Steam turbine governing valve self-evolution closed-loop control method, device, equipment and medium

Abstract

The present application belongs to the technical field of steam turbine control, and discloses a steam turbine regulating valve self-evolution closed-loop control method, device, equipment and storage medium. The scheme first collects real-time operation data of the unit, updates parameters and model structure online with the aid of a self-evolution regulating valve dynamic model, obtains a target evolution model adaptive to valve degradation characteristics and a main steam pressure prediction value; relies on multi-granularity working condition identification to output working condition labels and confidence, and then obtains comprehensive disturbance prediction results and confidence through multi-model disturbance prediction fusion; takes stable pressure, smooth valve action, wear inhibition and small opening vibration as optimization targets, uses multi-agent game optimization to cooperatively solve and generate a composite control vector, issues the DCS to complete closed-loop control, and starts arbitration and disturbance-free rollback when an exception occurs. The present scheme is adaptive to long-term degradation characteristics of the valve, compensates various disturbances in advance, takes into account pressure regulating effect and valve protection, and improves main steam pressure control precision, working condition self-adaptation capability and control system field operation safety.

Description

Technical Field

[0001] This application relates to the field of steam turbine control technology, and in particular to a method, device, equipment and medium for self-evolution closed-loop control of steam turbine regulating valve. Background Technology

[0002] With the advancement of the construction of new power systems, thermal power generating units are undertaking increasingly heavy tasks of deep peak shaving, rapid load change, and ancillary services for renewable energy consumption. As the core actuator connecting the boiler's heat output and the turbine's work, the control quality of the turbine regulating valve directly determines the stability of the unit's main steam pressure, load response speed, and operational safety.

[0003] However, the turbine control valve is a complex object characterized by strong nonlinearity, strong coupling, and significant time-varying characteristics. During long-term operation, the valve's true dynamic characteristics continuously degrade due to factors such as changes in valve stem friction, fluctuations in servo valve and hydraulic system oil pressure, dead zone expansion and hysteresis exacerbation caused by valve core and seat wear, and steam parameter drift. This degradation poses a severe challenge to traditional PID control methods based on fixed parameters or static models from the design phase: fixed models cannot track the actual valve state, leading to model mismatch, which in turn causes main steam pressure overshoot, response lag, frequent valve actuation, and even system oscillation.

[0004] To address the aforementioned problems, existing technologies have attempted to introduce intelligent algorithms such as Model Predictive Control (MPC), fuzzy control, or neural networks. However, in practical industrial applications, the following insurmountable drawbacks still exist: First, there is a lack of adaptive capability to the degradation characteristics of valves. Existing intelligent control schemes mostly use fixed structure models trained offline, which cannot perform online recursive estimation and structural evolution updates of degradation parameters such as dead zone and hysteresis of control valves based on real-time operating data, resulting in a gradual disconnect between the control model and the actual physical object.

[0005] Second, there is a lack of cross-timescale prediction capabilities for multi-source disturbances. Main steam pressure is affected by multiple factors coupled together, such as boiler combustion, feedwater disturbances, and grid commands, and these disturbances are characterized by suddenness and multiple timescales. Existing technologies mostly rely on feedback regulation after the deviation occurs, or only use single feedforward compensation, lacking the ability to predict and dynamically fuse short-term pulsations, medium-term trends, and sudden disturbances in parallel, resulting in regulation lag.

[0006] Third, there is a lack of multi-objective collaborative optimization mechanisms. Turbine control needs to simultaneously consider multiple conflicting objectives, such as main steam pressure stability, dynamic response speed, valve wear suppression, and small-opening jitter suppression. Existing technologies often compromise on one aspect while neglecting another; excessively pursuing pressure regulation speed leads to frequent and large valve movements that accelerate wear, while excessively restricting valve movement affects the unit's load response performance.

[0007] Fourth, the lack of a safety rollback mechanism in engineering. When external intelligent algorithms are integrated into a DCS (Distributed Control System), if there is a lack of abnormal arbitration and non-disruptive rollback strategies for communication timeouts, abnormal measurement points, abnormal model residuals, or optimization solution failures, the failure of the algorithm will lead to sudden changes in control output, seriously threatening the safe operation of the unit. Summary of the Invention

[0008] This application provides a self-evolutionary closed-loop control method, device, equipment, and medium for a steam turbine regulating valve. It aims to solve the technical problems existing in related technologies, such as the inability to simultaneously achieve multi-objective collaborative optimization of main steam pressure stability, dynamic response speed, valve life protection, and small opening jitter suppression under complex operating conditions of continuous degradation of regulating valve characteristics and multi-source disturbance coupling, and the lack of an engineering safety rollback mechanism when intelligent algorithms fail.

[0009] In a first aspect, embodiments of this application provide a self-evolving closed-loop control method for a steam turbine regulating valve, the method comprising: Collect real-time operating data of the steam turbine unit; Based on the real-time operating data, the time-varying degradation parameters of the control valve are recursively estimated and the structure is updated online through the self-evolving control valve dynamic model, so as to obtain the target evolution model and the predicted main steam pressure that reflect the current valve's true dynamic characteristics. Based on the real-time operating data, the target evolution model, and the predicted main steam pressure, a multi-granularity operating condition identification algorithm is used to identify the current operating condition and output the current operating condition label and confidence level. Based on the current operating condition label, confidence level, and real-time operating data, a multi-model disturbance prediction fusion algorithm is used to call multiple time scale predictors in parallel to estimate future disturbances, so as to obtain a comprehensive disturbance prediction result and a comprehensive confidence level. With the objective functions of minimizing main steam pressure deviation, smoothing valve action, suppressing valve wear, and suppressing small opening jitter, a composite control vector is output by collaboratively solving the problem based on a multi-agent game optimization algorithm, combined with the objective evolution model, comprehensive disturbance prediction results, comprehensive confidence level, and the current operating condition label. The composite control vector is sent to the DCS execution layer for closed-loop control, and when an abnormal fault in system operation or data interaction is detected, abnormal arbitration and uninterrupted rollback are performed.

[0010] In one embodiment, optionally, the time-varying degradation parameters of the control valve are recursively estimated and the structural evolution is updated online using a self-evolving control valve dynamic model, including: A gray-box state-space model is constructed that includes dead-zone nonlinearity and hysteresis nonlinearity, wherein the dead-zone nonlinearity is characterized by a dead-zone function and the hysteresis nonlinearity is characterized by a hysteresis mapping. The parameters of the gray box state space model are updated online using a recursive least squares method with engineering constraints, and saturation constraints and rate of change constraints are applied to the updated parameters. Maintain a model set consisting of multiple candidate sub-models, calculate the residual energy of each candidate sub-model within a sliding window, and dynamically allocate the weights of each candidate sub-model according to the residual energy using a preset fusion formula to generate the target evolution model.

[0011] In one embodiment, optionally, the method further includes: Based on the dead zone width parameter, hysteresis parameter and residual energy, the valve health index is calculated, and based on the valve health index, it is determined whether to update the parameters of the self-evolving control valve dynamic model and whether to reallocate the weights of the candidate sub-models. The valve health indicators include the current dead zone width parameter, hysteresis parameter, and optimal sub-model residual energy.

[0012] In one embodiment, optionally, the use of a multi-model perturbation prediction fusion algorithm, which calls predictors at multiple time scales in parallel to estimate future perturbations, includes: Multiple time-scale predictors are invoked in parallel to output the perturbation prediction value for a preset number of steps in the future. The multiple time-scale predictors include at least a short-term predictor, a medium-term predictor, and a mutation predictor. Based on the prediction error energy of each predictor within the historical sliding window and the gating function of the current operating condition label, the dynamic weights corresponding to each predictor are calculated. The outputs of the predictors at the multiple time scales are weighted and summed to obtain the comprehensive perturbation prediction result, and the comprehensive confidence level is calculated.

[0013] In one embodiment, optionally, the objective functions are minimizing main steam pressure deviation, smoothing valve action, suppressing valve wear, and suppressing small-opening jitter. A multi-agent game optimization algorithm is used, combined with the objective evolution model, comprehensive disturbance prediction results, comprehensive confidence level, and the current operating condition label, to collaboratively solve the problem and output a composite control vector, including: Construct a multi-agent decision-making architecture that includes a stress performance agent, a life protection agent, and a dynamic response agent; Within the prediction time domain, an objective function is constructed, which includes minimizing the main steam pressure deviation, smoothing valve action, suppressing valve wear, and suppressing small opening jitter, and each is assigned a corresponding weight coefficient. Based on the target evolution model and the comprehensive perturbation prediction results, the local optimal control sequence for each agent is solved respectively; The baseline values ​​of each weight coefficient in the objective function are adjusted according to the current operating condition label, and the valve action increment penalty weight and the maximum allowable action increment constraint are dynamically adjusted according to the comprehensive confidence level. Through Nash equilibrium game iteration, each local optimal control sequence is coordinated and adjusted until the preset system stability constraints and valve action constraints are met, thereby generating the global optimal control sequence. Based on the globally optimal control sequence, the composite control vector is generated, wherein the composite control vector includes: valve position control increment, baseline PID parameter correction, feedforward compensation value, and adaptive micro-opening suppression coefficient.

[0014] In one embodiment, optionally, the step of performing abnormal arbitration and uninterrupted rollback when a system operation or data interaction anomaly is detected includes: Real-time monitoring of the communication link status between the external intelligent computing layer and the DCS execution layer, the validity of the real-time running data, the residual level of the self-evolving regulating valve dynamic model, and the solution status and computation time of the multi-agent game optimization algorithm; When any one or more abnormal states are detected, such as communication timeout, abnormal measurement data, residual of the self-evolving regulating valve dynamic model exceeding the preset threshold, infeasibility of optimization solution, or calculation timeout, the composite control vector generated in the previous cycle remains unchanged and is issued as the control command for the current cycle, while reducing the upper limit of valve action change. If the abnormal state continues for more than a preset time, the control authority will be switched seamlessly to the original baseline PID control loop of the DCS execution layer, and the correction values ​​in the composite control vector will always obey the inherent amplitude limiting, speed limiting and protection interlocking logic of the DCS execution layer.

[0015] In one embodiment, optionally, before collecting the real-time operating data of the steam turbine unit, the method further includes: The self-evolving control valve dynamic model, the multi-granularity operating condition identification algorithm, the multi-model perturbation prediction and fusion algorithm, and the multi-agent game optimization algorithm are pre-trained and validated offline through a simulation learning layer, specifically including: Construct a full-condition coupled simulation model of boiler-turbine-regulating valve-EH actuator. The full-condition coupled simulation model includes at least steady-state condition, load increase / decrease condition, disturbance enhancement condition, oil pressure fluctuation condition, and valve dead zone hysteresis increase condition. Pre-set load disturbances, boiler-side input changes, and oil pressure factor fluctuations are injected into the full-condition coupled simulation model to simulate the multi-source disturbance scenario of the actual unit. The self-evolving regulating valve dynamic model, the multi-granularity working condition identification algorithm, the multi-model disturbance prediction and fusion algorithm, and the multi-agent game optimization algorithm are run to iteratively tune the control parameters and record the control performance index under boundary working conditions. The control parameters include the evolution parameters, working condition identification parameters, prediction and fusion parameters, and game optimization parameters in the gray box state space model. When the control performance indicators meet the preset stability, adjustment accuracy and valve life protection requirements, the adjusted control parameters and mapping relationship are solidified and deployed to the external intelligent computing layer.

[0016] Secondly, embodiments of this application provide a self-evolving closed-loop control device for a steam turbine regulating valve, the device comprising: The data acquisition module is used to collect real-time operating data of the steam turbine unit; The update module is used to perform online recursive estimation and structural evolution update of the time-varying degradation parameters of the control valve based on the real-time operating data and through the self-evolving control valve dynamic model, so as to obtain the target evolution model and predicted main steam pressure that reflect the current true dynamic characteristics of the valve. The identification module is used to identify multiple operating conditions based on the real-time operating data, the target evolution model, and the predicted main steam pressure, using a multi-granularity operating condition identification algorithm, and output the current operating condition label and confidence level. The prediction module is used to predict future disturbances by using a multi-model disturbance prediction fusion algorithm based on the current operating condition label, confidence level and real-time operating data, and calling multiple time scale predictors in parallel to obtain a comprehensive disturbance prediction result and a comprehensive confidence level. The output module is used to take minimizing the main steam pressure deviation, smoothing valve action, suppressing valve wear, and suppressing small opening jitter as objective functions. Based on the multi-agent game optimization algorithm, it combines the objective evolution model, comprehensive disturbance prediction results, comprehensive confidence and the current operating condition label to perform collaborative solution and output a composite control vector. The execution module is used to send the composite control vector to the DCS execution layer for closed-loop control, and to perform abnormal arbitration and uninterrupted rollback when an abnormal fault in system operation or data interaction is detected.

[0017] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described self-evolutionary closed-loop control method for turbine regulating valves.

[0018] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described self-evolutionary closed-loop control method for turbine regulating valves.

[0019] In the above-mentioned scheme implemented by the self-evolutionary closed-loop control method, device, equipment, and medium for steam turbine regulating valves, real-time operating data of the steam turbine unit is collected; based on the real-time operating data, the time-varying degradation parameters of the regulating valve are recursively estimated and the structure is updated online through a self-evolutionary regulating valve dynamic model to obtain a target evolution model reflecting the current true dynamic characteristics of the valve and a predicted main steam pressure; based on the real-time operating data, the target evolution model, and the predicted main steam pressure, multi-granularity operating condition identification is performed through a multi-granularity operating condition identification algorithm to output the current operating condition label and confidence level; based on the current operating condition label, confidence level, and real-time operating data... According to the algorithm, a multi-model disturbance prediction fusion algorithm is adopted to predict future disturbances in parallel using multiple time scales to obtain a comprehensive disturbance prediction result and a comprehensive confidence level. The algorithm uses minimizing main steam pressure deviation, smoothing valve action, suppressing valve wear, and suppressing small opening jitter as objective functions. Based on a multi-agent game optimization algorithm, it collaboratively solves the problem using the objective evolution model, the comprehensive disturbance prediction result, the comprehensive confidence level, and the current operating condition label, outputting a composite control vector. This composite control vector is then sent to the DCS execution layer for closed-loop control. When a system computation or data interaction anomaly is detected, anomaly arbitration and disturbance-free backoff are performed. In this invention, the self-evolving control valve dynamic model completes online parameter recursive estimation and structural evolution updates based on real-time operating data, adapting to various time-varying degradation characteristics of the valve and obtaining accurate predictions of main steam pressure, overcoming the model mismatch problem caused by fixed models. It relies on multi-granularity operating condition identification to output operating condition labels and confidence levels, supporting the fusion of multi-model disturbance predictions to complete multi-timescale disturbance predictions and obtain comprehensive disturbance prediction results and confidence levels. Pre-compensation based on prediction results improves regulation lag and pressure overshoot oscillation problems. Through multi-agent game optimization, aiming at minimizing pressure deviation, smoothing valve action, and suppressing wear and small-opening jitter, it combines the evolutionary model, disturbance prediction, and operating condition labels to collaboratively solve and generate a composite control vector, balancing main steam pressure control effects with valve protection requirements. The composite control vector is issued to the DCS for closed-loop execution. When the system experiences computational or data interaction anomalies, it triggers abnormal arbitration and uninterrupted backoff to prevent sudden changes in control output, improving the adaptive level, operational stability, and commissioning safety of the control system. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A schematic flowchart of a self-evolutionary closed-loop control method for a turbine regulating valve according to an embodiment of this application is shown.

[0022] Figure 2 A schematic flowchart of step S102 in a turbine regulating valve self-evolution closed-loop control method according to an embodiment of this application is shown.

[0023] Figure 3 A schematic flowchart of step S104 in a turbine regulating valve self-evolution closed-loop control method according to an embodiment of this application is shown.

[0024] Figure 4 A schematic flowchart of step S105 in a turbine regulating valve self-evolution closed-loop control method according to an embodiment of this application is shown.

[0025] Figure 5 The figure shows the simulation results of LSTM main steam pressure response, valve opening and adaptive state variables according to an embodiment of this application.

[0026] Figure 6 The simulation results of multi-source disturbance, oil pressure factor and feedforward compensation amount according to an embodiment of this application are shown.

[0027] Figure 7 A comparison chart of control performance indicators under different operating conditions according to an embodiment of this application is shown.

[0028] Figure 8 A block diagram of a turbine regulating valve self-evolution closed-loop control device according to an embodiment of this application is shown.

[0029] Figure 9 A block diagram of a computer device according to one embodiment of this application is shown. Detailed Implementation

[0030] To better understand the technical solution of this application, the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0031] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0032] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0033] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0034] This invention addresses the problems of strong nonlinearity, significant time-varying degradation, multiple disturbance sources spanning time scales, and coupling of control objectives (pressure stabilization / response / life protection) in the main steam pressure regulation of large thermal power units. It employs a "four-intelligent algorithm + three-layer collaborative architecture" to achieve closed-loop control. The four intelligent algorithms include: 1) SEM: Dynamic model of self-evolving control valve (parameter evolution + structural evolution) 2) Multi-granularity operating condition identification (operating condition + disturbance type + valve health status) 3) Multi-model perturbation prediction fusion (parallel short-term / medium-term / mutation prediction + dynamic weighting) 4) MAGO: Adaptive multi-objective optimization of multi-agent game (voltage regulation-response-lifetime coordination, outputting a composite control vector) The overall closed-loop link is fixed as follows: Data acquisition → SEM update → Multi-granularity working condition identification → Multi-model prediction fusion → MAGO multi-objective optimization → DCS execution → Safety arbitration rollback → Simulation learning layer verification update.

[0035] Three-tier collaborative architecture and data interaction 1) Control and Execution Layer (DCS / DEH / EH) The DCS does not perform the calculations for the four intelligent algorithms; it is only responsible for: real-time acquisition and uploading of main steam pressure. Valve position feedback ,load EH hydraulic pressure Valve stem / servo related quantities, etc.; Receive and execute composite control vectors issued by the intelligent server: Valve target opening degree / increment: or

[0036] Dynamic PID parameter correction: , ,

[0037] Feedforward compensation:

[0038] Micro-aperture suppression coefficient:

[0039] Operating condition switching signal:

[0040] Retain and prioritize the execution of safety logic: amplitude and speed limiting, overtravel, hydraulic pressure protection, and unit protection interlocking; and support uninterrupted baseline return control. Engineering reinforcement: DCS-side output execution follows the principle of "protection priority," with intelligent quantities participating in execution as control commands or correction quantities, and under no circumstances should the existing DCS protection interlocks be bypassed.

[0041] 2) External intelligent computing layer The intelligent server receives data frames at fixed intervals:

[0042] Run the four intelligent algorithms in a fixed order and output the results:

[0043] The communication interface can be OPC UA, Modbus TCP or a manufacturer's proprietary protocol, and the period can be set to 20–200ms, preferably 50–100ms.

[0044] To ensure operational feasibility, the intelligent server sets "timeout judgment" and "feasibility judgment" for each cycle of calculation, and the results are used for abnormal arbitration and rollback.

[0045] 3) Simulation Learning Layer 4) Used for training, verification, parameter tuning, and boundary condition testing; after simulation verification, the strategy parameters / mappings are solidified and deployed to the intelligent server to realize the engineering closed loop of "learning-verification-deployment".

[0046] Please see Figure 1 , Figure 1 A schematic flowchart of a self-evolutionary closed-loop control method for a turbine regulating valve according to an embodiment of this application is shown.

[0047] like Figure 1 As shown, a self-evolutionary closed-loop control method for turbine regulating valves includes: Step S101: Collect real-time operating data of the steam turbine unit; In this step, key process quantities reflecting the unit's operating status can be collected in real time through the DCS / DEH / EH system. The real-time operating data includes, but is not limited to: main steam pressure (the core controlled quantity reflecting the steam energy at the boiler outlet), regulating valve position feedback (a feedback quantity reflecting the actual position of the valve), load command (an external command reflecting the power grid's demand on the unit's output), and EH oil pressure (reflecting the operating condition of the power source of the hydraulic actuator).

[0048] By acquiring high-frequency, multi-dimensional data, the control system is able to capture subtle changes in the dynamic characteristics of the unit in real time, providing accurate and complete data support for subsequent model self-evolution, operating condition identification, and optimization decisions. This is a prerequisite for realizing closed-loop intelligent control.

[0049] Step S102: Based on the real-time operating data, the time-varying degradation parameters of the control valve are recursively estimated and the structure is updated online through the self-evolving control valve dynamic model to obtain the target evolution model and the predicted main steam pressure that reflect the current true dynamic characteristics of the valve. like Figure 2 As shown, in one embodiment, optionally, step S102 includes: Step S201: Construct a gray box state space model including dead zone nonlinearity and hysteresis nonlinearity, wherein the dead zone nonlinearity is characterized by a dead zone function and the hysteresis nonlinearity is characterized by a hysteresis mapping. This step aims to address the problem of valve physical properties degrading over time. A gray-box state-space model with dead zone and hysteresis nonlinearity is adopted:

[0050] For the model state; where For the set of evolutionary parameters, These are exogenous quantities (composed of load, oil pressure, etc.). A, B, and E are state-space matrices related to the parameters.

[0051]

[0052] The model predicts the main steam pressure; C and D are the output matrices. This represents the effective input acting on the control valve model at time k, preferably using a combination of dead zone and hysteresis. Degenerate nonlinearity is addressed through the effective input.

[0053] DZ( ) indicates dead zone nonlinearity; This is the dead zone width parameter that evolves over time; the output is 0 when the input amplitude is less than the dead zone threshold. This is the set of parameters for hysteresis evolution.

[0054] ( ) represents a hysteresis mapping; This is a set of hysteresis parameters (which may include loop width, slope, etc.); this formula indicates that the dead zone output is passed through a hysteresis stage to obtain an effective input, which is used to more closely approximate the actual response of the degraded valve.

[0055] Step S202: The parameters of the gray box state space model are updated online using the recursive least squares method with engineering constraints, and saturation constraints and rate of change constraints are applied to the updated parameters. Regression form:

[0056] (k) is the regression vector (composed of state, input, exogenous quantities, etc.); The parameter to be estimated; To model errors / noise, RLS updates are preferred.

[0057] This formula is a recursive update form (e.g., the unified notation for RLS / gradient-type updates); This is the gain matrix; the value in parentheses represents the current residual. Engineering constraints are then applied:

[0058] The above constraints make parameter updates controllable and reduce the risk of parameter mutations caused by noise or abnormal data. The first constraint is a parameter saturation constraint, which ensures that the parameters fall within the engineering confidence range; the second constraint is a parameter change rate constraint, which suppresses drastic parameter jumps caused by noise and improves operational stability.

[0059] Step S203: Maintain a model set consisting of multiple candidate sub-models, calculate the residual energy of each candidate sub-model within the sliding window, and dynamically allocate the weights of each candidate sub-model according to the residual energy through a preset fusion formula to generate the target evolution model.

[0060] Candidate model set Residuals and window costs: Regression form:

[0061] Let be the prediction residual of the i-th candidate sub-model; Output for the i-th sub-model. Soft fusion (used to reduce jitter): The residual energy (decidable quantity) within the window length W is used to evaluate the recent fit of each sub-model.

[0062] β is the soft fusion weight; M is the number of candidate models; the smaller the residual, the greater the weight, and the weight is naturally normalized.

[0063] Valve health indicators and evolution triggers (for use in operating condition identification and optimization) Constructing health indicators (example): Regression form:

[0064] in: Valve health indicators; , , This is the nominal reference value; , , As weight;

[0065] This represents the residual energy of the current optimal submodel. This metric reflects the degree of degradation and drives weighting / adjustment constraints and evolution triggers. Triggering conditions (determinable) include at least one of the following: residual statistics exceeding a threshold. or The preset number of times is reached continuously; a change in operating condition triggers: MS(k) changes and Health changes trigger: or When triggered, accelerated updates, candidate sub-model weight reallocation, or fusion enhancement can be performed, combined with... Suppress frequent switching.

[0066] Step S103: Based on the real-time operating data, the target evolution model, and the predicted main steam pressure, multi-granularity operating condition identification is performed using a multi-granularity operating condition identification algorithm, and the current operating condition label and confidence level are output. This step involves extracting feature vectors for identification.

[0067]

[0068] Where z(k) represents the feature vector used for operating condition identification in the k-th control cycle; P represents the main steam pressure. This represents the rate of change of main steam pressure, and L represents the unit load or load command. This indicates the load change rate, and H represents the EH oil pressure or oil pressure status quantity. This indicates the fluctuation characteristics of the main steam pressure within the sliding window. This vector represents the fluctuation characteristics of the valve position within the sliding window, where x represents the control valve position feedback, HI represents the valve health index, and ... represents other optional operating characteristics. This feature vector is used to identify the current operating condition, disturbance type, and valve health status.

[0069] Output multi-granularity operating condition labels: coarse-grained labels include steady state, load increase / decrease, and enhanced disturbance; fine-grained labels include oil pressure fluctuation type, boiler-side trend type, and load step type; ultra-fine-grained labels (valve health / abnormality) include micro-opening jitter, increased hysteresis, dead zone expansion, and response decay. Simultaneously, output confidence scores to characterize the degree of matching between the current feature vector and the standard operating condition template, for use by subsequent prediction and optimization modules.

[0070] The results of the working condition identification are used for: (1) As a weight gating function in prediction fusion Input; (2) As input to MAGO weighting / constraint (e.g., when switching operating conditions, a certain response enhancement is allowed but speed limit constraint is still applied). (3) MS(k) is a flag used for DCS-side policy switching or protection policy selection.

[0071] Step S104: Based on the current operating condition label, confidence level and real-time operating data, a multi-model disturbance prediction fusion algorithm is used to call multiple time scale predictors in parallel to predict future disturbances, so as to obtain a comprehensive disturbance prediction result and a comprehensive confidence level. like Figure 3 As shown, in one embodiment, optionally, step S104 includes: Step S301: Multiple time-scale predictors are invoked in parallel to output the perturbation prediction value for a preset number of steps in the future. The multiple time-scale predictors include at least a short-term predictor, a medium-term predictor, and a mutation predictor. Step S302: Calculate the dynamic weights of each predictor based on the prediction error energy of each predictor within the historical sliding window and the gating function of the current operating condition label. Step S303: The outputs of the predictors at the multiple time scales are weighted and summed to obtain the comprehensive perturbation prediction result, and the comprehensive confidence level is calculated.

[0072] Predictors with multiple time scales include, but are not limited to: short-term predictors, medium-term predictors, and burst predictors.

[0073] Parallel predictor set Covering perturbations at different time scales, outputting the future Step trend or equivalent disturbance:

[0074] Wherein, the j-th predictor outputs the future Step disturbance / trend prediction; This is a prediction step; the prediction can be an equivalent disturbance or a pressure trend.

[0075] Online error statistics:

[0076] in, Let J be the error energy of predictor j within window Wp; The actual main steam pressure at time τ; The replay forecast of pressure over j forecast periods is used to measure the forecast error of the forecaster within the recent window.

[0077] Dynamic weighting based on operating conditions and errors:

[0078] in, For weighting; This is the error sensitivity coefficient; Here, J is the operating condition gating function (preferring different predictors under different operating conditions); J is the total number of predictors; this formula ensures weight normalization and can be updated online.

[0079] Fusion prediction:

[0080] in, The fused perturbation prediction results are used as inputs for both the feedforward and optimizer predictions.

[0081]

[0082] Where C(k) is the fusion confidence (concentration index); when the weights are concentrated in a few predictors, C(k) is larger, indicating that the fusion is more certain; when the weights are dispersed, C(k) is smaller, indicating higher uncertainty, and is used for robust weight adjustment / tightening constraints.

[0083] Fusion prediction is used for: 1) Generate feedforward compensation FF(k); 2) Provide MAGO with predictive input and confidence level C(k) for weight and constraint adjustment to achieve robust control (e.g., increasing action penalty weights or tightening them when confidence decreases). ).

[0084] Step S105: With the objective functions of minimizing main steam pressure deviation, smoothing valve action, suppressing valve wear, and suppressing small opening jitter, a composite control vector is output by collaboratively solving the problem based on a multi-agent game optimization algorithm, combined with the objective evolution model, comprehensive disturbance prediction results, comprehensive confidence level, and the current operating condition label. like Figure 4 As shown, in one embodiment, optionally, step S105 includes: Step S401: Construct a multi-agent decision-making architecture including a stress performance agent, a life protection agent, and a dynamic response agent; Multiple objectives are assigned to multiple agents (e.g., stress performance agent, life protection agent, dynamic response agent). Each agent provides local policy suggestions based on SEM prediction and perturbation fusion prediction, and the final control vector is formed through a cooperative / game mechanism.

[0085] Step S402: In the prediction time domain, construct an objective function, which includes minimizing the main steam pressure deviation, smoothing valve action, suppressing valve wear, and suppressing small opening jitter, and assign corresponding weight coefficients to each. Step S403: Based on the target evolution model and the comprehensive perturbation prediction results, solve for the local optimal control sequence of each agent; Step S404: Adjust the baseline values ​​of each weight coefficient in the objective function according to the current operating condition label, and dynamically adjust the valve action increment penalty weight and the maximum allowable action increment constraint according to the comprehensive confidence level. Step S405: Through Nash equilibrium game iteration, the local optimal control sequences are coordinated and adjusted until the preset system stability constraints and valve action constraints are met, thereby generating the global optimal control sequence. Step S406: Generate the composite control vector based on the global optimal control sequence, wherein the composite control vector includes: valve position control increment, baseline PID parameter correction, feedforward compensation value, and adaptive micro-opening suppression coefficient.

[0086] Define error:

[0087] Output composite control vector:

[0088] Among them, the target valve opening degree / increment is the core decision variable, while feedforward, PID correction and micro-opening suppression are auxiliary variables.

[0089] Main steam pressure deviation; Set the pressure value.

[0090] Multi-objective cost function (voltage regulation + response + lifetime protection + jitter suppression) Minimize the overall cost over the prediction time domain N (example):

[0091] Where N is the prediction time domain; , , , As weight; For future error prediction based on SEM and perturbation prediction; For valve action increments; For wear and tear agents; This is a penalty for small opening jitter. This objective simultaneously covers voltage regulation, smooth operation, lifespan, and jitter suppression.

[0092]

[0093] in This represents the change in valve action, used for action smoothing and life constraint modeling.

[0094] Wear agent and jitter penalty (at least one or a combination) Wear agents (at least one or a combination):

[0095] Penalty for small opening shake:

[0096] Wear risk is characterized by the magnitude of movement, energy, and number of times the threshold is exceeded; , , 1 represents the coefficient; 1 represents the indicator function; This is the threshold for "major actions".

[0097] MAGO Game Theory Solution and Adaptive Weight Adjustment (Maintaining the "Game Theory" Approach) Multiple objectives are assigned to multiple agents (e.g., stress performance agent, life protection agent, dynamic response agent). Each agent provides local policy suggestions based on SEM prediction and perturbation fusion prediction, and the final control vector U(k) is formed through a cooperative / game mechanism.

[0098] Weights and boundaries adaptively adjust based on operating conditions, health, and prediction confidence (example):

[0099] The above approach enables the optimizer to favor conservative control when valve health declines or prediction uncertainty increases, reducing the risk of excessive movement and jitter.

[0100]

[0101] Output generation (keeping the output definition unchanged) Feedforward:

[0102] In the formula, Indicates the amount of feedforward compensation; Indicates feedforward gain; This represents the predicted value for the next disturbance or pressure trend. This formula indicates that the prediction results can be used to prepare compensation in advance, increasing the lead time for control.

[0103] PID correction:

[0104] In the formula, , , These represent the baseline PID parameters; , , These represent the online correction amounts; These represent the corrected PID parameters.

[0105] Micro-aperture suppression:

[0106] In the formula, Indicates the micro-aperture suppression coefficient; Indicates the nominal suppression coefficient; This represents the adjustment function related to operating conditions and valve health status. The formula indicates the dynamic adjustment of the small opening suppression intensity based on different operating conditions and valve health levels.

[0107] Step S106: The composite control vector is sent to the DCS execution layer for closed-loop control, and when an abnormal fault in system operation or data interaction is detected, abnormal arbitration and uninterrupted rollback are performed.

[0108] In one embodiment, optionally, the step of performing abnormal arbitration and uninterrupted rollback when a system operation or data interaction anomaly is detected includes: The system monitors in real time the communication link status between the external intelligent computing layer and the DCS execution layer, the validity of the real-time running data, the residual level of the self-evolving regulating valve dynamic model, and the solution status and computation time of the multi-agent game optimization algorithm. The external intelligent computing layer consists of an independent server unit that deploys self-evolving modeling, operating condition identification, prediction fusion, and multi-agent optimization algorithms; the DCS execution layer is the original distributed control system of the unit, responsible for the execution of field valves and basic protection logic; the model residual level is the difference between the predicted main steam pressure and the actual main steam pressure output by the self-evolving model; and the solution status is the operating condition under which the multi-agent optimization operation can successfully obtain a feasible solution.

[0109] When any one or more abnormal states are detected, such as communication timeout, abnormal measurement data, residual of the self-evolving regulating valve dynamic model exceeding the preset threshold, infeasibility of optimization solution, or calculation timeout, the composite control vector generated in the previous cycle remains unchanged and is issued as the control command for the current cycle, while reducing the upper limit of valve action change. Communication timeout refers to the data interaction between the intelligent computing layer and the DCS exceeding the set cycle limit; abnormal measurement point data refers to the distortion, absence, or jump of field-acquired parameters; infeasible optimization solution refers to the inability of multi-agent optimization to calculate a compliant control solution under the existing constraints; upper limit of valve action change refers to the maximum allowable change in valve opening in a single cycle.

[0110] Upon triggering any preset abnormal operating condition, the effective composite control vector from the previous cycle is latched for use in the current cycle's output, and the single-action amplitude limit of the valve is simultaneously tightened. In this way, large jumps in control commands are avoided in the early stages of an abnormality, and drastic valve fluctuations are suppressed by tightening the limit, achieving a smooth and instantaneous transition during the fault.

[0111] If the abnormal state continues for more than a preset time, the control authority will be switched seamlessly to the original baseline PID control loop of the DCS execution layer, and the correction values ​​in the composite control vector will always obey the inherent amplitude limiting, speed limiting and protection interlocking logic of the DCS execution layer.

[0112] The preset duration is the critical time threshold for determining instantaneous and persistent faults; the baseline PID control loop is the traditional PID basic control logic natively supplied to the DCS; the amplitude limiting, speed limiting, and protection interlocking logic are the valve opening upper and lower limits, action rate constraints, and unit safety interlocking protection built into the DCS.

[0113] If an anomaly persists for an extended period, the system smoothly switches back to the native baseline PID control. All corrective parameters in the intelligent optimization output must not override the original safety logic of the DCS. This ensures that the system completely escapes the control of the anomaly-based intelligent algorithm during prolonged faults, relying on mature original manufacturer control to guarantee unit safety, preventing equipment malfunctions caused by intelligent algorithm failures, and improving the reliability of on-site engineering commissioning.

[0114] Abnormal triggering conditions (determinable)

[0115] In the formula, This means that when an anomaly occurs, the control output at the current moment remains the feasible output from the previous moment; This represents the speed limit tightening factor, with a value range of [value missing]. ; This indicates that the upper limit of the control input variation is tightened simultaneously to improve control safety under abnormal operating conditions.

[0116] In one embodiment, optionally, the method further includes: Based on the dead zone width parameter, hysteresis parameter and residual energy, the valve health index is calculated, and based on the valve health index, it is determined whether to update the parameters of the self-evolving control valve dynamic model and whether to reallocate the weights of the candidate sub-models. The valve health indicators include the current dead zone width parameter, hysteresis parameter, and optimal sub-model residual energy.

[0117] Dead zone width parameter is a quantitative parameter characterizing the minimum input amplitude required for the control valve core to start its movement, which gradually increases with valve body wear; hysteresis parameter is a characteristic parameter reflecting the deviation of the valve's lifting and lowering stroke; residual energy is a cumulative statistical measure of the deviation between the model's predicted value and the actual operating value over a period of time; valve health index is a quantitative parameter formed by combining dead zone, hysteresis, and model fitting deviation, used to characterize the degree of degradation of the control valve; candidate sub-models are multiple backup models in the SEM model library adapted to different degrees of valve degradation.

[0118] In this step, valve health indicators are generated based on the real-time updated dead zone width parameters, hysteresis parameters, and residual energy corresponding to the optimal sub-model. These indicator values ​​serve as the criterion for triggering online parameter updates to the self-evolving model and re-weighting of candidate sub-models as needed. This approach autonomously triggers model evolution based on the actual degree of valve degradation, avoiding blindly updating parameters at fixed intervals. It allows the model to adaptively follow the valve's aging state, continuously reducing control errors caused by model mismatch.

[0119] In one embodiment, optionally, before collecting the real-time operating data of the steam turbine unit, the method further includes: The self-evolving control valve dynamic model, the multi-granularity operating condition identification algorithm, the multi-model perturbation prediction and fusion algorithm, and the multi-agent game optimization algorithm are pre-trained and validated offline through a simulation learning layer, specifically including: Construct a full-condition coupled simulation model of boiler-turbine-regulating valve-EH actuator. The full-condition coupled simulation model includes at least steady-state condition, load increase / decrease condition, disturbance enhancement condition, oil pressure fluctuation condition, and valve dead zone hysteresis increase condition. The full-condition coupled simulation model is an integrated simulation mathematical model that fully reproduces the linkage characteristics of boiler, turbine, regulating valve, and hydraulic actuator. Various operating conditions correspond to typical on-site operating scenarios such as stable unit operation, variable load operation, external disturbance impact, abnormal EH oil pressure, and valve body wear and deterioration.

[0120] In this step, a multi-device linkage simulation framework is built, covering all typical operating conditions of the unit, including daily operation and failure degradation. This fully replicates various on-site operating environments, ensuring a high degree of consistency between the simulation training scenario and the actual unit operating conditions, and improving the engineering applicability of the algorithm tuning parameters.

[0121] Pre-set load disturbances, boiler-side input changes, and oil pressure factor fluctuations are injected into the full-condition coupled simulation model to simulate the multi-source disturbance scenario of the actual unit. Load disturbance refers to sudden changes or fluctuations in unit load caused by changes in grid-side commands; boiler-side disturbances are caused by changes in boiler input, such as fuel and feedwater variations; oil pressure factor fluctuations are unsteady changes in the working pressure of the EH hydraulic system; and multi-source disturbance scenarios are the actual operating environment of the unit where multiple types of disturbances coexist. This step artificially superimposes multiple types of disturbance variables into the simulation model to simulate the complex operating conditions where multiple disturbances occur simultaneously in the field. In this way, the algorithm is debugged under complex disturbance environments, improving the algorithm's adaptability and anti-disturbance capability in the face of multi-source disturbances after it goes live.

[0122] The self-evolving regulating valve dynamic model, the multi-granularity working condition identification algorithm, the multi-model disturbance prediction and fusion algorithm, and the multi-agent game optimization algorithm are run to iteratively tune the control parameters and record the control performance index under boundary working conditions. The control parameters include the evolution parameters, working condition identification parameters, prediction and fusion parameters, and game optimization parameters in the gray box state space model. Iterative tuning involves repeatedly debugging and correcting various parameters until the control effect meets the target. The gray-box state-space model is a modeling method adopted by the SEM self-evolution model that takes into account both the mechanism and measured data. Control performance indicators are quantitative data used to evaluate the quality of pressure regulation accuracy, response speed, and valve action loss. Boundary conditions include harsh extreme operating conditions such as unit operating limits and severe valve degradation. This step involves running four core algorithms in the simulation environment throughout the entire process, iteratively optimizing various parameters, and simultaneously collecting various control indicators under extreme conditions. In this way, parameter refinement is completed under extreme conditions, avoiding control failures in critical unit operating conditions and optimizing control performance under all operating conditions.

[0123] When the control performance indicators meet the preset stability, adjustment accuracy and valve life protection requirements, the adjusted control parameters and mapping relationship are solidified and deployed to the external intelligent computing layer.

[0124] The parameter and mapping relationship are fixed, meaning the optimal parameters after simulation are achieved are locked and no longer subject to random changes. The external intelligent computing layer is a dedicated server hardware platform, independent of the DCS, that runs intelligent algorithms. In this step, after all performance indicators meet the standards, all optimal parameters and algorithm mapping logic are locked and deployed in batches to the on-site intelligent computing server. This enables seamless implementation of simulation verification results into the project, ensuring that the algorithm has qualified control performance from the initial stage of deployment and reducing the workload of on-site commissioning and debugging.

[0125] The following is a simulation and verification of the above embodiments of the present invention.

[0126] In a preferred embodiment, a coupled simulation model of boiler-turbine-regulating valve-EH actuator is established to verify the effectiveness of the four-intelligent closed-loop control method described in this invention. The simulation model includes main steam pressure dynamics, regulating valve opening actuation, load disturbance input, boiler-side input, oil pressure factor, feedforward compensation, and abnormal backoff logic.

[0127] Two control methods were compared in the simulation: baseline PID control and the four-intelligent algorithm control proposed in this invention. The four-intelligent algorithm control includes a SEM self-evolution model, multi-granularity operating condition identification, multi-model disturbance prediction fusion, and a MAGO multi-objective optimization module. The simulation sampling period was Ts. Based on data such as main steam pressure P(k), valve opening x(k), load disturbance, boiler-side input, and oil pressure status, the controller outputs online the target valve opening or increment, feedforward compensation, parameter correction, and abnormal backoff flag.

[0128] like Figure 5 As shown, under the same disturbance conditions, both baseline PID control and the four intelligent algorithm control of this invention can enable the main steam pressure to track the pressure setpoint, but their dynamic performance differs. Under baseline PID control, the main steam pressure fluctuates significantly, and the pressure curve still exhibits noticeable undulations near the steady state. The four intelligent algorithm control of this invention, through disturbance prediction and feedforward compensation, pre-corrects the control input, thereby reducing the main steam pressure fluctuation and resulting in a smoother pressure response. Figure 5 The middle section shows the valve opening variation. It can be seen that the four intelligent algorithms of this invention maintain the valve within the set upper and lower limits in the small opening range, and avoid frequent micro-movements of the valve through micro-opening suppression and action constraints. Compared to methods that simply rely on PID feedback regulation, this invention introduces valve action amplitude, action frequency, and small opening jitter as optimization constraints or penalty terms into the control process, which helps reduce the risk of valve wear. Figure 5The following section further presents the valve health index HI, the prediction fusion confidence level C(k), the upper limit of adaptive action constraints, and the anomaly backoff flag. Simulation results show that when the valve health status changes, the prediction confidence level fluctuates, or anomalies are triggered, this invention can perform safety arbitration through adaptive constraint adjustment and the anomaly backoff flag, ensuring that the control output does not undergo abrupt changes.

[0129] like Figure 6 As shown, load disturbances, boiler-side input changes, and oil pressure factor fluctuations are introduced during the simulation process to simulate operating scenarios with multiple disturbance sources and different time scales in actual units. Figure 6 The upper part displays the periodic fluctuations and trend changes of load disturbances and boiler-side inputs at different time periods; Figure 6 The middle section shows that the oil pressure factor decreases in some ranges, simulating EH oil pressure fluctuations or reduced execution capability. Figure 6 The lower section shows the changes in the feedforward compensation. It can be seen that when the front-end disturbance and oil pressure state change, the feedforward compensation adjusts accordingly. This indicates that the multi-model disturbance prediction fusion module can predict future pressure trends based on load disturbances, boiler-side inputs, and oil pressure changes, and convert the prediction results into feedforward control quantities. This process allows the control system to no longer rely entirely on feedback correction after pressure deviations occur, but rather to compensate in advance before the disturbance's impact is fully realized, thereby reducing lag in regulation.

[0130] like Figure 7 As shown, typical scenarios such as load increase / decrease, small disturbance, increased disturbance and oil pressure fluctuation, steady state, sudden disturbance, and valve dead zone hysteresis are selected to compare the baseline PID control and the four intelligent algorithm control of this invention. Figure 7 The top left shows a comparison of pressure error indicators. It can be seen that under various operating conditions, the pressure error controlled by the four intelligent algorithms of this invention is lower than that of the baseline PID control, indicating that self-evolutionary modeling, disturbance prediction fusion, and multi-objective optimization can improve pressure regulation accuracy. Figure 7 The top right corner shows a comparison of settling times. The two control methods have similar settling times, indicating that the present invention does not significantly sacrifice dynamic response speed while reducing pressure error and constraining valve action. Figure 7 The bottom left shows a comparison of confidence level indicators. The four intelligent algorithms of this invention maintain a high confidence level under multiple operating conditions, indicating that the operating condition identification and prediction fusion module can provide a relatively stable predictive input for the optimizer. Figure 7The bottom right corner shows a comparison of the number of valve disturbances or actions. In this figure, the number of actions under the control of this invention is higher than the baseline PID. This result should not be simply stated in the specification as "fewer actions," but rather as follows: Under the conditions of limiting amplitude and speed, upper and lower limits of small opening, and safety constraints, this invention improves the regulation participation through active feedforward and constraint optimization, in exchange for lower pressure error and stronger disturbance suppression capability; at the same time, valve action is still limited by minimum action interval, speed limit, and small opening protection constraints to avoid unconstrained high-frequency jitter.

[0131] comprehensive Figures 5 to 7 As can be seen, the four-intelligent algorithm closed-loop control method proposed in this invention can achieve closed-loop operation under scenarios such as multi-source disturbances, oil pressure fluctuations, valve degradation, and abnormal triggering. The SEM self-evolution model is used to reflect changes in valve characteristics, multi-granularity operating condition identification is used to determine the operating state, multi-model disturbance prediction fusion is used to form feedforward compensation, and MAGO multi-objective optimization is used to coordinate pressure control, dynamic response, valve protection, and small opening suppression. Compared with baseline PID control, this invention can reduce main steam pressure error and improve pressure fluctuation suppression under typical operating conditions, and achieve adaptive weighting, constraint adjustment, and safe backoff with the participation of state variables such as prediction confidence, valve health indicators, and abnormal backoff flags. Therefore, this simulation example demonstrates that this invention has good engineering feasibility and adaptability for field operation.

[0132] To address the problems in existing technologies, such as the inability of control valve models to adapt to degradation changes, insufficient prediction of multi-source disturbances, inadequate identification of operating conditions and valve health status, weak multi-objective coordination capabilities, and imperfect engineering safety fallback mechanisms, this application proposes a closed-loop control scheme based on "self-evolutionary modeling—multi-granularity operating condition identification—multi-model disturbance prediction fusion—adaptive multi-objective optimization—DCS safe execution," which can bring the following beneficial effects: 1) Improve the adaptability of the control valve model to the actual operating conditions and reduce the control error caused by model mismatch.

[0133] This application constructs a self-evolving dynamic model of a control valve, incorporating operational degradation factors such as increased valve friction, dead zone expansion, hysteresis aggravation, and EH oil pressure fluctuations into the online modeling process, and continuously updating the model parameters based on operational data. When model residuals, operating conditions change, or valve health indicators reach trigger conditions, the system can switch or merge sub-models, gradually bringing the control model closer to the actual dynamic characteristics of the valve.

[0134] Therefore, compared with fixed model or fixed parameter control methods, this application can reduce the model mismatch problem caused by the drift of valve operating characteristics over a long period of time, and improve the main steam pressure regulation accuracy and long-term control stability.

[0135] 2) Improve the main steam pressure's ability to resist disturbances and reduce lag regulation and overshoot oscillation.

[0136] To address the issue that the main steam pressure is affected by multiple factors such as boiler-side input, load changes, feedwater disturbances, steam quality changes, and EH oil pressure fluctuations, this application sets up a multi-model disturbance prediction fusion mechanism to predict short-term disturbances, medium-term trends, and sudden disturbances, and dynamically adjusts the weights of each predictor based on the operating condition identification results, prediction errors, and confidence levels.

[0137] In this way, the control system no longer relies entirely on passive feedback adjustment after pressure deviation occurs, but can generate feedforward compensation in advance using fused prediction results, providing predictive input for MAGO optimization, thereby reducing regulation lag, pressure overshoot and oscillation, and improving disturbance rejection performance under complex operating conditions.

[0138] 3) Improve the adaptability of the control strategy to different operating conditions and valve health status.

[0139] This application introduces a multi-granularity operating condition identification mechanism, which identifies states such as steady state, load change rate, load change, EH oil pressure, valve position fluctuation and valve health indicators based on data such as main steam pressure, pressure change rate, load change, EH oil pressure, valve position fluctuation and valve health indicators.

[0140] The identification results are not only used to determine the current operating scenario, but also further participate in predictive fusion weight allocation, optimization of target weight adjustment, and action constraint boundary adjustment. Therefore, the control strategy can automatically change the control strength and protection strategy according to the actual operating conditions, reducing frequent manual tuning and improving the system's adaptability and operational convenience under multiple operating conditions.

[0141] 4) It simultaneously ensures pressure stability, response speed, and valve life protection.

[0142] This application incorporates factors such as main steam pressure deviation, dynamic response, valve action amplitude, action frequency, cumulative stroke, and small opening jitter penalty into an adaptive multi-objective optimization framework, and combines valve amplitude and speed limits, minimum action interval, EH oil pressure availability, and safety boundary for constraint solution.

[0143] This avoids the problems of "pursuing rapid pressure stabilization leading to frequent valve actuation" or "overly restricting valve actuation affecting pressure response" in traditional control systems. Through multi-objective coordination, this application can suppress unnecessary high-frequency actuation and small-opening jitter while meeting pressure regulation requirements, thereby reducing valve wear risk and improving valve lifespan.

[0144] 5) Improve the engineering deployability and operational security of intelligent control systems.

[0145] This application adopts a layered architecture consisting of a DCS execution layer, an external intelligent computing layer, and a simulation learning layer. The DCS execution layer retains the original protection interlocking logic, the external intelligent computing layer is responsible for running self-evolutionary modeling, condition identification, predictive fusion, and multi-objective optimization algorithms, and the simulation learning layer is used for parameter tuning and boundary condition verification. This structure avoids the safety risks caused by intelligent algorithms directly replacing DCS protection logic, allowing intelligent control quantities to participate in execution as control commands or correction quantities, and always complying with the original protection constraints of the DCS. Therefore, this application improves control performance while also ensuring field system compatibility and operational safety.

[0146] 6) Improve control reliability under abnormal operating conditions and reduce the risk of sudden changes in control output.

[0147] To address engineering anomalies such as communication timeouts, abnormal measurement points, abnormal model residuals, infeasible optimization, or computation timeouts, this application establishes an anomaly arbitration and non-disruptive rollback mechanism. When an anomaly occurs, the system can maintain the feasible control output from the previous moment, tighten action constraints, or switch to the baseline control strategy without disruption.

[0148] The above mechanism can prevent sudden changes in valve control output caused by external intelligent algorithm malfunctions, ensuring a smooth transition in the control process and improving unit operation safety and field reliability.

[0149] 7) Reduce reliance on manual tuning and improve the convenience of long-term operation and maintenance.

[0150] This application utilizes self-evolutionary model updates, operating condition identification, predictive fusion, and adaptive weighting and constraint adjustment to enable control parameters and action boundaries to automatically adjust according to valve status and operating conditions. Compared to the traditional method of relying on manual experience to repeatedly tune PID parameters, feedforward parameters, and amplitude and speed limiting parameters, this application can reduce the workload of manual trial and error, and improve the long-term consistency and maintenance convenience of the control system.

[0151] In summary, this application improves model adaptability through self-evolutionary models, enhances disturbance prediction capabilities through multi-model prediction fusion, coordinates pressure stabilization, response, and valve protection through multi-objective optimization, and ensures safe project implementation through hierarchical deployment and anomaly rollback mechanisms. Therefore, this application can achieve better main steam pressure regulation performance, stronger adaptability to complex operating conditions, greater valve life friendliness, and more reliable on-site commissioning results in the closed-loop control of turbine regulating valves in large thermal power units.

[0152] like Figure 8 As shown, in a third aspect, embodiments of this application provide a self-evolving closed-loop control device 80 for a steam turbine regulating valve, comprising: The data acquisition module 81 is used to acquire real-time operating data of the steam turbine unit; The update module 82 is used to perform online recursive estimation and structural evolution update of the time-varying degradation parameters of the control valve based on the real-time operating data and through the self-evolving control valve dynamic model, so as to obtain the target evolution model and predicted main steam pressure that reflect the current true dynamic characteristics of the valve. The identification module 83 is used to identify multiple operating conditions based on the real-time operating data, the target evolution model and the predicted main steam pressure, using a multi-granularity operating condition identification algorithm, and output the current operating condition label and confidence level. The prediction module 84 is used to predict future disturbances in parallel by using a multi-model disturbance prediction fusion algorithm based on the current operating condition label, confidence level and real-time operating data, so as to obtain a comprehensive disturbance prediction result and a comprehensive confidence level. Output module 85 is used to perform collaborative solution based on the objective functions of minimizing main steam pressure deviation, smoothing valve action, suppressing valve wear, and suppressing small opening jitter, according to a multi-agent game optimization algorithm, combined with the objective evolution model, comprehensive disturbance prediction results, comprehensive confidence and the current operating condition label, and output a composite control vector. The execution module 86 is used to send the composite control vector to the DCS execution layer for closed-loop control, and to perform abnormal arbitration and uninterrupted rollback when an abnormal fault in system operation or data interaction is detected.

[0153] In one embodiment, optionally, the updating module includes: The first construction unit is used to construct a gray box state-space model including dead-zone nonlinearity and hysteresis nonlinearity, wherein the dead-zone nonlinearity is characterized by a dead-zone function and the hysteresis nonlinearity is characterized by a hysteresis mapping. The constraint unit is used to update the parameters of the gray box state space model online using a recursive least squares method with engineering constraints, and to apply saturation constraints and rate of change constraints to the updated parameters. The model generation unit is used to maintain a model set consisting of multiple candidate sub-models, calculate the residual energy of each candidate sub-model within a sliding window, and dynamically allocate the weights of each candidate sub-model according to the residual energy through a preset fusion formula to generate the target evolution model.

[0154] In one embodiment, optionally, the apparatus further includes: The determination module is used to calculate the valve health index based on the dead zone width parameter, hysteresis parameter and residual energy, and determine whether to update the parameters of the self-evolving control valve dynamic model and whether to reallocate the weights of the candidate sub-models based on the valve health index. The valve health indicators include the current dead zone width parameter, hysteresis parameter, and optimal sub-model residual energy.

[0155] In one embodiment, optionally, the prediction module includes: The prediction output unit is used to call multiple time-scale predictors in parallel and output the perturbation prediction values ​​for a preset number of steps in the future, wherein the multiple time-scale predictors include at least: a short-term predictor, a medium-term predictor and a mutation predictor. The first calculation unit is used to calculate the dynamic weights of each predictor based on the prediction error energy of each predictor within the historical sliding window and the gating function of the current operating condition label. The second calculation unit is used to perform weighted summation of the outputs of the predictors at the multiple time scales to obtain the comprehensive perturbation prediction result, and to calculate the comprehensive confidence level.

[0156] In one embodiment, optionally, the output module includes: The second building unit is used to build a multi-agent decision-making architecture that includes a stress performance agent, a life protection agent, and a dynamic response agent. The third construction unit is used to construct an objective function in the prediction time domain. The objective function includes minimizing the main steam pressure deviation, smoothing valve action, suppressing valve wear, and suppressing small opening jitter, and assigns corresponding weight coefficients to each. The solution unit is used to solve the local optimal control sequence for each agent based on the target evolution model and the comprehensive perturbation prediction results. The adjustment unit is used to adjust the baseline values ​​of each weight coefficient in the objective function according to the current operating condition label, and to dynamically adjust the valve action increment penalty weight and the maximum allowable action increment constraint according to the comprehensive confidence level. The sequence generation unit is used to coordinately adjust each locally optimal control sequence through Nash equilibrium game iteration until the preset system stability constraints and valve action constraints are met, thereby generating the globally optimal control sequence. The vector generator is used to generate the composite control vector based on the globally optimal control sequence, wherein the composite control vector includes: valve position control increment, baseline PID parameter correction, feedforward compensation value, and adaptive micro-opening suppression coefficient.

[0157] In one embodiment, optionally, the execution module includes: The monitoring unit is used to monitor in real time the communication link status between the external intelligent computing layer and the DCS execution layer, the validity of the real-time running data, the residual level of the dynamic model of the self-evolving regulating valve, and the solution status and computation time of the multi-agent game optimization algorithm. The first processing unit is used to maintain the composite control vector generated in the previous cycle unchanged and issue it as the control command for the current cycle when any one or more abnormal states are detected, such as communication timeout, abnormal measurement data, residual of the self-evolving regulating valve dynamic model exceeding the preset threshold, infeasibility of optimization solution or calculation timeout, while reducing the upper limit of valve action change. The second processing unit is used to switch the control authority to the original baseline PID control loop of the DCS execution layer without disturbance if the abnormal state continues for more than a preset time. The correction values ​​in the composite control vector always obey the inherent amplitude limiting, speed limiting and protection interlocking logic of the DCS execution layer.

[0158] In one embodiment, optionally, before collecting the real-time operating data of the steam turbine unit, the method further includes: The verification module is used to perform offline pre-training and verification of the self-evolving regulating valve dynamic model, the multi-granularity working condition identification algorithm, the multi-model perturbation prediction and fusion algorithm, and the multi-agent game optimization algorithm through a simulation learning layer. Specifically, it is used for: Construct a full-condition coupled simulation model of boiler-turbine-regulating valve-EH actuator. The full-condition coupled simulation model includes at least steady-state condition, load increase / decrease condition, disturbance enhancement condition, oil pressure fluctuation condition, and valve dead zone hysteresis increase condition. Pre-set load disturbances, boiler-side input changes, and oil pressure factor fluctuations are injected into the full-condition coupled simulation model to simulate the multi-source disturbance scenario of the actual unit. The self-evolving regulating valve dynamic model, the multi-granularity working condition identification algorithm, the multi-model disturbance prediction and fusion algorithm, and the multi-agent game optimization algorithm are run to iteratively tune the control parameters and record the control performance index under boundary working conditions. The control parameters include the evolution parameters, working condition identification parameters, prediction and fusion parameters, and game optimization parameters in the gray box state space model. When the control performance indicators meet the preset stability, adjustment accuracy and valve life protection requirements, the adjusted control parameters and mapping relationship are solidified and deployed to the external intelligent computing layer.

[0159] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described self-evolutionary closed-loop control method for turbine regulating valves.

[0160] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described self-evolutionary closed-loop control method for turbine regulating valves.

[0161] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the turbine regulating valve self-evolution closed-loop control device and its modules described above can be referred to the corresponding process in the aforementioned embodiment of the turbine regulating valve self-evolution closed-loop control method, and will not be repeated here.

[0162] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the model training device and each module described above can be referred to the corresponding processes in the aforementioned embodiments of the self-evolution closed-loop control method for turbine regulating valves, and will not be repeated here.

[0163] The aforementioned self-evolving closed-loop control device for turbine regulating valves can be implemented as a computer program, which can, for example... Figure 9 It runs on the computer device shown.

[0164] Figure 9 A block diagram of a computer device according to one embodiment of this application is shown.

[0165] See Figure 9 The computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include storage media and internal memory.

[0166] The storage medium may store an operating system and a computer program. The computer program includes program instructions that, when executed, cause the processor to perform any of the multi-source data self-evolutionary closed-loop control methods for turbine regulating valves provided in the embodiments of this application.

[0167] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0168] The internal memory provides an environment for the execution of a computer program stored in the storage medium. When executed by a processor, this program enables the processor to perform any multi-source data self-evolutionary closed-loop control method for turbine regulating valves. The storage medium can be non-volatile or volatile.

[0169] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0170] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0171] In addition, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions for performing the steps of the method in the first aspect embodiment.

[0172] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or electronic device described above can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0173] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0174] It should be understood that although the terms "first," "second," etc., may be used to describe the setting units in the embodiments of this application, these setting units should not be limited to these terms. These terms are only used to distinguish the setting units from each other. For example, without departing from the scope of the embodiments of this application, the first setting unit may also be referred to as the second setting unit, and similarly, the second setting unit may also be referred to as the first setting unit.

[0175] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0176] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0177] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in a combination of hardware and software functional units.

[0178] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0179] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A self-evolving closed-loop control method for a steam turbine regulating valve, characterized in that, The method includes: Collect real-time operating data of the steam turbine unit; Based on the real-time operating data, the time-varying degradation parameters of the control valve are recursively estimated and the structure is updated online through the self-evolving control valve dynamic model, so as to obtain the target evolution model and the predicted main steam pressure that reflect the current valve's true dynamic characteristics. Based on the real-time operating data, the target evolution model, and the predicted main steam pressure, a multi-granularity operating condition identification algorithm is used to identify the current operating condition and output the current operating condition label and confidence level. Based on the current operating condition label, confidence level, and real-time operating data, a multi-model disturbance prediction fusion algorithm is used to call multiple time scale predictors in parallel to estimate future disturbances, so as to obtain a comprehensive disturbance prediction result and a comprehensive confidence level. With the objective functions of minimizing main steam pressure deviation, smoothing valve action, suppressing valve wear, and suppressing small opening jitter, a composite control vector is output by collaboratively solving the problem based on a multi-agent game optimization algorithm, combined with the objective evolution model, comprehensive disturbance prediction results, comprehensive confidence level, and the current operating condition label. The composite control vector is sent to the DCS execution layer for closed-loop control, and when an abnormal fault in system operation or data interaction is detected, abnormal arbitration and uninterrupted rollback are performed.

2. The method according to claim 1, characterized in that, Online recursive estimation and structural evolution updates of the time-varying degradation parameters of the control valve are performed using a self-evolving dynamic model, including: A gray-box state-space model is constructed that includes dead-zone nonlinearity and hysteresis nonlinearity, wherein the dead-zone nonlinearity is characterized by a dead-zone function and the hysteresis nonlinearity is characterized by a hysteresis mapping. The parameters of the gray box state space model are updated online using a recursive least squares method with engineering constraints, and saturation constraints and rate of change constraints are applied to the updated parameters. Maintain a model set consisting of multiple candidate sub-models, calculate the residual energy of each candidate sub-model within a sliding window, and dynamically allocate the weights of each candidate sub-model according to the residual energy using a preset fusion formula to generate the target evolution model.

3. The method according to claim 2, characterized in that, The method further includes: Based on the dead zone width parameter, hysteresis parameter and residual energy, the valve health index is calculated, and based on the valve health index, it is determined whether to update the parameters of the self-evolving control valve dynamic model and whether to reallocate the weights of the candidate sub-models. The valve health indicators include the current dead zone width parameter, hysteresis parameter, and optimal sub-model residual energy.

4. The method according to claim 1, characterized in that, The method employs a multi-model perturbation prediction fusion algorithm, which calls predictors at multiple time scales in parallel to estimate future perturbations, including: Multiple time-scale predictors are invoked in parallel to output the perturbation prediction value for a preset number of steps in the future. The multiple time-scale predictors include at least a short-term predictor, a medium-term predictor, and a mutation predictor. Based on the prediction error energy of each predictor within the historical sliding window and the gating function of the current operating condition label, the dynamic weights corresponding to each predictor are calculated. The outputs of the predictors at the multiple time scales are weighted and summed to obtain the comprehensive perturbation prediction result, and the comprehensive confidence level is calculated.

5. The method according to claim 1, characterized in that, With the objective functions of minimizing main steam pressure deviation, smoothing valve action, suppressing valve wear, and suppressing small-opening jitter, a multi-agent game optimization algorithm is used to collaboratively solve the problem, combining the objective evolution model, comprehensive disturbance prediction results, comprehensive confidence level, and the current operating condition label. The resulting composite control vector includes: Construct a multi-agent decision-making architecture that includes a stress performance agent, a life protection agent, and a dynamic response agent; Within the prediction time domain, an objective function is constructed, which includes minimizing the main steam pressure deviation, smoothing valve action, suppressing valve wear, and suppressing small opening jitter, and each is assigned a corresponding weight coefficient. Based on the target evolution model and the comprehensive perturbation prediction results, the local optimal control sequence for each agent is solved respectively; The baseline values ​​of each weight coefficient in the objective function are adjusted according to the current operating condition label, and the valve action increment penalty weight and the maximum allowable action increment constraint are dynamically adjusted according to the comprehensive confidence level. Through Nash equilibrium game iteration, each local optimal control sequence is coordinated and adjusted until the preset system stability constraints and valve action constraints are met, thereby generating the global optimal control sequence. Based on the globally optimal control sequence, the composite control vector is generated, wherein the composite control vector includes: valve position control increment, baseline PID parameter correction, feedforward compensation value, and adaptive micro-opening suppression coefficient.

6. The method according to claim 1, characterized in that, The step of performing abnormal arbitration and uninterrupted rollback when a system operation or data interaction anomaly is detected includes: Real-time monitoring of the communication link status between the external intelligent computing layer and the DCS execution layer, the validity of the real-time running data, the residual level of the self-evolving regulating valve dynamic model, and the solution status and computation time of the multi-agent game optimization algorithm; When any one or more abnormal states are detected, such as communication timeout, abnormal measurement data, residual of the self-evolving regulating valve dynamic model exceeding the preset threshold, infeasibility of optimization solution, or calculation timeout, the composite control vector generated in the previous cycle remains unchanged and is issued as the control command for the current cycle, while reducing the upper limit of valve action change. If the abnormal state continues for more than a preset time, the control authority will be switched seamlessly to the original baseline PID control loop of the DCS execution layer, and the correction values ​​in the composite control vector will always obey the inherent amplitude limiting, speed limiting and protection interlocking logic of the DCS execution layer.

7. The method according to claim 1, characterized in that, Before collecting the real-time operating data of the steam turbine unit, the method further includes: The self-evolving control valve dynamic model, the multi-granularity operating condition identification algorithm, the multi-model perturbation prediction and fusion algorithm, and the multi-agent game optimization algorithm are pre-trained and validated offline through a simulation learning layer, specifically including: Construct a full-condition coupled simulation model of boiler-turbine-regulating valve-EH actuator. The full-condition coupled simulation model includes at least steady-state condition, load increase / decrease condition, disturbance enhancement condition, oil pressure fluctuation condition, and valve dead zone hysteresis increase condition. Pre-set load disturbances, boiler-side input changes, and oil pressure factor fluctuations are injected into the full-condition coupled simulation model to simulate the multi-source disturbance scenario of the actual unit. The self-evolving regulating valve dynamic model, the multi-granularity working condition identification algorithm, the multi-model disturbance prediction and fusion algorithm, and the multi-agent game optimization algorithm are run to iteratively tune the control parameters and record the control performance index under boundary working conditions. The control parameters include the evolution parameters, working condition identification parameters, prediction and fusion parameters, and game optimization parameters in the gray box state space model. When the control performance indicators meet the preset stability, adjustment accuracy and valve life protection requirements, the adjusted control parameters and mapping relationship are solidified and deployed to the external intelligent computing layer.

8. A self-evolving closed-loop control device for a steam turbine regulating valve, characterized in that, The device includes: The data acquisition module is used to collect real-time operating data of the steam turbine unit; The update module is used to perform online recursive estimation and structural evolution update of the time-varying degradation parameters of the control valve based on the real-time operating data and through the self-evolving control valve dynamic model, so as to obtain the target evolution model and predicted main steam pressure that reflect the current true dynamic characteristics of the valve. The identification module is used to identify multiple operating conditions based on the real-time operating data, the target evolution model, and the predicted main steam pressure, using a multi-granularity operating condition identification algorithm, and output the current operating condition label and confidence level. The prediction module is used to predict future disturbances by using a multi-model disturbance prediction fusion algorithm based on the current operating condition label, confidence level and real-time operating data, and calling multiple time scale predictors in parallel to obtain a comprehensive disturbance prediction result and a comprehensive confidence level. The output module is used to take minimizing the main steam pressure deviation, smoothing valve action, suppressing valve wear, and suppressing small opening jitter as objective functions. Based on the multi-agent game optimization algorithm, it combines the objective evolution model, comprehensive disturbance prediction results, comprehensive confidence and the current operating condition label to perform collaborative solution and output a composite control vector. The execution module is used to send the composite control vector to the DCS execution layer for closed-loop control, and to perform abnormal arbitration and uninterrupted rollback when an abnormal fault in system operation or data interaction is detected.

9. A computer device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, the instructions being configured to perform the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The device stores computer-executable instructions for performing the method as described in any one of claims 1 to 7.

Images