Agent-based optimization method and system for starting control of a circulating unit

Abstract

This application relates to the field of cyclic unit control optimization technology, and discloses an agent-based cyclic unit startup control optimization method and system, including: analyzing historical startup situations based on pre-acquired cyclic unit status data and environmental parameters to construct a startup case library; responding to the cyclic unit startup command, analyzing parameter changes during startup based on the current cyclic unit status to generate a first startup curve; configuring a multi-agent collaborative mechanism, decomposing the first startup curve to obtain multiple control tasks, analyzing the control capabilities and real-time status of each executing agent through a central agent, matching at least one control task to each executing agent to obtain a first control strategy; controlling the startup process of the cyclic unit, analyzing the deviation between the real-time operating status and the first startup curve, and dynamically adjusting the control parameters of the agents to optimize the first control strategy; this application can adaptively adjust the startup strategy according to the unit status, improving the overall performance of the unit.

Description

Technical Field

[0001] This application relates to the field of circulating unit control optimization technology, and more specifically to agent-based circulating unit start-up control optimization method and system. Background Technology

[0002] Currently, the start-up control of gas-steam combined cycle units largely relies on fixed templates or manual experience, resulting in problems such as rigid start-up curves, insufficient subsystem coordination, and inadequate utilization of historical data. Although traditional DCS systems can achieve sequential control, each control loop operates independently, lacking multi-variable collaborative optimization capabilities, and exhibiting delayed response to real-time deviations, leading to long start-up times, high energy consumption, and significant safety risks.

[0003] The existing technology has the following problems: the allocation of control tasks for each subsystem during startup is rigid, lacks coordination and adaptive capabilities, and is difficult to dynamically adjust according to real-time unit status, environmental parameters, and historical best practices; the allocation of control tasks is static, and it is impossible to dynamically schedule control resources according to real-time equipment status, load changes, and boundary conditions, and it lacks coordination and conflict resolution mechanisms; historical startup data has not been systematically mined, and a reusable knowledge base cannot be formed; in order to solve at least one of the above problems, this application proposes an agent-based cyclic unit startup control optimization method and system. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this application is to provide an agent-based optimization method and system for cyclic unit start-up control, which can effectively solve the problems in the background technology. The specific technical solution of this application is as follows:

[0005] Agent-based optimization methods for cyclic unit start-up control include:

[0006] Based on the pre-acquired circulating unit status data and environmental parameters, analyze historical startup scenarios and build a startup case library;

[0007] In response to the start-up command of the circulating unit, based on the current status of the circulating unit, at least one historical case is matched in the start-up case library, the parameter changes during the start-up process are analyzed, and the first start-up curve is generated.

[0008] A multi-agent collaborative mechanism is configured, the first start-up curve is decomposed to obtain multiple control tasks, the control capabilities and real-time status of each executing agent are analyzed by the central agent, and at least one control task is matched for each executing agent to obtain the first control strategy. The agents include the central agent and multiple executing agents.

[0009] The startup process of the circulating unit is controlled according to the first control strategy. Real-time operating data is collected, the deviation between the real-time operating status and the first startup curve is analyzed, and the control parameters of the intelligent agent are dynamically adjusted to optimize the first control strategy and optimize the startup process of the circulating unit.

[0010] Specifically, the step of analyzing historical startup data and environmental parameters based on pre-acquired circulating unit status data and constructing a startup case library includes:

[0011] Based on the pre-acquired cyclic unit status data and environmental parameters, the start and end times of each historical startup process are identified, and multiple startup segments are obtained to construct the first startup segment set.

[0012] According to the preset startup success criteria, the startup process of each startup segment in the first startup segment set is analyzed, and the startup segments that start successfully are selected to obtain the second startup segment set.

[0013] Analyze the boundary condition features of each startup fragment in the second startup fragment set, cluster the boundary condition features, and construct a startup case library.

[0014] Specifically, in response to the circulating unit start-up command, based on the current circulating unit status, at least one historical case is matched in the start-up case library, the parameter changes during the start-up process are analyzed, and a first start-up curve is generated, including:

[0015] In response to the cyclic unit start-up command, the features of the current start-up task are extracted based on the current cyclic unit status, and a feature vector is constructed.

[0016] Based on the feature vector, at least one historical case is matched in the startup case library to construct a first case set;

[0017] Analyze the case relationships and parameter changes during the case startup process in the first case set to generate the first startup curve.

[0018] Specifically, the analysis of case associations and parameter changes during case initiation within the first case set, generating a first initiation curve, includes:

[0019] Extract the static and dynamic parameters corresponding to each case in the first case set, and construct the case feature vector;

[0020] Calculate the similarity between case feature vectors, and then weight and fuse the parameters of cases with similarity greater than a preset similarity threshold to obtain a second case set;

[0021] The case with the highest similarity between the feature vectors of the cases in the second case set is selected as the base case. The difference points between each case in the second case set and the base case with parameter differences greater than a preset difference threshold are identified. The difference points are then corrected by interpolation to obtain the third case set.

[0022] The parameter points corresponding to each case in the third case set are merged to generate the first startup curve.

[0023] Specifically, the multi-agent cooperative mechanism includes:

[0024] Multiple control tasks are obtained by decomposing the first start-up curve through the central intelligent agent. The parameter coupling in the control tasks is analyzed, the parameters are aggregated, and a set of control tasks is constructed.

[0025] The central agent analyzes the control capabilities and real-time status of each executing agent, matches each control task in the control task set with an executing agent, and each executing agent matches at least one control task to obtain the first control strategy.

[0026] Specifically, the process involves decomposing the first startup curve through a central intelligent agent to obtain multiple control tasks, analyzing the parameter coupling within these control tasks, aggregating the parameters, and constructing a control task set, including:

[0027] Multiple control tasks are obtained by decomposing the first start-up curve through the central intelligent agent. The slope of parameter change in each control task is analyzed, and the coupling strength of parameters at adjacent time nodes is calculated.

[0028] Parameters with coupling strength greater than a preset coupling threshold are aggregated to obtain a parameter linkage set. By analyzing the number of parameters and the parameter coupling complexity, the parameter linkage set is split to obtain the first subtask group.

[0029] Analyze the phase lag and amplitude attenuation between parameters in the parameter linkage set, select key parameters as independent control parameter nodes, and obtain the second subtask group;

[0030] By combining the first subtask group and the second subtask group, a set of control tasks is obtained.

[0031] Specifically, the process involves analyzing the control capabilities and real-time status of each executing agent through a central agent, matching each control task in the control task set with an executing agent, and matching each executing agent with at least one control task to obtain a first control strategy, including:

[0032] The central agent analyzes the control capabilities and real-time status of each executing agent, calculates the corresponding capability score, and analyzes the priority of each control task in the control task set.

[0033] According to priority order, each control task is matched with an execution agent that meets the corresponding task parameters. Execution agents not assigned to control tasks are matched with control tasks of corresponding priority from high to low according to their ability scores, and the first matching result is obtained.

[0034] The task pairs with timing conflicts and resource competition in the first matching result are identified. The task with the lower priority in the task pair is rematched by the execution agent to obtain the second matching result.

[0035] Based on the second matching result, the control task of each executing agent is determined, and the first control strategy is obtained.

[0036] Specifically, the step of controlling the start-up process of the circulating unit according to the first control strategy, collecting real-time operating data, analyzing the deviation between the real-time operating status and the first start-up curve, and dynamically adjusting the control parameters of the intelligent agent to optimize the first control strategy includes:

[0037] The startup process of the circulating unit is controlled according to the first control strategy. Real-time operating data is collected, and the deviation between the real-time operating status and the first startup curve is analyzed to obtain the status deviation.

[0038] The deviation level is determined based on the state deviation, and the control parameters of the agent are dynamically adjusted according to the deviation level to optimize the first control strategy.

[0039] Specifically, the step of determining the deviation level based on the state deviation, dynamically adjusting the control parameters of the agent according to the deviation level, and optimizing the first control strategy includes:

[0040] The deviation level is determined based on the state deviation, and the corresponding optimization parameters are matched in the preset parameter optimization library according to the deviation level to obtain the set of optimization parameters;

[0041] The control parameters of the agent are dynamically adjusted according to the set of optimized parameters to optimize the first control strategy.

[0042] An agent-based cyclic unit start-up control optimization system is used to implement the agent-based cyclic unit start-up control optimization method, including:

[0043] The startup case library construction module analyzes historical startup data and environmental parameters based on the pre-acquired cyclic unit status data and constructs a startup case library.

[0044] The startup curve analysis module responds to the start-up command of the circulating unit, matches at least one historical case in the startup case library based on the current status of the circulating unit, analyzes the parameter changes during the startup process, and generates the first startup curve.

[0045] The multi-agent collaboration module is configured with a multi-agent collaboration mechanism. It decomposes the first start-up curve to obtain multiple control tasks. The central agent analyzes the control capabilities and real-time status of each executing agent and matches at least one control task to each executing agent to obtain the first control strategy. The agents include the central agent and multiple executing agents.

[0046] The control optimization module controls the startup process of the circulating unit according to the first control strategy, collects real-time operating data, analyzes the deviation between the real-time operating status and the first startup curve, dynamically adjusts the control parameters of the intelligent agent to optimize the first control strategy, and optimizes the startup process of the circulating unit.

[0047] The beneficial effects of this application are as follows: A case library is built based on historical successful startup segments. Through feature matching, case fusion, and interpolation correction, a dynamic startup curve adapted to real-time operating conditions is generated. Tasks are dynamically decomposed by a central intelligent agent, and optimized matching is performed by combining the capabilities and states of the executing intelligent agents. The coupling strength, phase, and amplitude relationships between parameters are identified, and control tasks are intelligently aggregated or split to improve control accuracy and stability. Startup deviations are monitored online, and intelligent agent control parameters are adaptively adjusted according to the deviation level to achieve closed-loop optimization. By optimizing the startup curve and task scheduling, startup time is shortened, gas consumption and auxiliary power consumption are reduced, and the startup strategy can be adaptively adjusted according to changes in unit health status, environmental conditions, fuel characteristics, etc., thereby improving the overall performance of the unit. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the agent-based cyclic unit start-up control optimization method in the embodiments of this application.

[0049] Figure 2 This is a flowchart illustrating the first startup curve analysis process in an embodiment of this application.

[0050] Figure 3 This is a schematic diagram illustrating the collaborative process between the central intelligent agent and the executive intelligent agent in an embodiment of this application;

[0051] Figure 4 This is a schematic diagram of the agent-based cyclic unit start-up control optimization system in the embodiments of this application. Detailed Implementation

[0052] The present application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0053] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0054] Hereinafter, the terms "first," "second," and other generic terms are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0055] refer to Figure 1 The image shows a specific implementation of the agent-based cyclic unit start-up control optimization method of this application, including:

[0056] S101. Analyze historical startup data and environmental parameters based on the pre-acquired circulating unit status data and construct a startup case library;

[0057] S102. In response to the cycle unit start-up command, based on the current cycle unit status, match at least one historical case in the start-up case library, analyze the parameter changes during the start-up process, and generate the first start-up curve.

[0058] S103. Configure a multi-agent collaborative mechanism, decompose the first start-up curve to obtain multiple control tasks, analyze the control capabilities and real-time status of each executing agent through the central agent, match at least one control task for each executing agent, and obtain the first control strategy. The agents include the central agent and multiple executing agents.

[0059] S104. Control the startup process of the circulating unit according to the first control strategy, collect real-time operating data, analyze the deviation between the real-time operating status and the first startup curve, dynamically adjust the control parameters of the intelligent agent to optimize the first control strategy, and optimize the startup process of the circulating unit.

[0060] In this embodiment, time-series data of the entire startup process is extracted from the power plant's real-time database and performance calculation module. This includes, but is not limited to, key system parameters and corresponding environmental parameters. Key system parameters include, but are not limited to, gas turbine speed, exhaust temperature, steam drum pressure, and vibration values. Environmental parameters include, but are not limited to, atmospheric temperature and pressure. Based on the pre-acquired circulating unit status data and environmental parameters, historical startup situations are analyzed to detect the time interval between the unit load rising from zero to the grid-connected load. The start and end times of each startup are automatically identified, and independent startup segments are defined. Based on preset startup success criteria, startup data with abnormalities or failures are eliminated, and a startup case library is constructed. By constructing the startup case library, the scattered and unstructured historical operating data is transformed into a systematic and searchable startup knowledge base, which allows for rapid matching of corresponding case references. Data filtering ensures the quality of cases in the case library, providing accurate data support for subsequent intelligent startup. It also allows for setting corresponding startup control strategies by referring to historical best cases, improving the efficiency and effectiveness of strategy formulation.

[0061] Specifically, in response to the circulating unit start-up command, based on the current circulating unit status, including but not limited to metal temperature and downtime, the similarity between the current unit status characteristics and the boundary condition characteristics of each case in the start-up case library is calculated. At least one historical case is matched in the start-up case library, and the parameter sequences of the historical cases are weighted and fused. The case most similar to the current state is selected as the base case. The differences between other cases and the base case are identified, and the differences are corrected using an interpolation algorithm to generate the first start-up curve. By dynamically generating the first start-up curve, it can adapt to the dynamic changes of the circulating unit and improve environmental adaptability. Through multi-case fusion and interpolation correction, the resulting first start-up curve incorporates the inherent laws of historical successful experience and adapts to the current real-time operating conditions, providing the unit with a safe, efficient, and achievable start-up target path, thus improving the scientific nature and effectiveness of the start-up process.

[0062] Furthermore, a multi-agent collaborative mechanism is configured, comprising a central agent and multiple executing agents. The central agent decomposes the first startup curve in terms of timing and system dimensions, analyzes the parameter coupling relationships within and between each control task, quantifies the coupling strength between parameters, filters composite and independent tasks to obtain a set of control tasks, analyzes the control capability status of each executing agent, calculates the corresponding capability score, analyzes the physical timing requirements and safety priorities of the control tasks, sets execution priorities for each task, and matches tasks to the corresponding executing agents according to priority order to obtain the first control strategy. Optimizing the allocation of control resources and tasks can avoid the problem of poor coordination in centralized control. Through the collaboration and conflict resolution of multiple agents, the orderliness and coordination of the startup process can be ensured, improving the overall control performance and effectiveness of the startup process.

[0063] Specifically, the startup process of the circulating unit is controlled according to the first control strategy. Real-time operating data is collected, and the central intelligent agent compares the real-time values ​​of the parameters with the corresponding expected values ​​in the first startup curve. It analyzes the deviation between the real-time operating state and the first startup curve, determines the deviation level based on the calculated magnitude and trend of the deviation, and dynamically adjusts the control parameters of the intelligent agent to optimize the first control strategy, thereby optimizing the startup process of the circulating unit. By sensing deviations in real time and dynamically adjusting the control strategy, disturbances can be actively suppressed and nonlinearities compensated, ensuring that the startup process proceeds smoothly and accurately along the optimized path even under uncertain environmental conditions, thus improving the safety, speed, and economy of the startup control process.

[0064] This application constructs a case library based on historical successful startup segments. Through feature matching, case fusion, and interpolation correction, it generates dynamic startup curves adapted to real-time operating conditions. The central intelligent agent dynamically decomposes tasks and optimizes matching by combining the capabilities and states of the executing intelligent agents. It identifies the coupling strength, phase, and amplitude relationships between parameters, intelligently aggregates or splits control tasks, improves control accuracy and stability, monitors startup deviations online, and adaptively adjusts the intelligent agent control parameters according to the deviation level to achieve closed-loop optimization. By optimizing the startup curve and task scheduling, it shortens startup time, reduces gas consumption and auxiliary machine power consumption, and can adaptively adjust startup strategies according to changes in unit health status, environmental conditions, and fuel characteristics, thereby improving the overall performance of the unit.

[0065] Furthermore, based on the pre-acquired circulating unit status data and environmental parameters, historical startup scenarios are analyzed to construct a startup case library, including:

[0066] S201. Based on the pre-acquired circulating unit status data and environmental parameters, identify the start and end times of each historical startup process, divide it into multiple startup segments, and construct the first startup segment set.

[0067] S202. According to the preset startup success judgment criteria, analyze the startup process of each startup segment in the first startup segment set, filter out the startup segments that have started successfully, and obtain the second startup segment set.

[0068] S203. Analyze the boundary condition features of each startup fragment in the second startup fragment set, cluster the boundary condition features, and construct a startup case library.

[0069] In this embodiment, long-term, time-synchronized circulating unit status data and environmental parameters are pre-acquired from the power plant's real-time historical database. The circulating unit status data mainly includes parameters directly reflecting the unit's operating stage, including but not limited to gas turbine speed, generator active power, unit load command, and turbine steam pressure. Environmental parameters include but are not limited to atmospheric temperature, atmospheric pressure, and relative humidity. Based on the start-up control accuracy requirements, a persistence criterion for start and end time points is set to identify the start and end time points where the unit's operating state, as represented by key parameters, undergoes a step change. For example, the start and end time points where the unit load command changes from zero to a positive value are selected. The initial moment when the gas turbine speed starts to rise continuously from the turning gear speed is taken as the start-up time point; the time point when the generator grid connection signal is stable and the unit's active power first reaches and stabilizes at more than 10% of the rated load for more than 5 minutes is selected as the start-up end time point. Historical data is traversed to detect all time point pairs that meet the start-up time and start-up end time, and the continuous data stream is divided into multiple independent start-up segments; each start-up segment contains a complete time sequence of all relevant state parameters and environmental parameters from the start to the end of the segment. All independent start-up segments are integrated to form the first start-up segment set.

[0070] It should be noted that by using specific time points for filtering and state recognition based on multi-parameter linkage, complete startup segments can be quickly identified from the startup process of massive historical data. This provides accurate and complete data support for subsequent batch analysis and case construction, ensuring the integrity of each startup segment and improving the accuracy of the analysis results.

[0071] Specifically, based on the operating procedures, safety guidelines, and expert experience of the circulating unit, multi-dimensional start-up success criteria are established, including a rule set of multiple sub-rules logically combined. For example, analysis is performed to determine process safety if no process protection signals causing tripping are generated within the entire start-up segment's time range; analysis is performed to determine parameter compliance if key operating parameters do not exceed their safe operating limits; and analysis is performed to determine process integrity if the total start-up time is within a reasonable range. The reasonable range can be set according to the start-up control accuracy requirements. For each start-up segment in the first start-up segment set, the start-up success criteria are used for judgment. When the segment data record meets all preset success judgment rules, the segment is judged as a successfully started start-up segment. By filtering out the successfully started start-up segments, a second start-up segment set is obtained, eliminating failure data caused by fault interruption, operational abnormalities, or poor performance.

[0072] It should be noted that by setting strict and multi-dimensional startup success criteria, historical startup data can be effectively cleaned and purified, ensuring that the data sources used to build the subsequent case library are all derived from successful operational practices, eliminating interference from failed or poor cases, improving the safety and effectiveness of subsequent startup curve generation, and enhancing the effectiveness and reliability of the startup control process.

[0073] For each successful startup segment in the second startup segment set, features at the startup initiation time are extracted to construct a boundary condition feature vector. These features reflect the startup initiation scenario, including but not limited to ambient temperature, atmospheric pressure, gas turbine metal temperature, unit shutdown duration, turbine upper and lower cylinder temperature difference, and boiler drum pressure. Each startup segment is treated as a feature point in a multi-dimensional feature space. These feature points are clustered using the K-means clustering algorithm. Based on the iterative calculation of the Euclidean distance between feature points, multiple clusters are obtained. After clustering, all successful startup segments within each cluster are considered to be similar cases with similar boundary conditions. Each category is defined as a case category, and the feature center of that category is used as the representative of the case category. By combining multiple case category representatives, a startup case library is constructed.

[0074] It should be noted that by clustering the initiation conditions, scattered initiation fragments can be organized into case types with corresponding categories. When faced with a new initiation task, the system can quickly locate the most similar case category based on the current actual boundary conditions by calculating the feature distance with each case category. This improves the efficiency and accuracy of historical experience matching, provides case references for formulating corresponding initiation curves, and enhances the generalization and adaptability of the case matching process and the initiation curve formulation process.

[0075] Furthermore, in response to the circulating unit start-up command, based on the current circulating unit status, at least one historical case is matched in the start-up case library, the parameter changes during the start-up process are analyzed, and a first start-up curve is generated, including:

[0076] S301. In response to the cyclic unit start-up command, extract the features of the current start-up task based on the current cyclic unit status and construct a feature vector;

[0077] S302. Based on the feature vector, match at least one historical case in the startup case library to construct a first case set;

[0078] S303. Analyze the case associations and parameter changes during the case startup process in the first case set, and generate the first startup curve.

[0079] In this embodiment, in response to the operator's command to start the cyclic unit, parameter values ​​that can comprehensively characterize the initial conditions of the start-up task are obtained from the real-time database and the unit monitoring system. The parameter selection process is aligned with the boundary condition feature dimensions used when constructing the start-up case library, improving the fairness and accuracy of the matching process. For example, the current status collected by the system includes the current ambient temperature, current atmospheric pressure, current temperature of key metal components of the gas turbine, continuous shutdown duration from the last shutdown to the current moment, current temperature difference between the upper and lower cylinders of the high-pressure cylinder of the steam turbine, and current pressure of the boiler drum. The collected parameter values ​​are integrated in the same order as the boundary condition features to obtain a feature vector.

[0080] It should be noted that by standardizing and structuring feature extraction, the complex and multidimensional real-time status of the scene is transformed into corresponding feature vectors, providing accurate data support for the subsequent case matching process and quantifying the differences between the current task and historical tasks in terms of initial conditions.

[0081] Specifically, based on the feature vector, similarity calculation is performed between the current feature vector and the boundary condition feature vector of each historical case pre-stored in the startup case library. The similarity calculation includes, but is not limited to, Euclidean distance and cosine similarity. For Euclidean distance, the Euclidean distance between the current feature vector and the boundary condition feature vector of each historical case is calculated. The smaller the distance value, the more similar the two are in terms of boundary conditions such as environment and initial state. According to the unit startup control accuracy requirements, a distance threshold is set, and historical cases with Euclidean distance greater than the distance threshold are selected as the first case set. This includes all the information of several historical successful cases that are closest to the current startup in terms of initial conditions, including boundary condition features and the complete startup process parameter sequence of the case that changes over time.

[0082] It should be noted that by calculating feature similarity for case screening, the most relevant successful cases to the current scenario can be quickly located from historical data. This avoids blindly or randomly referencing historical data and ensures that the cases used to generate the subsequent startup curve are highly relevant and referable to the current state. This enhances the reference value of the selected cases and improves the pertinence and effectiveness of the startup decision-making process.

[0083] Specifically, the analysis of case associations and parameter changes during case initiation within the first case set generates a first initiation curve. By deeply fusing data from multiple similar cases and intelligently correcting based on differences, the randomness and inaccuracy inherent in directly applying single historical cases are avoided. Data fusion enhances the robustness and adaptability of the initiation curve. Through interpolation correction, the generated initiation curve can adapt to the differences between current boundary conditions and historical average conditions. Combining commonalities from historical successes with current operating conditions yields the first initiation curve, providing an accurate and reliable setpoint trajectory for subsequent intelligent control decision-making processes.

[0084] like Figure 2 As shown, the analysis examines the case associations and parameter changes during the case initiation process in the first case set, generating the first initiation curve, including:

[0085] S401. Extract the static and dynamic parameters corresponding to each case in the first case set, and construct the case feature vector;

[0086] S402. Calculate the similarity between case feature vectors, and perform weighted fusion on the parameters corresponding to cases with similarity greater than a preset similarity threshold to obtain a second case set;

[0087] S403. Select the case with the highest similarity between the case feature vectors in the second case set as the base case, identify the difference points between each case in the second case set and the base case whose parameter difference is greater than the preset difference threshold, and perform interpolation correction on the difference points to obtain the third case set.

[0088] S404. Merge the parameter points corresponding to each case in the third case set to generate the first startup curve.

[0089] In this embodiment, for each case in the first case set, static parameters and dynamic parameters are extracted. Static parameters represent the boundary condition state values ​​near the zero point of time at the start of the case, including but not limited to ambient temperature, atmospheric pressure, gas turbine metal temperature, and shutdown duration at startup. Dynamic parameters represent the statistical situation of the change sequence of key controlled variables or state variables throughout the startup process, reflecting the process behavior pattern, including but not limited to extracting the average rate of increase, maximum rate of increase, and duration of the warm-up speed plateau from the gas turbine rate of increase curve of the case. The extracted static parameter values ​​and dynamic feature values ​​are combined according to a predefined order and structure to obtain the case feature vector. By extracting static and dynamic parameters, the similarity comparison of cases is transformed from matching a single initial condition to considering the startup process behavior pattern. This allows for the identification of successful cases with slightly different initial conditions but highly similar startup process control strategies and response characteristics, providing more reliable and accurate data support for subsequent case fusion and avoiding missing high-quality reference cases due to minor differences in initial conditions.

[0090] Specifically, the cosine similarity between each pair of feature vectors of all cases is calculated. A similarity threshold is set based on the startup control precision requirements, and cases with cosine similarities exceeding this threshold are classified as highly similar case groups. For each such case group, the original, complete time-series data for each case within the group is weighted and averaged at the same timestamp. The weights can be set as the similarity between each case and the initial feature vector of the current task, or equal weights can be used. This results in a fused curve and related features, forming a new case. This new case replaces the corresponding highly similar case, resulting in a second case set. The fused high-quality case represents a group of highly similar cases. By fusing data from highly similar cases, random fluctuations or measurement noise in individual historical cases can be effectively smoothed, extracting more stable and fundamental startup process patterns under this similarity model. This reduces the number of cases that need to be processed subsequently, lowers computational complexity, and improves the quality and representativeness of the reference cases, providing accurate data support for generating smooth and reliable startup curves.

[0091] Furthermore, from the second case set, the case with the highest similarity to the feature vector constructed in S301 is selected as the base case. This case represents the historical success pattern closest to the current operating conditions. Each other case in the second case set is compared with the base case, comparing the values ​​of the same parameters at the same time stamp along the time axis. When the difference between a parameter value of a case at a certain time point and the corresponding value of the base case exceeds a preset difference threshold, this point is marked as a difference point. The difference threshold can be set according to the startup control accuracy requirements, including but not limited to a temperature difference exceeding 5 degrees Celsius. Difference points reflect differences in process response caused by different initial conditions. The difference points are processed through interpolation correction, replacing the corresponding difference point value with the mean of adjacent values ​​to correct the difference point value. The corrected value then replaces the corresponding case in the second case set, resulting in the third case set. By identifying and correcting difference points, the error caused by directly applying historical curves can be reduced. Interpolation correction combines historical success patterns and adapts to startup conditions to obtain a fused startup curve, improving the applicability and accuracy of the generated curve.

[0092] Specifically, the time-series data of all cases in the third case set are aligned on the same time coordinate system. For each discrete time point in the startup process, the values ​​of all cases for the same parameter at that time point are compiled into a set. Weights are assigned to the corresponding parameters based on the final comprehensive similarity between each case and the current task. The parameter values ​​in the set are then weighted and synthesized to obtain the final target value of that parameter at that time point. These final target values ​​are then integrated chronologically to obtain a complete sequence of target values ​​in both time and parameter dimensions, serving as the first startup curve. This provides optimized setpoints for each key parameter at every moment from startup to completion for subsequent control processes. By fusing multiple cases point-by-point in the time series, information from various calibrated successful paths can be integrated to obtain the optimal target value under the current conditions. Parameter fusion fully utilizes collective experience, improving the robustness and accuracy of the startup curve, and providing an accurate setpoint benchmark for the safe, rapid, and economical automatic startup of the unit.

[0093] Furthermore, multi-agent cooperative mechanisms include:

[0094] S501. Multiple control tasks are obtained by decomposing the first start-up curve through the central intelligent agent, the parameter coupling in the control tasks is analyzed, the parameters are aggregated, and a set of control tasks is constructed.

[0095] S502. By analyzing the control capabilities and real-time status of each executing agent through the central agent, an executing agent is matched for each control task in the control task set, and each executing agent is matched with at least one control task to obtain the first control strategy.

[0096] In this embodiment, multiple control tasks are obtained by decomposing the first startup curve through a central intelligent agent. Along the time axis, the parameter coupling in the control tasks is analyzed and the parameters are aggregated to construct a set of control tasks. By deconstructing the startup curve and performing parameter coupling analysis, the complex multivariate coordinated control problem is decomposed into independent or internally strongly coupled subsystem control tasks, determining the control boundaries and interaction relationships. This provides accurate data support for subsequent efficient and conflict-free task allocation, avoids oscillation problems caused by the lack of consideration for coupling between control loops, and improves the rationality and stability of the entire control process.

[0097] like Figure 3 As shown, the central agent analyzes the control capabilities and real-time status of each executing agent. Executing agents include, but are not limited to, gas turbine controller agents, waste heat boiler feedwater controller agents, and auxiliary machine sequence controller agents. Each control task in the control task set is matched with an executing agent, and each executing agent is matched with at least one control task, resulting in a first control strategy. Dynamic matching based on the real-time requirements of the tasks and the real-time status of the executing agents fully utilizes the capabilities of each agent, balancing the system load. When the performance of a single executing agent degrades or fails, tasks can be reallocated through the central agent, enhancing the system's reliability and robustness. The resulting first control strategy is consistent, resource allocation is reasonable, and runtime conflicts are avoided, ensuring the orderly and efficient startup process.

[0098] Furthermore, by decomposing the first startup curve through the central intelligent agent, multiple control tasks are obtained. The parameter coupling in the control tasks is analyzed, the parameters are aggregated, and a set of control tasks is constructed, including:

[0099] S601. The first start-up curve is decomposed by the central intelligent agent to obtain multiple control tasks. The slope of parameter change in each control task is analyzed, and the coupling strength of parameters at adjacent time nodes is calculated.

[0100] S602. Aggregate the parameters whose coupling strength is greater than the preset coupling threshold to obtain the parameter linkage set. By analyzing the number of parameters and the parameter coupling complexity, split the parameter linkage set to obtain the first subtask group.

[0101] S603. Analyze the phase lag and amplitude attenuation between parameters in the parameter linkage set, select key parameters as independent control parameter nodes, and obtain the second subtask group.

[0102] S604. Combining the first subtask group and the second subtask group, we obtain the control task set.

[0103] In this embodiment, a time interval is set according to the required start-up control accuracy. According to the time interval, the central agent discretely samples the first start-up curve on the time axis to obtain discrete parameter value data points. For each controlled parameter, the instantaneous rate of change at each sampling time point is calculated using the central difference method as the slope, reflecting the dynamic change trend of each parameter during the start-up process. A pre-trained linear correlation model using a large amount of historical operating data is used to analyze the impact of parameter changes on parameter changes at the next time point, calculating and outputting an estimated influence coefficient to reflect the coupling strength between parameters. The corresponding coupling strength is calculated for all parameter pairs to be examined, resulting in a coupling strength matrix that quantifies the tightness of the dynamic response between any two parameters under the corresponding start-up curve. By analyzing the dynamic correlation of parameter change rates, the real-time interaction strength between parameters during the start-up dynamic process can be accurately captured, providing accurate data support for parameters requiring coordinated action and independent control. This avoids the limitations of relying solely on experience or static models to judge coupling relationships, improving the accuracy of the analysis results.

[0104] Specifically, a coupling threshold is set according to the startup control precision requirements. The coupling strength matrix is ​​traversed to identify parameter pairs whose coupling strength is greater than the coupling threshold. Based on these strong connections, these parameter nodes are divided into one or more parameter linkage sets through clustering. There is strong coupling between parameters within each parameter linkage set, which requires joint control. A parameter linkage set contains many parameters, and the coupling relationships between parameters are relatively complex. The parameter coupling complexity of the set is analyzed by calculating the average connectivity of the parameter nodes. A complexity threshold is set according to the startup control precision requirements. If the parameter coupling complexity exceeds the complexity threshold, the central agent identifies the position with the minimum coupling strength within the set based on the coupling strength, splits the set into subsets, and each subset serves as a control subtask. The set that does not need to be split and the split subsets are integrated to obtain the first subtask group.

[0105] It should be noted that by analyzing the coupling strength and coupling complexity of parameters and intelligently decomposing the parameters, we can avoid creating control tasks with too many variables and overly complex internal interactions. Complex control tasks are difficult to design and tune in actual control and are prone to instability. Through reasonable decomposition, we obtain multiple control subtasks of appropriate size and high internal cohesion. Each subtask can be efficiently processed by an execution agent, which can clarify the interaction between subtasks and improve the control effectiveness and stability of the control system.

[0106] Specifically, for each parameter linkage set, the central agent analyzes the phase lag and amplitude decay between parameters within the set. By analyzing the cross-correlation function between parameters in historical data, it analyzes the delay time of the gas turbine exhaust temperature response after the fuel flow command changes, as well as the ratio of the response amplitude to the command change amplitude. It identifies parameter nodes with phase lag much greater than other parameters or abnormal amplitude transfer function gain. These parameters are not suitable to be simply bundled with other parameters in a fast coordinated control process. The corresponding parameters are treated as independent control parameter nodes, corresponding to safety constraints that require special attention or slow processes with unique dynamic characteristics. The extracted independent control parameter nodes are treated as independent control tasks, forming the second subtask group.

[0107] It should be noted that by identifying key parameters with significant lag and abnormal gain and controlling them independently, the ability to protect the important safety and performance boundaries of the unit can be enhanced. Separating these parameters from rapid collaborative tasks allows for the design of dedicated and more robust control processes, avoiding control conflicts or instability caused by bundling them with the fast-response main control loop. By refining and optimizing control tasks, the overall startup performance of the system can be improved, as well as the control accuracy and effectiveness of key safety parameters can be enhanced.

[0108] Specifically, the first and second subtask groups are merged. During the merging process, the central agent checks and ensures that there are no duplicate parameter assignments between the two subtask groups, and that each controlled parameter belongs to exactly one control task, resulting in a set of control tasks. Through task integration, it is ensured that all control requirements in the first startup curve are decomposed and mapped to specific executable tasks. This provides accurate data support for the central agent to perform efficient and reasonable task matching based on the expertise and status of each executing agent, thereby improving the effectiveness of the multi-agent collaborative process.

[0109] Furthermore, by analyzing the control capabilities and real-time state of each executing agent through the central agent, an executing agent is matched for each control task in the control task set, and each executing agent is matched with at least one control task, thus obtaining the first control strategy, including:

[0110] S701. Analyze the control capabilities and real-time status of each executing agent through the central agent, calculate the corresponding capability score, and analyze the priority of each control task in the control task set.

[0111] S702. According to the priority order, match the execution agent that meets the corresponding task parameters for each control task. The execution agents that are not assigned to control tasks are matched to the corresponding control tasks with the highest to lowest priority according to the ability score from high to low, and the first matching result is obtained.

[0112] S703. Identify task pairs in the first matching result that have timing conflicts and resource competition, and re-match the low-priority task in the task pair with the execution agent to obtain the second matching result.

[0113] S704. Determine the control task of each executing agent based on the second matching result to obtain the first control strategy.

[0114] In this embodiment, the central agent analyzes the control capabilities and real-time status of each executing agent through a pre-defined communication interface. Control capability refers to the inherent, relatively static attributes of the agent, including but not limited to the types of physical quantities it can control, the types supported by the control algorithm library, and the list of specific actuator devices it can access and drive. Real-time status includes dynamic information, including but not limited to the current computing resource utilization rate and the health status of underlying devices. The central agent analyzes the capabilities of the executing agents and calculates capability scores based on a pre-trained random forest model using a large amount of historical running data from the executing agents. The central agent analyzes the set of control tasks and sets the priority of each task based on its safety criticality, its temporal priority in the startup process, and the impact of task failure on the overall startup objective. Through agent capability scoring and task priority allocation, accurate data support is provided for dynamic task scheduling, distinguishing the importance of different tasks and the capabilities of different agents, avoiding blind and arbitrary task allocation, and ensuring the safety and orderliness of the startup control process.

[0115] Specifically, all control tasks are sorted according to their priority from high to low, and each task is processed sequentially. For the currently processed task, the central agent selects all executing agents whose control capabilities cover the parameter type of the task and whose status is idle. For example, if the task requires controlling the main steam pressure, executing agents with pressure control capabilities are included in the candidates. From the candidate agents, the one with the highest capability score is selected, and the current task is assigned to the corresponding executing agent. After allocating all tasks in priority order, for executing agents that have not been assigned tasks, the central agent sorts these idle agents in priority from high to low, and at the same time, lists the high-priority tasks of the already assigned executing agents again. According to the corresponding priority order and capability score order, the high-capability idle agents are used as auxiliary agents and matched to the corresponding tasks to enhance the execution reliability of critical tasks, thus obtaining the first matching result.

[0116] It should be noted that dynamic task matching ensures that high-priority tasks are matched with the most capable execution agents, reducing the risk of critical tasks failing due to insufficient execution unit capabilities. By performing secondary matching of high-capability idle agents to high-priority tasks, auxiliary execution guarantees are dynamically added to critical tasks, improving the execution guarantee capability for core tasks and enhancing the robustness and effectiveness of the startup control process.

[0117] Specifically, the central agent analyzes the first matching result, identifying timing conflicts where tasks are scheduled to execute within the same time period, and resource contention where two or more tasks need to operate on the same physical actuator resource within the same time period. For conflicting task pairs, the priorities of the tasks are compared. The original allocation of the high-priority task remains unchanged, while the current agent matching for the lower-priority task is removed, and a new, available execution agent is re-matched for the low-priority task. Simultaneously, it is ensured that the new matching scheme does not create similar conflicts with any other tasks, including but not limited to changing the execution agent for the low-priority task and fine-tuning its planned execution time window within permissible limits to avoid conflicts. After coordinating and resolving the conflicts, the second matching result is obtained. Through conflict detection and resolution, problems such as instruction conflicts, resource contention, and logical sequence disorder that occur in multi-agent distributed execution environments can be avoided, improving the safety and orderliness of the execution process, ensuring the accurate execution of the control process, and enhancing the practicality and reliability of the control system.

[0118] Specifically, based on the second matching result, the central agent generates a corresponding control task list for each executing agent, including but not limited to the target parameters of the task, the specific data of the first start-up curve corresponding to the task, the planned start conditions and time points of the task, and the completion conditions of the task. For tasks involving the collaboration of multiple agents, it also includes necessary collaboration instructions and data interface descriptions. The central agent integrates the control tasks of all executing agents and adds corresponding timing coordination information to obtain the first control strategy. The optimized task allocation result is transformed into a set of operation instructions that the executing agents can clearly understand and follow, ensuring the accurate execution of the control strategy and ensuring that the complex start-up process of the gas-steam combined cycle unit can be automatically, in parallel, collaboratively, and precisely controlled.

[0119] Furthermore, the startup process of the circulating unit is controlled according to the first control strategy, real-time operating data is collected, the deviation between the real-time operating status and the first startup curve is analyzed, and the control parameters of the intelligent agent are dynamically adjusted to optimize the first control strategy, including:

[0120] S801. Control the startup process of the circulating unit according to the first control strategy, collect real-time operating data, analyze the deviation between the real-time operating status and the first startup curve, and obtain the status deviation.

[0121] S802. Determine the deviation level based on the state deviation, dynamically adjust the control parameters of the agent according to the deviation level, and optimize the first control strategy.

[0122] In this embodiment, each executing agent begins to execute its own control task according to the instructions of the first control strategy. This includes, but is not limited to, the gas turbine controller agent controlling the fuel valve according to the curve setting, and continuously and periodically collecting real-time operating data of key unit parameters, including but not limited to the actual gas turbine speed, exhaust temperature, and actual steam drum pressure. The central agent compares the real-time data with the corresponding expected value of the first start-up curve at the same time, calculates the difference between the real-time value and the expected value of each monitored parameter, and obtains the state deviation. By monitoring the start-up process in real time, the deviation of the operation process is quantitatively analyzed, providing accurate deviation data for the control process. It can be adjusted based on the continuous deviation of multiple parameters throughout the entire process, providing accurate data support for the subsequent control optimization process.

[0123] Specifically, the deviation level is determined based on the state deviation, and the control parameters of the intelligent agent are dynamically adjusted according to the deviation level to optimize the first control strategy. Based on actual operation feedback, the control parameters are intelligently adjusted to respond to actual conditions in real time, improving the robustness of the control process in the face of object nonlinearity, time-varying parameters, and external disturbances. This ensures that the startup process is always on the optimal or near-optimal trajectory in uncertain environments, improving the safety, speed, and stability of the startup control process.

[0124] Furthermore, the deviation level is determined based on the state deviation, and the control parameters of the agent are dynamically adjusted according to the deviation level to optimize the first control strategy, including:

[0125] S901. Determine the deviation level based on the state deviation, and match the corresponding optimization parameters in the preset parameter optimization library according to the deviation level to obtain the set of optimization parameters;

[0126] S902. Dynamically adjust the control parameters of the agent according to the optimized parameter set to optimize the first control strategy.

[0127] In this embodiment, a parameter optimization library is constructed by data mining and inductive learning of historical successful intervention cases. The parameter optimization library uses deviation level as the main index condition. Under each condition, for the controlled object and the specific deviation direction, one or more sets of optimization parameters are stored. The optimization parameters include, but are not limited to, the adjustment amounts of the proportional, integral, and derivative gains of the PID controller. Based on the magnitude, trend, and duration of the state deviation, the deviation level is determined. After identifying the deviation of the current control parameters and determining its level, the central intelligent agent searches the parameter optimization library according to the controlled object identifier, deviation direction, and deviation level to quickly match the corresponding combination of optimization parameters and obtain the set of optimization parameters.

[0128] It should be noted that, through the pre-built optimization knowledge base, corresponding corrective measures can be quickly matched based on deviation identification, avoiding the need for complex online calculations or searches every time a deviation occurs, shortening the decision response time, ensuring the reliability and consistency of adjustment suggestions, and improving the effectiveness of the control process.

[0129] Specifically, the central agent generates specific control parameter adjustment instructions based on the optimized parameter set and sends them to the corresponding execution agent through the communication interface. For example, the current value of the proportional gain of the gas turbine speed PID control module is multiplied by a coefficient of 0.95, and the integral time is increased by 10%. After receiving the instructions, the execution agent dynamically updates the corresponding parameter variables, directly replacing the original parameter values. After the parameters are adjusted, the control behavior of the execution agent changes immediately, and the output control instructions are also adjusted accordingly, thus acting on the controlled object. At the same time, the central agent records the parameter adjustment process in the log of the currently executing first control strategy, dynamically updating the strategy.

[0130] It should be noted that by dynamically updating the first control strategy, it can learn from the real-time operating results and immediately fine-tune its own behavior. By dynamically adjusting the underlying control parameters, it can effectively compensate for the errors of the unit model, adapt to the slow time-varying characteristics of the equipment, and suppress unknown external disturbances, ensuring that the actual startup process can closely track the first startup curve and improve the control accuracy, stability and robustness of the startup process when facing uncertainties.

[0131] like Figure 4 As shown, the agent-based cyclic unit start-up control optimization system is used to implement the agent-based cyclic unit start-up control optimization method, including:

[0132] The startup case library construction module analyzes historical startup data and environmental parameters based on the pre-acquired cyclic unit status data and constructs a startup case library.

[0133] The startup curve analysis module responds to the start-up command of the circulating unit, matches at least one historical case in the startup case library based on the current status of the circulating unit, analyzes the parameter changes during the startup process, and generates the first startup curve.

[0134] The multi-agent collaboration module is configured with a multi-agent collaboration mechanism. It decomposes the first start-up curve to obtain multiple control tasks. The central agent analyzes the control capabilities and real-time status of each executing agent and matches at least one control task to each executing agent to obtain the first control strategy. The agents include the central agent and multiple executing agents.

[0135] The control optimization module controls the startup process of the circulating unit according to the first control strategy, collects real-time operating data, analyzes the deviation between the real-time operating status and the first startup curve, dynamically adjusts the control parameters of the intelligent agent to optimize the first control strategy, and optimizes the startup process of the circulating unit.

[0136] In this embodiment, the startup case library construction module analyzes historical startup data, automatically extracts successful startup segments, and clusters them according to boundary conditions to obtain a startup case library. This transforms scattered startup operations into searchable and reusable standardized cases, solving the problem of the difficulty in systematically accumulating and inheriting historical experience, and providing a high-quality data foundation and knowledge reserve for intelligent startup. The startup curve analysis module responds to the startup command, matches similar historical cases in the startup case library according to the real-time unit status, and generates the first startup curve through multi-case fusion and correction. It generates the current optimal startup target trajectory through data-driven generation, solving the problems of poor adaptability and poor economy of traditional startup methods, and providing the unit with safe, efficient, and customized startup path guidance.

[0137] Specifically, the multi-agent collaboration module decomposes the startup curve into specific control tasks and coordinates multiple execution agents through a central agent to allocate tasks and resolve conflicts, resulting in an executable first control strategy. By constructing a collaborative distributed control execution network, it avoids the rigidity of traditional centralized control and the poor coordination between subsystems, achieving optimal scheduling and efficient parallel execution of control resources, and improving the coordination and reliability of complex startup processes. The control optimization module monitors operational deviations in real time during strategy execution and dynamically adjusts the control parameters of each agent according to the deviation level, performing online optimization of the first control strategy. This improves the real-time self-tuning and adaptive capabilities of the control process, solving the performance degradation problem of fixed-parameter control strategies when facing model errors and external disturbances. Through closed-loop optimization, it ensures that the actual startup process closely and smoothly tracks the optimal curve, ultimately guaranteeing the safety, speed, and economy of the startup process.

[0138] The above description is merely a preferred embodiment of this application. The scope of protection of this application is not limited to the above embodiments. All technical solutions falling within the scope of this application's concept are within the scope of protection of this application. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of this application should also be considered within the scope of protection of this application.

Claims

1. An agent-based optimization method for cyclic unit start-up control, characterized in that, include: Based on the pre-acquired circulating unit status data and environmental parameters, analyze historical startup scenarios and build a startup case library; In response to the start-up command of the circulating unit, based on the current status of the circulating unit, at least one historical case is matched in the start-up case library, the parameter changes during the start-up process are analyzed, and the first start-up curve is generated. A multi-agent collaborative mechanism is configured, the first start-up curve is decomposed to obtain multiple control tasks, the control capabilities and real-time status of each executing agent are analyzed by the central agent, and at least one control task is matched for each executing agent to obtain the first control strategy. The agents include the central agent and multiple executing agents. The startup process of the circulating unit is controlled according to the first control strategy. Real-time operating data is collected, the deviation between the real-time operating status and the first startup curve is analyzed, and the control parameters of the intelligent agent are dynamically adjusted to optimize the first control strategy and optimize the startup process of the circulating unit.

2. The agent-based cyclic unit start-up control optimization method according to claim 1, characterized in that, The step involves analyzing historical startup data and environmental parameters based on pre-acquired circulating unit status data to construct a startup case library, including: Based on the pre-acquired cyclic unit status data and environmental parameters, the start and end times of each historical startup process are identified, and multiple startup segments are obtained to construct the first startup segment set. According to the preset startup success criteria, the startup process of each startup segment in the first startup segment set is analyzed, and the startup segments that start successfully are selected to obtain the second startup segment set. Analyze the boundary condition features of each startup fragment in the second startup fragment set, cluster the boundary condition features, and construct a startup case library.

3. The agent-based cyclic unit start-up control optimization method according to claim 1, characterized in that, In response to the circulating unit start-up command, based on the current circulating unit status, at least one historical case is matched in the start-up case library, the parameter changes during the start-up process are analyzed, and a first start-up curve is generated, including: In response to the cyclic unit start-up command, the features of the current start-up task are extracted based on the current cyclic unit status, and a feature vector is constructed. Based on the feature vector, at least one historical case is matched in the startup case library to construct a first case set; Analyze the case relationships and parameter changes during the case startup process in the first case set to generate the first startup curve.

4. The agent-based cyclic unit start-up control optimization method according to claim 3, characterized in that, The analysis of case associations and parameter changes during case initiation within the first case set generates a first initiation curve, including: Extract the static and dynamic parameters corresponding to each case in the first case set, and construct the case feature vector; Calculate the similarity between case feature vectors, and then weight and fuse the parameters of cases with similarity greater than a preset similarity threshold to obtain a second case set; The case with the highest similarity between the feature vectors of the cases in the second case set is selected as the base case. The difference points between each case in the second case set and the base case with parameter differences greater than a preset difference threshold are identified. The difference points are then corrected by interpolation to obtain the third case set. The parameter points corresponding to each case in the third case set are merged to generate the first startup curve.

5. The agent-based cyclic unit start-up control optimization method according to claim 1, characterized in that, The multi-agent cooperative mechanism includes: Multiple control tasks are obtained by decomposing the first start-up curve through the central intelligent agent. The parameter coupling in the control tasks is analyzed, the parameters are aggregated, and a set of control tasks is constructed. The central agent analyzes the control capabilities and real-time status of each executing agent, matches each control task in the control task set with an executing agent, and each executing agent matches at least one control task to obtain the first control strategy.

6. The agent-based cyclic unit start-up control optimization method according to claim 5, characterized in that, The process involves decomposing the first startup curve through a central intelligent agent to obtain multiple control tasks, analyzing the parameter coupling within these tasks, aggregating the parameters, and constructing a control task set, including: Multiple control tasks are obtained by decomposing the first start-up curve through the central intelligent agent. The slope of parameter change in each control task is analyzed, and the coupling strength of parameters at adjacent time nodes is calculated. Parameters with coupling strength greater than a preset coupling threshold are aggregated to obtain a parameter linkage set. By analyzing the number of parameters and the parameter coupling complexity, the parameter linkage set is split to obtain the first subtask group. Analyze the phase lag and amplitude attenuation between parameters in the parameter linkage set, select key parameters as independent control parameter nodes, and obtain the second subtask group; By combining the first subtask group and the second subtask group, a set of control tasks is obtained.

7. The agent-based cyclic unit start-up control optimization method according to claim 6, characterized in that, The process involves analyzing the control capabilities and real-time status of each executing agent through a central agent, matching each control task in the control task set with an executing agent, and each executing agent matching at least one control task, to obtain a first control strategy, including: The central agent analyzes the control capabilities and real-time status of each executing agent, calculates the corresponding capability score, and analyzes the priority of each control task in the control task set. According to priority order, each control task is matched with an execution agent that meets the corresponding task parameters. Execution agents not assigned to control tasks are matched with control tasks of corresponding priority from high to low according to their ability scores, and the first matching result is obtained. The task pairs with timing conflicts and resource competition in the first matching result are identified. The task with the lower priority in the task pair is rematched by the execution agent to obtain the second matching result. Based on the second matching result, the control task of each executing agent is determined, and the first control strategy is obtained.

8. The agent-based cyclic unit start-up control optimization method according to claim 1, characterized in that, The process of controlling the start-up of the circulating unit according to the first control strategy, collecting real-time operating data, analyzing the deviation between the real-time operating status and the first start-up curve, and dynamically adjusting the control parameters of the intelligent agent to optimize the first control strategy includes: The startup process of the circulating unit is controlled according to the first control strategy. Real-time operating data is collected, and the deviation between the real-time operating status and the first startup curve is analyzed to obtain the status deviation. The deviation level is determined based on the state deviation, and the control parameters of the agent are dynamically adjusted according to the deviation level to optimize the first control strategy.

9. The agent-based cyclic unit start-up control optimization method according to claim 8, characterized in that, The step of determining the deviation level based on the state deviation, dynamically adjusting the control parameters of the agent according to the deviation level, and optimizing the first control strategy includes: The deviation level is determined based on the state deviation, and the corresponding optimization parameters are matched in the preset parameter optimization library according to the deviation level to obtain the set of optimization parameters; The control parameters of the agent are dynamically adjusted according to the set of optimized parameters to optimize the first control strategy.

10. An agent-based cyclic unit start-up control optimization system, characterized in that, To implement the agent-based cyclic unit start-up control optimization method as described in any one of claims 1 to 9, comprising: The startup case library construction module analyzes historical startup data and environmental parameters based on the pre-acquired cyclic unit status data and constructs a startup case library. The startup curve analysis module responds to the start-up command of the circulating unit, matches at least one historical case in the startup case library based on the current status of the circulating unit, analyzes the parameter changes during the startup process, and generates the first startup curve. The multi-agent collaboration module is configured with a multi-agent collaboration mechanism. It decomposes the first start-up curve to obtain multiple control tasks. The central agent analyzes the control capabilities and real-time status of each executing agent and matches at least one control task to each executing agent to obtain the first control strategy. The agents include the central agent and multiple executing agents. The control optimization module controls the startup process of the circulating unit according to the first control strategy, collects real-time operating data, analyzes the deviation between the real-time operating status and the first startup curve, dynamically adjusts the control parameters of the intelligent agent to optimize the first control strategy, and optimizes the startup process of the circulating unit.

Images