association rule set, the digital twin model under the current operating condition is regenerated using the current operating condition cluster identifier as the filtering condition.

[0135] Step S402: Deploy the updated digital twin model as a new digital twin model, dynamically instantiate it, and reconnect it to the parallel simulation process.

[0136] In this embodiment, the updated digital twin model structure definition file is read, and the double-precision floating-point state variable storage area is reallocated to the state nodes according to the computational dependency topology order of the node-to-primary-key organization. Corresponding function pointers are set for dynamic association rules, including linear single-input single-output, piecewise linear single-input single-output, and linear multi-input single-output. The sensor measured values ​​corresponding to the state nodes at the current time are read, and the initial state of the new digital twin model instance is assigned according to the sensor measured values. After the initial state assignment is completed, the memory object pointer of the new digital twin model instance is registered to the active model pool of the parallel simulation scheduler, and the pointer position of the original digital twin model instance is atomically replaced. At the beginning of the next control cycle, that is, at the rising edge of the first 1-second synchronization pulse after the completion of this deployment, the scheduler synchronously distributes the real-time load command issued by the automatic power generation control of the power grid to the new digital twin model instance and the physical unit. The new digital twin model instance is officially re-integrated into the parallel simulation cycle.

[0137] Step S403: Based on the real-time peak-shaving instructions issued by the power generation control group of the power grid to which the thermal power unit is connected, generate a control instruction sequence through the simulation results of the new digital twin model.

[0138] In this embodiment, the real-time peak-shaving command issued by the power grid generation control group is used as input, including the target power generation setpoint, the target load change rate limit, and the command issuance timestamp. The real-time simulation output values ​​of the current state node are read from the new digital twin model, including but not limited to main steam pressure, main steam temperature, power generation, furnace negative pressure, furnace flue gas oxygen content, and nitrogen oxide concentration. Assuming a prediction time domain length of 60s, a control time domain length of 30s, and a simulation step size of 300ms, within the prediction time domain, the current state is used as the initial condition, and within the control time domain, sequences such as coal feed rate command sequence, turbine valve opening command sequence, feedwater flow rate command sequence, total air volume command sequence, and ammonia injection rate command sequence are used as variables to be optimized. The simulation output values ​​of the current state node are read from the new digital twin model, including but not limited to main steam pressure, main steam temperature, power generation, furnace negative pressure, furnace flue gas oxygen content, and nitrogen oxide concentration. The simulation interface of the digital twin model calculates the state trajectory of the thermal power unit under real-time peak shaving commands and determines the control command sequence. The control command sequence is weighted and summed according to the sum of squares of power generation tracking error, the sum of squares of total coal feed rate change, and the sum of squares of turbine valve actuation amplitude. The summation results are sorted from smallest to largest, for example, the first three control command sequences are used as output. Among them, the sum of squares of power generation tracking error has the largest weight because it directly corresponds to the automatic power generation control command of the power grid; the sum of squares of total coal feed rate change has the second largest weight because drastic fluctuations in coal feed can induce furnace pressure fluctuations and combustion instability; the sum of squares of turbine valve actuation amplitude has the smallest weight, although its frequent actuation affects valve life, it is not a major limiting factor under normal peak shaving frequency.

[0139] Step S404: Perform rolling time-domain optimization on the control command sequence, and perform peak-shaving control on the thermal power unit based on the optimization results.

[0140] In this embodiment, a control instruction is selected from the control instruction sequence as the optimized solution for the current control cycle. The coal feed rate instruction, turbine valve opening instruction, feedwater flow rate instruction, total air volume instruction, and ammonia injection rate instruction corresponding to the first control step in the control instruction are written into the setpoint register of the corresponding analog output channel of the peak shaving control system. In the next control cycle, the peak shaving control system sends the control instructions to the coal feeder frequency converter, the turbine digital electro-hydraulic control system, the feedwater pump turbine governor, the blower impeller actuator, and the denitrification ammonia injection regulating valve via the fieldbus. The instructions of the remaining steps in the control instruction sequence are cached in memory as hot backup data for use as an initial solution reference during the rolling optimization in the next control cycle. At the beginning of the next control cycle, the real-time peak shaving instructions issued by the grid power generation control group are reread. The updated measured state of the unit is used as the initial condition, and the control instruction generation process is repeated to achieve rolling optimization peak shaving. It should be noted that the rolling time-domain optimization is based on the standard implementation of model predictive control, combined with digital twin model, dynamic activator graph, and causal discovery-driven update mechanism.

[0141] Example 2

[0142] Please see Figure 4 One embodiment of the present invention provides a deep peak-shaving control system for thermal power units based on digital twins, comprising a state perception module, a simulation module, a model evolution module, and a control module.

[0143] The status perception module is used to collect operating data of thermal power units, extract equipment operating parameters, set dynamic association rules that change with operating conditions, and construct a unit operating status diagram.

[0144] The simulation module is used to instantiate the unit's operating status diagram, establish a digital twin model, and perform parallel simulations of the thermal power unit and the digital twin model in each control cycle of the thermal power unit, calculating the real-time residual between the simulation output and the actual physical output.

[0145] The model evolution module is used to set evolution instructions based on real-time residuals, collect the operating data of the thermal power unit when the evolution instructions are executed, learn and generate a data causal graph from the operating data, match and calculate it with the unit operating status diagram, and obtain a set of undefined causal relationship edges.

[0146] The control module is used to update the digital twin model based on the causal relationship edge set and to perform peak-shaving control of the thermal power unit in conjunction with the control loop of the thermal power unit.

[0147] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0148] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of the present invention without departing from the spirit and scope of the present invention, and all of these modifications are within the protection scope of the present invention.

Claims

1. A deep peak-shaving control method for thermal power units based on digital twins, characterized in that, include: Collect operating data of thermal power units, extract equipment operating parameters, set dynamic correlation rules that change with operating conditions, and construct a unit operating status diagram; Based on the unit's operating status diagram, an instantiation process is performed to establish a digital twin model. In each control cycle of the thermal power unit, parallel simulations are conducted on the thermal power unit and the digital twin model to calculate the real-time residual between the simulation output and the actual physical output. Based on the real-time residual, the evolution command is set, the operating data of the thermal power unit is collected when the evolution command is executed, the data cause-effect graph is generated from the operating data, and it is matched and calculated with the unit operating status diagram to obtain the set of undefined causal relationship edges. The digital twin model is updated based on the edge set of causal relationships, and the unit's peak-shaving control is carried out in conjunction with the control loop of the thermal power unit.

2. The method for deep peak shaving control of thermal power units based on digital twins according to claim 1, characterized in that, The process of collecting thermal power unit operating data, extracting equipment operating parameters, setting dynamic correlation rules that change with operating conditions, and constructing a unit operating status diagram includes: Real-time data streams from the physical equipment of thermal power units are collected according to a preset data sampling frequency, and spatiotemporal alignment is performed to obtain an operational data sequence. Based on the process design principles of thermal power units, the operating data sequence is used to construct entity nodes representing physical equipment and state nodes representing equipment operating parameters, and to assign static attributes and dynamic real-time attributes. Clustering is used to identify the current operating conditions and operating data sequences of thermal power units under the corresponding historical and real-time data slices. The dynamic correlation strength between state nodes is calculated and quantified as an edge with a correlation function attribute. The correlation function is used to describe the dynamic mathematical relationship between nodes. The unit operation status diagram is constructed based on entity nodes, status nodes, and edges with correlation function attributes. The edges with correlation function attributes are defined as dynamic correlation rules that change with operating conditions.

3. The method for deep peak shaving control of thermal power units based on digital twins according to claim 2, characterized in that, The process involves clustering to identify the current operating conditions and data sequences of the thermal power unit, along with historical and real-time data slices under the corresponding operating conditions. It then calculates the dynamic correlation strength between state nodes and quantifies this strength as edges with correlation function properties. These correlation functions describe the dynamic mathematical relationships between nodes, including: The operating conditions of the running data sequence are identified by online clustering to obtain the current operating condition cluster identifier. Data slices that match the current operating condition cluster identifier are then extracted from historical and real-time data streams. Based on the prior knowledge of the physical topology of the thermal power unit, state node pairs with physical coupling relationships are selected to form a set of node pairs to be evaluated. For each pair of nodes in the set to be evaluated, the correlation strength is calculated based on the corresponding time-series data in the data slice using correlation metrics. The association strength is filtered according to the preset association strength threshold, and the node pairs with the association strength greater than the association strength threshold are judged to have dynamic association; For node pairs determined to have dynamic relationships, their corresponding time series data are used to generate a correlation function describing the quantitative relationship between the two through parameterized fitting, and this function is used as an attribute of the edge connecting the node pairs.

4. The method for deep peak shaving control of thermal power units based on digital twins according to claim 3, characterized in that, The process involves instantiating the unit's operating status diagram to establish a digital twin model. During each control cycle of the thermal power unit, parallel simulations are performed on both the thermal power unit and the digital twin model. The real-time residual between the simulation output and the actual physical output is calculated, including: Based on the current operating conditions, activate the matching dynamic association rules from the unit operating status diagram, record copies of relevant nodes and edge structures, and generate instances of activated subgraphs; Based on the association model encapsulated by each dynamic association rule in the activation subgraph, a digital twin model matching the current control cycle is instantiated and generated. Based on the real-time commands of the thermal power unit control system, parallel state calculations with a preset step size are performed on the physical thermal power unit and the digital twin model to obtain the actual physical output of the thermal power unit and the simulation output of the digital twin model. Vector difference and statistical analysis are performed on the actual physical output of the thermal power unit and the simulation output of the digital twin model to obtain the real-time residual vector representing the multidimensional state differences.

5. The method for deep peak shaving control of thermal power units based on digital twins according to claim 4, characterized in that, The vector difference and statistical analysis of the actual physical output of the thermal power unit and the simulation output of the digital twin model are performed to obtain a real-time residual vector representing multidimensional state differences, including: Based on the sensor measured values ​​of the thermal power unit and the synchronous calculation output values ​​of the digital twin model, the real-time residual vector of the current control cycle is calculated through multi-channel vector difference. Based on the time series formed by the real-time residual vector in the current and past consecutive control cycles, the mean shift, variance and autocorrelation characteristics of each residual component are calculated through sliding window statistics and trend analysis. For the mean shift, variance and autocorrelation characteristics of each residual component, parameter calibration is triggered by a preset decision threshold, and the calibration result is used as a real-time residual vector representing the multidimensional state differences.

6. The method for deep peak shaving control of thermal power units based on digital twins according to claim 5, characterized in that, The process involves setting evolution commands based on real-time residuals, collecting operating data of the thermal power unit during the execution of these commands, learning and generating a data causal graph from the operating data, matching and calculating this graph with the unit's operating status diagram, and obtaining a set of undefined causal relationship edges, including: Evolution instructions are set based on real-time residuals, whereby the evolution instructions encapsulate the associated abnormal residual characteristics, specific state nodes, and the corresponding dynamic association rules in the unit operation status diagram. Collect operating data of thermal power units when evolution commands are executed, and construct a variable set for causal relationship learning by windowing the operating data; By analyzing the conditional independence relationships between variables in a variable set, causal inferences are made, and a data causal graph is constructed. Based on the definitions of the corresponding nodes and edges in the data causal graph and the unit operation status graph, the set of causal relationship edges that exist in the data causal graph but are not defined in the dynamic association rules of the unit operation status graph is identified through subgraph isomorphic matching and difference calculation. The physical rationality of the causal relationship edge set is judged to obtain the filtered causal relationship edge set and its quantified relationship.

7. The method for deep peak shaving control of thermal power units based on digital twins according to claim 6, characterized in that, The step involves identifying a set of causal relationship edges that exist in the data causal graph but are not defined in the dynamic association rules of the unit operation state diagram, based on the definitions of corresponding nodes and edges in the data causal graph and the unit operation state diagram, through subgraph isomorphic matching and difference calculation. This includes: By aligning and mapping nodes semantically, a correspondence between nodes is established between the data cause-effect graph and the unit operation status graph. Traverse the edges in the data cause-effect graph according to the node correspondence relationship, and search for edges in the unit operation status graph that have the same source node and target node, and whose dynamic association rules cover the current operating condition. If the corresponding edge does not exist, the edge in the data cause-effect graph is marked as a candidate causal relationship edge. If the corresponding edge exists, the causal strength and direction of the edge in the data cause-effect graph are compared with the corresponding edge in the unit operation status graph. The significance of causal strength of candidate causal relationship edges in the data causal graph is statistically analyzed, and a comprehensive score is obtained by combining the proximity of the source node and the target node in the physical thermal power unit structure, and a set of candidate causal relationship edges is established. The candidate causal relationship edge set is integrated with the candidate conflicting causal relationship edge set to obtain the causal relationship edge set.

8. The method for deep peak shaving control of thermal power units based on digital twins according to claim 7, characterized in that, The step involves statistically analyzing the causal strength of candidate causal edges in the data causal graph, and combining this with a comprehensive score based on the proximity of the source node and the target node within the physical structure of the thermal power unit. This results in the establishment of a set of candidate causal edges, including: Calculate the topological score and statistical score of the candidate causal relationship edges based on the statistical significance of the centrality and causal strength of the connecting nodes in the data causal graph. The physical proximity score is calculated based on the distance between the source node and the target node of the candidate causal relationship edge in the physical topology of the thermal power unit. A comprehensive score is calculated based on topological score, statistical score, and physical proximity score, and a set of candidate causal edge pairs is obtained by combining the score with a preset screening threshold.

9. A deep peak-shaving control method for thermal power units based on digital twins according to claim 8, characterized in that, The process of updating the digital twin model based on the causal relationship edge set and combining it with the control loop of the thermal power unit for peak shaving control includes: Based on the new causal relationship edges and their quantification relationships, the state equation and input-output mapping relationship of the digital twin model are reconstructed to generate a digital twin model with updated structure. The updated digital twin model is deployed as a new digital twin model, dynamically instantiated, and reconnected to the parallel simulation process; Based on the real-time peak-shaving instructions issued by the power generation control group of the power grid to which the thermal power unit is connected, a control instruction sequence is generated through the simulation results of the new digital twin model; The control command sequence is optimized in the rolling time domain, and peak-shaving control of thermal power units is performed based on the optimization results.

10. A deep peak-shaving control system for thermal power units based on digital twins, used to implement the deep peak-shaving control method for thermal power units based on digital twins as described in any one of claims 1-9, characterized in that, It includes a state awareness module, a simulation module, a model evolution module, and a control module: The status perception module is used to collect operating data of thermal power units, extract equipment operating parameters, set dynamic association rules that change with operating conditions, and construct a unit operating status diagram. The simulation module is used to instantiate the unit's operating status diagram, establish a digital twin model, and perform parallel simulation of the thermal power unit and the digital twin model in each control cycle of the thermal power unit, and calculate the real-time residual between the simulation output and the actual physical output. The model evolution module is used to set evolution instructions based on real-time residuals, collect the operating data of the thermal power unit when the evolution instructions are executed, learn and generate a data causal graph from the operating data, match and calculate it with the unit operating status diagram, and obtain a set of undefined causal relationship edges. The control module is used to update the digital twin model based on the causal relationship edge set and to perform peak-shaving control of the thermal power unit in conjunction with the control loop of the thermal power unit.

Images