A flexible gain scheduling control method for a deep peak regulation unit

Abstract

The application discloses a flexible gain scheduling control method for a deep peak-regulating unit, which solves the problem that a single model cannot accurately describe the time-varying dynamic characteristics of the unit in a wide load range by performing open-loop step response tests at multiple typical working points and identifying a transfer function model. The method adopts a model integrated active disturbance rejection controller, integrates the information of the identified transfer function model into an extended state observer, so that the extended state observer can more accurately identify the inherent dynamic characteristics of the system. For a deterministic dynamic link, the control command can be directly fed forward through the model information, reducing the estimation lag of the observer and greatly improving the response speed of the load tracking, pressure regulation and other control tasks. For nonlinear or time-varying characteristics, the model information can assist the extended state observer to more accurately separate the internal dynamics from the external disturbance / model mismatch, thereby improving the control accuracy.

Description

Technical Field

[0001] This invention belongs to the field of deep peak-shaving unit control technology, and particularly relates to a flexible gain scheduling control method for deep peak-shaving units. Background Technology

[0002] With the increasing proportion of new energy grid connection, thermal power units need to undertake deep peak shaving tasks through a coordination system (also known as a "coordination control system"). These thermal power units are also called deep peak shaving units, whose operating load needs to frequently switch within the range of 30%-100% of rated load, and they face internal and external disturbances (such as fluctuations in coal feed and main steam pressure). The coordination system refers to the boiler-turbine coordination control system, which is the core of the deep peak shaving unit control. It receives load commands from the grid and coordinates the control of the boiler side (such as coal feed, water feed, and air volume regulation) and the turbine side (such as turbine valve opening regulation). It must achieve rapid response of the unit to load commands while maintaining stable internal parameters such as main steam pressure and separator outlet temperature. It is a key system for deep peak shaving units to adapt to wide load operation and cope with disturbances. Existing deep peak shaving unit coordination control systems mainly adopt the following schemes:

[0003] Traditional PID control relies on fixed parameters and struggles to adapt to changes in the dynamic characteristics of the unit over a wide load range. It is prone to problems such as large overshoot, long settling time, and poor disturbance rejection performance. For example, the existing technology with publication number "CN115327890A" provides a method for optimizing the main steam pressure of a thermal power deep peak-shaving unit using an improved population search algorithm. Although this method optimizes PID parameters through a reverse differential evolution mechanism and an adaptive t-distribution strategy, it is still essentially based on the PID control framework. When frequently switching between 30% and 100% rated load, the optimized PID parameters cannot adapt to the dynamic characteristics of the unit's main steam pressure changing with load in real time, and regulation lag may still occur under low-load conditions.

[0004] Fixed parameter active disturbance rejection control: Although it can suppress disturbances, it does not incorporate unit model information and the parameters are fixed values, which cannot cope with frequent fluctuations in operating conditions. The control quality deteriorates significantly under low load. For example, the prior art with publication number "CN116719232A" provides an intelligent control strategy for the boiler-generator coupling process of ultra-supercritical units based on cascaded error active disturbance rejection control. This method constructs a cascaded ESO to improve the disturbance estimation accuracy, but it does not fully integrate the three-input three-output model information of the boiler-generator coupling system. Moreover, the core parameters such as ESO bandwidth and controller bandwidth are fixed configurations. When the load changes frequently, it cannot adapt to the nonlinear dynamic changes of the unit, and the main steam pressure control deviation increases under low load. The prior art with publication number "CN113391552A" provides a cascaded adaptive volumetric Kalman active disturbance rejection control method to improve the flexible operation of thermal power units. This method cascades adaptive volumetric Kalman filtering with EADRC, but the ACKF-EADRC controller also requires manual parameter tuning and does not take into account the differences in the dynamic model of the unit under different loads. Under disturbances such as coal feed fluctuations and sudden changes in load commands, the active power tracking accuracy and main steam pressure stability decrease significantly under low load.

[0005] Control combining mechanism modeling and system identification: Mechanism modeling has problems such as large errors and complex structures, while system identification is only used for modeling a single operating condition and cannot achieve dynamic parameter scheduling, making it difficult to meet the requirements of precise control under wide loads.

[0006] The above schemes result in insufficient control accuracy and poor robustness of the deep-load generator units during wide-load operation and operating condition switching, making it impossible to meet the peak-shaving control requirements under the grid connection of new energy. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention proposes a flexible gain scheduling control method for deep peak-shaving units.

[0008] The technical solution of the present invention is as follows:

[0009] A flexible gain dispatch control method for a deep peak-shaving unit, wherein the deep peak-shaving unit is a generator unit that adjusts its output within a wide load range and meets the deep peak-shaving requirements of the power grid through a coordination system, the method comprising:

[0010] Open-loop step response tests were conducted under multiple typical operating conditions to identify the transfer function model of the coordinated system under each typical operating condition; each typical operating condition has a corresponding load setpoint.

[0011] Based on the transfer function models of various typical operating conditions, a model-integrated active disturbance rejection controller is designed for each typical operating condition. The model-integrated active disturbance rejection controller estimates and compensates for the total disturbance in real time through an extended state observer with transfer function model information.

[0012] Based on the transfer function model, a closed-loop control loop is constructed, and the particle swarm optimization algorithm is used to tune the parameters of the model integrated active disturbance rejection controller for each typical working condition, so as to obtain the optimal controller parameters corresponding to each typical working condition.

[0013] The load setpoint corresponding to each typical working condition is used as input, and the optimal controller parameters corresponding to each typical working condition are used as output to train a Gaussian regression process model to construct the mapping relationship between load commands and controller parameters, so as to realize the real-time scheduling of model-integrated active disturbance rejection controller parameters under changing working conditions.

[0014] The optimal controller parameters for the target operating condition are output from the trained Gaussian regression process model, and flexible gain scheduling control of the deep peak-shaving unit is performed.

[0015] Furthermore, before conducting open-loop step response tests under multiple typical operating conditions, the collected raw data is preprocessed. The preprocessing includes removing outliers using the 3σ criterion and smoothing the data using the moving average method.

[0016] Furthermore, the specific method for the open-loop step response test includes: stabilizing the deep peak-shaving unit under typical operating conditions, switching to basic operating mode, disconnecting the controller and engaging manual mode; applying a positive step disturbance to the input control quantity, and after stabilization, applying a reverse step disturbance of the same amplitude; recording the input control quantity and output controlled quantity throughout the entire test process. The input control quantity includes coal feed rate uB, feedwater flow rate dfw, and turbine valve opening ut, while the output controlled quantity includes the unit's actual generated power Ne, main steam pressure pst, and steam-water separator specific enthalpy hm.

[0017] Furthermore, the coordination system is a three-input three-output system, and the input-output pairing relationship is as follows: turbine control valve opening ut - actual unit power Ne, coal feed rate uB - main steam pressure pst, feedwater flow rate dfw - steam specific enthalpy hm of steam-water separator.

[0018] Furthermore, the transfer function model of the coordination system under various typical working conditions is identified using a swarm intelligence optimization algorithm. The transfer function of the nine channels formed by the three-input three-output pairing of the coordination system is identified separately, and the structure of the identified transfer function model is uniformly selected as a second-order system.

[0019] Furthermore, the expression for the extended state observer with transfer function model information in the model-integrated active disturbance rejection controller is as follows:

[0020] ;

[0021] in,

[0022] In the formula, , and These are the first, second, and third gains of the extended state observer, respectively. To expand the bandwidth of the state observer; , and These are the observed values ​​of the output variable y, the first derivative of the output variable y, and the total disturbance f, respectively. , and They are respectively , and The first derivative; The actual output of the controlled object; For the input of the controlled object; and These are the 0th and 1st model parameters, respectively, which contain information about the transfer function model. This is the estimated value of the model gain.

[0023] Furthermore, in the expression of the extended state observer with transfer function model information, the 0th model parameter... and the first model parameters The dynamic equations for the transfer function model of the controlled object are obtained by establishing the transfer function model of the controlled object. The dynamic equation is established as follows: In the formula, The steady-state gain of the controlled object; For inertial time; For Lagrangian operators; and The actual output of the controlled object The first and second derivatives; This is the sum of unmodeled dynamics and disturbances.

[0024] Furthermore, the specific method for tuning the parameters of the model-integrated active disturbance rejection controller includes:

[0025] The state-space expression of the extended state observer, which contains transfer function model information, is transformed into a transfer function using Laplace transform. The equivalent transfer functions of the virtual control quantity U0(s) and the output Y(s) are then derived. ;

[0026] Therefore, the expected closed-loop dynamic tracking transfer function is: In the formula, r c The expected tracking factor;

[0027] The equivalent closed-loop dynamic tracking transfer function is ;

[0028] The feedback gain is solved by comparing equations, where the proportional gain is... Differential gain .

[0029] Furthermore, the specific methods of the particle swarm optimization algorithm include:

[0030] Treating the loop coupling of the coordination system as the total disturbance and neglecting the off-diagonal elements of the transfer function matrix, model-integrated active disturbance rejection controllers (ADCs) are designed for the actual unit power output Ne, main steam pressure pst, and steam-water separator enthalpy hm loops. Each ADC contains five parameters to be optimized, namely the model gain estimate. Bandwidth of the extended state observer Expected tracking factor Model 0 parameters and the first model parameters ;

[0031] Design independent objective functions, including power control loops. Main steam pressure control circuit Enthalpy control loop ;in, , and These are the overshoot amounts for load, main steam pressure, and separator specific enthalpy, respectively. , and These are the adjustment times for load, main steam pressure, and separator specific enthalpy, respectively.

[0032] A parallel optimization approach is adopted to generate particle populations for the three integrated active disturbance rejection controllers, and calculate fitness and convergence judgment separately without interference between them.

[0033] Furthermore, the Gaussian regression process model selects the squared exponential kernel function, and predicts the parameters to be optimized for the model-integrated active disturbance rejection controller using the covariance matrix between sample loads under typical operating conditions and the covariance vector between the load setpoint and the sample load; wherein the covariance matrix corresponds to any sample load Nesp i With another sample load Nesp j The covariance is: In the formula, l represents the signal variance; l represents the length scale. The standard deviation of noise; For the Kronecker function;

[0034] For load setting value With sample loading The covariance is ;

[0035] The prediction method for predicting the parameters to be optimized in a model-integrated active disturbance rejection controller is as follows:

[0036] , , , and ,in For covariance The covariance matrix is ​​composed of elements. covariance The covariance vector matrix composed of elements; , , , and These are the model gain estimates. Bandwidth of the extended state observer Expected tracking factor Model 0 parameters and the first model parameters The predicted value; These are the model gain estimates. Bandwidth of the extended state observer Expected tracking factor Model 0 parameters and the first model parameters The column vector formed by these.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] This invention proposes a flexible gain scheduling control method for deep peak-shaving units. This method solves the problem of time-varying dynamic characteristics of units over a wide load range, which cannot be accurately described by a single model, by conducting open-loop step response tests at multiple typical operating points and identifying the transfer function model. This enables the acquisition of an accurate object model covering 30%-100% of the rated load range, laying a solid foundation for the design of subsequent high-performance controllers and thus improving model accuracy and operating condition adaptability.

[0039] This invention employs a Model Integrated Active Disturbance Rejection Controller (MADRC). This MADRC integrates the identified transfer function model information into the extended state observer (ESO), enabling the ESO to more accurately identify the inherent dynamic characteristics of the system. For deterministic dynamic elements (such as known model components), control commands can be directly fed forward through model information, reducing the observer's estimation lag and significantly improving the response speed of control tasks such as load tracking and pressure regulation. For nonlinear or time-varying characteristics (such as combustion instability under low load conditions), model information can assist the ESO in more accurately separating "intra-model dynamics" from "external disturbances / model mismatch," thereby improving control accuracy. Compared to MADRCs with fixed parameters or without model information, this invention better adapts to the operating characteristics of deep peak-shaving units, which feature "large operating range and time-varying dynamic characteristics," achieving superior control performance over a wide load range.

[0040] This invention employs a strategy based on particle swarm optimization (PSO) algorithm to tune controller parameters in a closed-loop simulation circuit. This technique transforms the complex multi-parameter, multi-objective tuning problem into a highly efficient and automated optimization process, overcoming the drawbacks of manual tuning, which relies on experience, is time-consuming and labor-intensive, and struggles to obtain optimal solutions. By performing parallel and independent optimization of the power, pressure, and specific enthalpy circuits, it ensures that each circuit achieves optimal control parameters under typical operating conditions, significantly reducing system overshoot and shortening settling time.

[0041] This invention employs a Gaussian process regression (GPR) model, leveraging GPR's powerful nonlinear fitting capabilities to learn and establish a nonlinear mapping relationship between the "load setpoint" and the "optimal controller parameters." This solves the problem that fixed-parameter controllers cannot adapt to frequent fluctuations in operating conditions, enabling the control system to smoothly and in real-time dispatch a set of control parameters best suited to the current operating conditions when the unit's load command changes. Ultimately, adaptive optimization control is achieved across the entire operating range, ensuring that the unit maintains high precision, fast response, and strong robustness during deep peak shaving. Attached Figure Description

[0042] Figure 1 This is a diagram of the coupling model of the coordination system in the embodiment;

[0043] Figure 2 This is a block diagram of the active disturbance rejection control structure with model information in the embodiment;

[0044] Figure 3 This is a flowchart of the MADRC controller parameter PSO optimization process in the embodiment;

[0045] Figure 4 This is a diagram of the GPR parameter scheduling control loop structure in the embodiment;

[0046] Figure 5 This is a comparison chart of the identification model output and actual open-loop data under 50% rated load in the application example;

[0047] Figure 6(a) is one of the schematic diagrams showing the influence of MADRC controller parameters on the control effect of the transfer function identified under 50% rated load in the application embodiment. In Figure 6(a), the other four parameters are fixed. Take different values;

[0048] Figure 6(b) is the second schematic diagram showing the influence of MADRC controller parameters on the control effect of the transfer function identified under 50% rated load in the application embodiment. In Figure 6(b), the other four parameters are fixed. Take different values;

[0049] Figure 6(c) is the third schematic diagram showing the influence of MADRC controller parameters on the control effect of the transfer function identified under 50% rated load in the application embodiment. In Figure 6(c), the other four parameters are fixed. Take different values;

[0050] Figure 6(d) is the fourth schematic diagram showing the influence of MADRC controller parameters on the control effect of the transfer function identified under 50% rated load in the application example. In Figure 6(d), the other four parameters are fixed. Take different values;

[0051] Figure 6(e) is the fifth schematic diagram showing the influence of MADRC controller parameters on the control effect of the transfer function identified under 50% rated load in the application embodiment. In Figure 6(e), the other four parameters are fixed. Take different values;

[0052] Figure 7(a) is one of the output response curves and control signal diagrams of different control strategies under 50% rated load in the application embodiment. Figure 7(a) corresponds to the main steam pressure comparison.

[0053] Figure 7(b) is one of the output response curves and control signal diagrams of different control strategies under 50% rated load in the application embodiment. Figure 7(b) corresponds to the comparison of the specific enthalpy of the separator steam.

[0054] Figure 7(c) is one of the output response curves and control signal diagrams of different control strategies under 50% rated load in the application embodiment. Figure 7(c) corresponds to the load comparison.

[0055] Figure 8(a) is one of the output line diagrams of different control algorithms based on the simulation platform under varying operating conditions in the application embodiment. Figure 8(a) is the actual load output curve.

[0056] Figure 8(b) is the second schematic diagram of the output line of different control algorithms under varying operating conditions based on the simulation platform in the application embodiment. Figure 8(b) corresponds to the actual main steam pressure output curve.

[0057] Figure 8(c) is one of the output line diagrams of different control algorithms based on the simulation platform under varying operating conditions in the application embodiment. Figure 8(c) corresponds to the actual steam specific enthalpy output curve of the separator. Detailed Implementation

[0058] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0059] Example 1:

[0060] This invention discloses a flexible gain dispatch control method for deep peak-shaving generating units. Deep peak-shaving generating units refer to generating units that can rapidly adjust their output within a wide load range to meet the deep peak-shaving requirements of the power grid. Their deep peak-shaving objectives are achieved through a coordination system. This method treats the coordination system of the deep peak-shaving generating unit as the controlled object, and includes the following steps:

[0061] Open-loop step response tests were conducted under multiple typical operating conditions to identify the transfer function model of the coordinated system under each typical operating condition; each typical operating condition has a corresponding load setpoint.

[0062] Based on the transfer function models of various typical working conditions, a model-integrated active disturbance rejection controller (ADRC) corresponding to each typical working condition is designed. The model-integrated ADRC estimates and compensates for the total disturbance in real time through an extended state observer with transfer function model information.

[0063] A closed-loop control loop is constructed based on the transfer function model, and the particle swarm optimization algorithm is used to tune the parameters of the model integrated active disturbance rejection controller for each typical working condition, so as to obtain the optimal controller parameters corresponding to each typical working condition.

[0064] The Gaussian regression process model is trained by taking the load setpoint corresponding to each typical working condition as input and the optimal controller parameters corresponding to each typical working condition as output, so as to build the mapping relationship between load command and controller parameters and realize the real-time scheduling of model integrated active disturbance rejection controller parameters under changing working conditions.

[0065] The optimal controller parameters for the target operating condition are output from the trained Gaussian regression process model, and flexible gain scheduling control of the deep peak-shaving unit is performed.

[0066] Example 2:

[0067] This embodiment is further designed based on Embodiment 1. In this example, before conducting open-loop step response tests under multiple typical operating conditions, the collected raw data is preprocessed. The preprocessing includes using the 3σ criterion to remove outliers and using the moving average method to smooth the data.

[0068] Furthermore, the 3σ criterion is used to eliminate outlier data points caused by simulation fluctuations, where σ is the standard deviation of the data;

[0069] The moving average method can be specifically used to smooth the preprocessed data by employing a 5-point moving average method, thereby eliminating high-frequency noise.

[0070] Example 3:

[0071] This embodiment, based on Embodiment 1, further designs the following: the specific method for the open-loop step response test includes: stabilizing the deep peak-shaving unit under typical operating conditions; switching the unit's operating mode to basic operation; disconnecting the controller and engaging manual operation; applying a positive step disturbance to the input control quantity; and applying a reverse step disturbance of the same magnitude after the coordinated system stabilizes; recording the input control quantity and output controlled quantity throughout the test process; and finally identifying the complete transfer function model of the controlled object in the coordinated system. The input control quantity includes coal feed rate uB, feedwater flow rate dfw, and turbine valve opening ut; the output controlled quantity includes the unit's actual generated power Ne, main steam pressure pst, and steam-water separator specific enthalpy hm.

[0072] Example 4:

[0073] This embodiment, based on Embodiment 1, further designs the coordination system as a three-input, three-output system. The input control variables are coal feed rate uB, feedwater flow rate dfw, and turbine valve opening ut; the output controlled variables are actual unit power output Ne, main steam pressure pst, and steam-water separator enthalpy hm. To ensure the variable load rate, the turbine valve opening ut should be paired with the actual unit power output Ne. From the perspective of the unit's dynamic characteristics, selecting fuel feed rate uB and main steam pressure pst as a pair of input and output variables is beneficial for maintaining stable main steam pressure. Finally, feedwater flow rate dfw and steam-water separator enthalpy hm are selected as a pair of input and output variables. The coupling model of the coordination system is as follows: Figure 1 As shown.

[0074] The final input-output pairing relationship is as follows: turbine valve opening ut - actual unit power Ne (to ensure load change rate), coal feed uB - main steam pressure pst (to maintain pressure stability), and feedwater flow rate dfw - steam-water separator steam specific enthalpy hm (to control steam-water quality).

[0075] Example 5:

[0076] This embodiment is a further design based on embodiment four. In this example, the transfer function model of the coordination system under various typical working conditions is identified using a swarm intelligence optimization algorithm. The transfer function of the nine channels formed by the three-input three-output pairing of the coordination system is identified separately, and the structure of the identified transfer function model is uniformly selected as a second-order system.

[0077] Furthermore, the nine channels include:

[0078] Turbine valve opening ut - actual unit power output Ne, turbine valve opening ut - main steam pressure pst, turbine valve opening ut - steam enthalpy of steam separator hm, coal feed rate uB - main steam pressure pst, coal feed rate uB - actual unit power output Ne, coal feed rate uB - steam enthalpy of steam separator hm, feedwater flow rate dfw - steam enthalpy of steam separator hm, feedwater flow rate dfw - main steam pressure pst, and feedwater flow rate dfw - actual unit power output Ne.

[0079] The parameterized representations are "uB→pst", "uB→Ne", "uB→hm", "dfw→hm", "dfw→pst", "dfw→Ne", "ut→Ne", "ut→pst", and "ut→hm".

[0080] Example 6:

[0081] This control structure adopts a control architecture that includes a linear error feedback control law, an extended state observer (MESO), and the controlled object. The functions and signal flows of each part are as follows:

[0082] The reference input signal r(t) is divided into two paths: one path is processed by gain a0, and the other path is subtracted from the output y(t) of the controlled object and then passed through a proportional element k. p Processing yields the intermediate virtual control quantity u0(t);

[0083] The intermediate virtual control quantity u0(t) and the outputs Z3(t) and Z2(t) of the extended state observer (MESO) are compared through the differential gain k. d The difference is calculated after feedback, and then the input u(t) of the controlled object is obtained through a 1 / b0 gain circuit;

[0084] Controlled object G p It receives input u(t) and external disturbance d(t), and outputs system output y(t);

[0085] The Extended State Observer (MESO) receives the input u(t) and output y(t) of the controlled object and outputs observation signals Z1(t), Z2(t), and Z3(t).

[0086] This embodiment is further designed based on Embodiment 1, as follows: Figure 2As shown, in this example, the expression for the Extended State Observer (MESO) with transfer function model information in the Model Integrated Active Disturbance Rejection Controller is:

[0087] ;

[0088] in,

[0089] In the formula, , and These are the first, second, and third gains of the extended state observer, respectively, and represent the polynomial coefficients. To expand the bandwidth of the state observer; , and These are the observed values ​​of the output variable y, the first derivative of the output variable y, and the total disturbance f, respectively, representing the first, second, and third outputs of the extended state observer; , and They are respectively , and The first derivative; The actual output of the controlled object; The input to the controlled object, i.e., the input to the coordination system; and For the 0th and 1st model parameters, which respectively possess transfer function model information, further, , ; The estimated value of the model gain b is an adjustable key parameter. Indicates the time.

[0090] Furthermore, the expression for the linear state feedback law in the model-integrated active disturbance rejection controller is:

[0091] ;

[0092] In the formula, k p k d Let r be the proportional gain and the derivative gain, respectively, and r be the set value.

[0093] Example 7:

[0094] This embodiment, based on Embodiment Six, is further designed in that the 0th model parameter in the expression of the extended state observer with transfer function model information in this example... and the first model parameters The dynamic equations are obtained by establishing the transfer function model of the controlled object (second-order system). In the formula, The steady-state gain of the controlled object; For inertial time; For Lagrangian operators;

[0095] The dynamic equation is established as follows ;

[0096] and The outputs of the controlled object are respectively The first and second derivatives; This is the sum of unmodeled dynamics and disturbances. ; and These represent the uncertain parts of the two model parameters, respectively. This represents the uncertain part of the model gain; External disturbance;

[0097] Furthermore, the above dynamic equations are transformed into extended state equations with model information:

[0098] ;

[0099] in, ;

[0100] In the formula, , and These are the first, second, and third state variables of the controlled object, respectively, and represent the output variable y, the first derivative of the output variable y, and the total disturbance f, respectively; , and They are respectively , and The first derivative; for First derivative; superscript This is the transpose marker.

[0101] Therefore, the corresponding extended state observer can be designed as follows:

[0102]

[0103] in, Indicates the observer gain; z i For state variable x i The estimated state.

[0104] At this point, the characteristic equation of the extended state observer can be expressed as:

[0105]

[0106] To achieve ideal estimation performance, the observer poles can be placed at the same location. Therefore, the design rules for the observer gain can be derived as follows:

[0107]

[0108] When the state estimation is effective, the final control action can be derived based on the linear state feedback law as follows:

[0109]

[0110] Where, k p k d These are the proportional gain and the derivative gain, respectively.

[0111] Example 8:

[0112] This embodiment, based on Embodiment Seven, further designs the parameter tuning method for the model-integrated active disturbance rejection controller, which includes:

[0113] The state-space expression of the Extended State Observer (MESO) with transfer function model information is transformed into a transfer function using Laplace transform, yielding the expressions for Z1 and Z2 as follows:

[0114]

[0115] In the formula, and These are the input signal and the output signal of the controlled object, respectively. , and These represent the observed values ​​of the system output y(t) and the first derivative of the system output, respectively. The observed values ​​of the system and the observed values ​​of the total system disturbance (the sum of unmodeled dynamics and external disturbances);

[0116] Combining the above estimated states The equivalent transfer functions of the virtual control quantity U0(s) and the output Y(s) can then be derived. ; Pass functions to the controlled object;

[0117] according to Then, referring to the equivalent transfer function between the input signal R(s) and the output Y(s), that is... To rationally configure the proportional and integral actions, for a delay-free object, the closed-loop dynamic tracking transfer function of the linear second-order integrated active disturbance rejection controller can be specified as follows: In the formula, r c The expected tracking factor;

[0118] By comparing the two formulas above, we can obtain the following relationship. ;

[0119] The feedback gain is solved by comparing equations, where the proportional gain is... Differential gain .

[0120] Therefore, by using the expected tracking factor r c By adjusting the feedback gain, the closed-loop dynamic tracking transfer function of MADRC can be shaped into the desired transfer function.

[0121] Example 9:

[0122] This embodiment is further designed based on embodiment eight, as follows: Figure 3 As shown, the specific methods of the particle swarm optimization algorithm in this example include:

[0123] Based on the transfer functions identified above, control loops are constructed. However, since the coordination system is a three-input, three-output coupled system, the loop coupling of the coordination system is considered as the total disturbance, which is then estimated and compensated. In the process of designing MESOs for each loop individually for observation, off-diagonal elements of the transfer function matrix are ignored, and MADRC controller design is performed only for the diagonal transfer functions. Model-integrated active disturbance rejection controllers (MADRCs) are designed for the unit's actual power output Ne, main steam pressure pst, and steam-water separator enthalpy hhm loops respectively. Each MADRC contains five parameters to be optimized, namely the model gain estimate. Bandwidth of the extended state observer Expected tracking factor Model 0 parameters and the first model parameters The objective function of each sub-controller contains only the "overshoot" and "set-up time" of its own output, and assigns weights based on the process priority of that output.

[0124] Design independent objective functions, including the MADRC controller objective function (power control loop) for controlling the load. Objective function of MADRC controller for controlling main steam pressure (main steam pressure control loop) The objective function of the MADRC controller (specific enthalpy control loop) for controlling the specific enthalpy of the separator steam. ;in, , and These are the overshoot amounts for load, main steam pressure, and separator specific enthalpy, respectively. , and These are the adjustment times for load, main steam pressure, and separator specific enthalpy, respectively.

[0125] Since the three objective functions are independent, particle swarm optimization adopts a parallel optimization approach. Independent particle populations are generated for each of the three model-integrated active disturbance rejection controllers, and fitness and convergence criteria are calculated separately. When calculating the fitness of the load controller particles, only the load control loop simulation is substituted, using J... L Evaluation; Pressure controller particles use only J P The evaluation process continues in this manner; finally, each sub-controller independently determines convergence without interfering with each other.

[0126] Example 10:

[0127] This embodiment is further designed based on embodiment nine, as follows: Figure 4 As shown, the parameter scheduling strategy for the Gaussian regression process model in this example is as follows:

[0128] Gaussian Regression Process (GPR) is a nonparametric probabilistic regression model based on Gaussian processes. Its core principle is to model the relationship between input and output through the distribution of a function. The goal of GPR is to use known training data (X). n y n ), n samples, X n Let y be an n×d characteristic matrix. n Given an n×1 output vector, the output y* is predicted to correspond to the new input X*. That is, the input training data reflects the key parameter load setpoints of the system under different operating conditions and the optimal parameters of the MADRC controller under each operating condition. The output is the MADRC controller parameters predicted based on the new load setpoints.

[0129] The performance of the Gaussian Process Rejection Regulator (GPR) is almost entirely determined by its kernel function. The mean function is typically simplified, and both together define the shape of the Gaussian process. This invention selects the quadratic exponential kernel function and predicts the parameters to be optimized in a model-integrated active disturbance rejection controller using the following method:

[0130] Calculate the covariance between sample loads under any two typical operating conditions, for sample load Nesp i With sample loading Nesp j The covariance between them is: In the formula, l represents the signal variance; l represents the length scale. The standard deviation of noise; Let i be the Kronecker function, i, j = 1, 2, ..., N, where i = j and j is 1, otherwise 0.

[0131] Calculate load setpoint With sample loading Covariance: ;

[0132] Based on the calculated covariance matrix (its elements are) ), covariance vector (its elements are) To predict the current load The following parameters of the integrated active disturbance rejection controller need to be optimized:

[0133] ;

[0134] In the formula, For covariance The covariance matrix is ​​composed of elements. covariance The covariance vector matrix composed of elements; , , , and These are the model gain estimates. Bandwidth of the extended state observer Expected tracking factor Model 0 parameters and the first model parameters The predicted value; These are the model gain estimates. Bandwidth of the extended state observer Expected tracking factor Model 0 parameters and the first model parameters The column vector formed by these.

[0135] For the coordinated control system model, six operating conditions are selected as sample points. At each sample point, the approximate transfer function is obtained by identifying the model. Through optimization, the optimal MADRC parameters for the six typical operating conditions are obtained. Then, the controller parameter dataset is used as sample data for the GPR model. Finally, the corresponding MADRC controller parameters are scheduled and predicted using the load setpoint.

[0136] Specifically, such as Figure 4 As shown, the dataset first includes several sample parameters (load setpoint Nesp, controller parameters). (where i=1,2,3 represent different MADRCs) are input into GPR for training, and then the MADRCi controller parameters predicted based on the new load setpoint are output. The data is then sent to the corresponding controllers. MADRC1 controls the load; the deviation between the desired load setpoint Nesp* and the actual load Ne is input to MADRC1, and the output is the control quantity ut. MADRC2 controls the main steam pressure; the desired pressure setpoint pst... sp The deviation from the actual pressure pst is input to MADRC2, and the output is the control quantity uB. MADRC3 controls the specific enthalpy of the separator, with the desired specific enthalpy setpoint hm. sp The deviation from the actual enthalpy hm is input into MADRC3, which outputs the control quantity dfw. The control quantities (ut, uB, dfw) output by the last three MADRCs are input into the coordinated control system model. The model outputs the actual controlled quantities (Ne, pst, hm) based on these control quantities. These actual values ​​are then fed back to their respective deviation comparison links, forming a closed-loop control circuit.

[0137] Example 11:

[0138] An electronic device includes a memory and a processor. The memory stores a computer program, and the processor is used to invoke and run the computer program stored in the memory to perform the methods of any of the above embodiments.

[0139] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above embodiments.

[0140] Application Examples:

[0141] This example applies the flexible gain scheduling control method for deep peak-shaving units of the present invention to the coordination system of a simulation platform for a 1000MW ultra-supercritical thermal power unit, thereby verifying the control of the method of the present invention. The specific process is as follows:

[0142] 1) An open-loop step response test was conducted based on the simulation platform of the 1000MW ultra-supercritical thermal power unit to establish a complete transfer function model of the controlled object of the coordinated system under different operating conditions. The controlled object of the coordinated system can be described as a three-input, three-output system. The input control variables are coal feed rate uB, feedwater flow rate dfw, and turbine valve opening ut; the output control variables are the actual generated power Ne, main steam pressure pst, and steam-water separator enthalpy hm. To ensure the variable load rate, the turbine valve opening ut should be paired with the actual generated power Ne. From the perspective of the unit's dynamic characteristics, selecting fuel feed rate uB and main steam pressure pst as a pair of input and output variables is beneficial for maintaining the stability of the main steam pressure. Finally, feedwater flow rate dfw and steam-water separator enthalpy hm are selected as a pair of input and output variables.

[0143] First, an open-loop step response test was conducted to stabilize the unit at the target operating point. The unit's operating mode was switched to basic mode, the controller was disconnected, and manual operation was initiated. Then, a positive step disturbance of a certain magnitude was applied to the corresponding control quantity. After the object stabilized, a reverse step disturbance of the same magnitude was applied, and all relevant data were recorded. Before model identification, the collected raw data needed to be preprocessed: ① Outlier removal: The 3σ criterion (σ is the standard deviation of the data) was used to remove outliers caused by simulation fluctuations; ② Data smoothing: The 5-point moving average method was used to eliminate high-frequency noise. Finally, a swarm intelligence optimization algorithm was used to identify the complete transfer function model of the controlled object of the coordinated system. The transfer function of each of the nine channels "uB→pst", "uB→Ne", "uB→hm", "dfw→hm", "dfw→pst", "dfw→Ne", "ut→Ne", "ut→pst", and "ut→hm" was identified. The identification model structure was selected as a second-order system.

[0144] The identified model was stimulated with the same step input signal as in the experiment, and its response curve was compared with the corresponding actual output data. Figure 5 As shown in Table 1, the transfer functions at different operating points are obtained.

[0145] Table 1

[0146]

[0147] 2) Design a model-integrated active disturbance rejection controller based on the second-order transfer function model of the controlled object obtained in step 1):

[0148]

[0149] Where K and T represent the model gain and inertial time constant, respectively. The dynamic equations of the controlled object can be described as follows:

[0150]

[0151] In the formula: and For the 0th and 1st model parameters, which respectively possess transfer function model information, further, , ; The estimated value of the model gain b is an adjustable key parameter; This is the sum of unmodeled dynamics and disturbances. ; and These represent the uncertain parts of the two model parameters, respectively. This represents the uncertain part of the model gain; This is an external disturbance.

[0152] This can be further written as the following extended state equation with model information:

[0153]

[0154] in, .

[0155] The corresponding extended state observer can be designed as follows:

[0156]

[0157] in, Indicates the observer gain; z i For state variable x i The estimated state. At this point, the characteristic equation of the MESO can be expressed as:

[0158]

[0159] To achieve ideal estimation performance, the observer poles can be placed at the same location. ( (where is the observer bandwidth), from which the design rules for the observer gain can be derived as follows:

[0160]

[0161] When the state estimation is effective, the final control action can be derived based on the linear state feedback law as follows:

[0162]

[0163] Where, k p k d These are the proportional gain and the derivative gain, respectively.

[0164] 3) Based on the closed-loop system circuit and combined with the particle swarm optimization algorithm, the controller parameters in step 2) are tuned and optimized:

[0165] First, the MESO state-space expression is transformed into transfer function form using Laplace transform, resulting in the expressions for Z1 and Z2 as follows:

[0166]

[0167] Combining the above estimated states We can obtain the equivalent transfer function between U0(s) and Y(s), that is...

[0168]

[0169] according to We can obtain the equivalent transfer function between R(s) and Y(s), that is...

[0170]

[0171] To properly configure the proportional and integral actions, for a delay-free object, the closed-loop dynamic tracking transfer function of a linear second-order MADRC can typically be specified as follows:

[0172]

[0173] In the formula, The desired tracking factor.

[0174] By comparing the two formulas above, we can obtain the following relationship.

[0175]

[0176] Finally, the feedback gain can be obtained.

[0177]

[0178] Therefore, by using the expected tracking factor r c By adjusting the feedback gain, the closed-loop dynamic tracking transfer function of MADRC can be shaped into the desired transfer function.

[0179] Then, based on the identified transfer functions as shown in Table 1, a control loop is constructed. However, since the coordinated system is a three-input, three-output coupled system, the coupling between each loop is considered as part of the total disturbance, and then estimated and compensated for. In the process of designing a separate MESO for each loop for observation, the off-diagonal transfer functions in the system transfer function matrix are ignored, and only the diagonal transfer functions are considered for MADRC controller design. Each MADRC has five parameters to be optimized, including... Each sub-controller's objective function contains only its own output's "overshoot" and "set-up time," and assigns weights based on the process priority of that output.

[0180] The objective function of the MADRC controller for controlling the load: Objective function of the MADRC controller for controlling main steam pressure: The objective function of the MADRC controller for controlling the specific enthalpy of the separator steam is: .

[0181] Since the three objective functions are independent, particle swarm optimization can be performed in parallel. First, independent particle populations are generated for each of the three MADRC sub-controllers. Then, when calculating the fitness of the load controller particles, only the load control loop simulation is substituted, using J... L Evaluation; Pressure controller particles use only JP The evaluation process continues in this manner; finally, each sub-controller independently determines convergence without interfering with each other.

[0182] Test 1: Taking the 50% operating condition as an example, the transfer function under this condition was identified using steps 1) to 3) above, and the optimal controller parameters for MADRC were obtained as shown in Table 2. Therefore, the impact of changes in each controller parameter on the control effect was first observed, as shown in Figures 6(a) to 6(e). Secondly, taking the transfer function under the current operating condition as an example, MADRC was compared with the traditional PID control strategy. The comparison results of the control effects of the two control strategies are shown in Figures 7(a) to 7(c), and their index calculations are shown in Table 3.

[0183] Table 2

[0184] Table 3

[0185] 4) Repeat steps 2) and 3) to obtain the optimal MADRC controller parameters under different operating conditions, as shown in Table 2. Use this as a dataset to train the GPR model, and then extract the load setpoint in real time to predict the scheduling MADRC parameters.

[0186] Gaussian Regression (GPR) is a nonparametric probabilistic regression model based on Gaussian processes. Its core principle is to model the relationship between input and output through the distribution of a function. The performance of GPR is almost entirely determined by its kernel function, while the mean function is typically simplified. Together, these two functions define the form of the Gaussian process. This invention chooses the squared exponential kernel function, thus the sample loading Nesp... i and Nesp j The covariance between them is:

[0187]

[0188] In the formula: —Signal variance; l —Length scale parameter; —Noise standard deviation; —Kronecker function; i, j=1,2,..,N, is 1 when i=j, otherwise it is 0.

[0189] Load setpoint With sample loading The covariance is:

[0190]

[0191] Based on the calculated covariance matrix (its elements are) ), covariance vector (its elements are) The current load setpoint is predicted using the sample data in Table 2 and the data in Table 2. MADRC parameters below and .

[0192] ;

[0193] In the formula, For the sample data in Table 2 The column vector formed by these.

[0194] 5) The flexible gain scheduling control system designed above was used in a simulation platform for a 1000MW ultra-supercritical unit.

[0195] Test 2: The active disturbance rejection controller (ADRC) designed according to steps 1) to 4) was deployed to a simulation platform of a 1000MW ultra-supercritical unit for testing. The control strategies participating in the comparative simulation were: GPR-ADRC, ADRC and PID.

[0196] First, the load change rate was set to 1.5% / min. Then, the unit load command was varied between 320MW and 950MW. After each load command change, the system was required to operate stably for a period of time before issuing a new command, as shown in Figures 8(a) to 8(c). To better quantitatively compare the control performance of different control strategies, the integral performance indicators of each control strategy were recorded, as shown in Table 4.

[0197] Table 4

[0198] PI 1669.8 4379.8 ADRC 794.5 3301.5 GPR-ADRC 402.0 2734.9

[0199] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A flexible gain dispatch control method for a deep peak-shaving unit, wherein the deep peak-shaving unit is a generator unit that adjusts its output within a wide load range and meets the deep peak-shaving requirements of the power grid through a coordination system, characterized in that... The method includes: Open-loop step response tests were conducted under multiple typical operating conditions to identify the transfer function model of the coordinated system under each typical operating condition; each typical operating condition has a corresponding load setpoint. Based on the transfer function models of various typical operating conditions, a model-integrated active disturbance rejection controller is designed for each typical operating condition. The model-integrated active disturbance rejection controller estimates and compensates for the total disturbance in real time through an extended state observer with transfer function model information. Based on the transfer function model, a closed-loop control loop is constructed, and the particle swarm optimization algorithm is used to tune the parameters of the model integrated active disturbance rejection controller for each typical working condition, so as to obtain the optimal controller parameters corresponding to each typical working condition. The load setpoint corresponding to each typical working condition is used as input, and the optimal controller parameters corresponding to each typical working condition are used as output to train a Gaussian regression process model to construct the mapping relationship between load commands and controller parameters, so as to realize the real-time scheduling of model-integrated active disturbance rejection controller parameters under changing working conditions. The optimal controller parameters for the target operating condition are output from the trained Gaussian regression process model, and flexible gain scheduling control of the deep peak-shaving unit is performed.

2. The flexible gain scheduling control method for deep peak-shaving units according to claim 1, characterized in that, Before conducting open-loop step response tests under multiple typical operating conditions, the process also includes preprocessing the collected raw data. The preprocessing includes removing outliers using the 3σ criterion and smoothing the data using the moving average method.

3. The flexible gain scheduling control method for deep peak-shaving units according to claim 1, characterized in that, The specific method of the open-loop step response test includes: stabilizing the deep peak-shaving unit under typical operating conditions, switching to basic operating mode, disconnecting the controller and engaging manual mode; applying a positive step disturbance to the input control quantity, and after stabilization, applying a reverse step disturbance of the same amplitude; recording the input control quantity and output controlled quantity throughout the test process. The input control quantity includes coal feed rate uB, feedwater flow rate dfw, and turbine valve opening ut, while the output controlled quantity includes the unit's actual generated power Ne, main steam pressure pst, and steam-water separator specific enthalpy hm.

4. The flexible gain scheduling control method for deep peak-shaving units according to claim 3, characterized in that, The coordination system is a three-input, three-output system. The input-output pairing relationship is as follows: turbine valve opening ut - actual unit power Ne, coal feed rate uB - main steam pressure pst, feedwater flow rate dfw - steam specific enthalpy hm of steam-water separator.

5. The flexible gain scheduling control method for deep peak-shaving units according to claim 4, characterized in that, The transfer function model of the coordination system under various typical operating conditions is identified using a swarm intelligence optimization algorithm. The transfer function of the nine channels formed by the three-input and three-output pairing of the coordination system is identified separately, and the structure of the identified transfer function model is uniformly selected as a second-order system.

6. The flexible gain scheduling control method for deep peak-shaving units according to claim 1, characterized in that, The expression for the extended state observer with transfer function model information in the model-integrated active disturbance rejection controller is: ； in, ; In the formula, , and These are the first, second, and third gains of the extended state observer, respectively. To expand the bandwidth of the state observer; , and These are the observed values ​​of the output variable y, the first derivative of the output variable y, and the total disturbance f, respectively. , and They are respectively , and The first derivative; The actual output of the controlled object; For the input of the controlled object; and These are the 0th and 1st model parameters, respectively, which contain information about the transfer function model. This is the estimated value of the model gain.

7. The flexible gain scheduling control method for deep peak-shaving units according to claim 6, characterized in that, The zeroth model parameter in the expression of the extended state observer with transfer function model information and the first model parameters The dynamic equations for the transfer function model of the controlled object are obtained by establishing the transfer function model of the controlled object. The dynamic equation is established as follows: In the formula, The steady-state gain of the controlled object; For inertial time; For Lagrangian operators; and The actual output of the controlled object The first and second derivatives; This is the sum of unmodeled dynamics and disturbances.

8. The flexible gain scheduling control method for deep peak-shaving units according to claim 7, characterized in that, The specific method for tuning the parameters of the model-integrated active disturbance rejection controller includes: The state-space expression of the extended state observer, which contains transfer function model information, is transformed into a transfer function using Laplace transform. The equivalent transfer functions of the virtual control quantity U0(s) and the output Y(s) are then derived. ; Therefore, the expected closed-loop dynamic tracking transfer function is: In the formula, r c The expected tracking factor; The equivalent closed-loop dynamic tracking transfer function is ; The feedback gain is solved by comparing equations, where the proportional gain is... Differential gain .

9. The flexible gain scheduling control method for deep peak-shaving units according to claim 8, characterized in that, The specific methods of the particle swarm optimization algorithm include: Treating the loop coupling of the coordination system as the total disturbance and neglecting the off-diagonal elements of the transfer function matrix, model-integrated active disturbance rejection controllers (ADCs) are designed for the actual unit power output Ne, main steam pressure pst, and steam-water separator enthalpy hm loops. Each ADC contains five parameters to be optimized, namely the model gain estimate. Bandwidth of the extended state observer Expected tracking factor Model 0 parameters and the first model parameters ; Design independent objective functions, including power control loops. Main steam pressure control circuit Enthalpy control loop ;in, , and These are the overshoot amounts for load, main steam pressure, and separator specific enthalpy, respectively. , and These are the adjustment times for load, main steam pressure, and separator specific enthalpy, respectively. A parallel optimization approach is adopted to generate particle populations for the three integrated active disturbance rejection controllers, and calculate fitness and convergence judgment separately without interference between them.

10. The flexible gain scheduling control method for deep peak-shaving units according to claim 9, characterized in that, The Gaussian regression process model selects the squared exponential kernel function, and predicts the parameters to be optimized for the model-integrated active disturbance rejection controller by using the covariance matrix between sample loads under typical operating conditions and the covariance vector between the load setpoint and the sample load; wherein the covariance matrix corresponds to any sample load Nesp i With another sample load Nesp j The covariance is: In the formula, l represents the signal variance; l represents the length scale. The standard deviation of noise; For the Kronecker function; For load setting value With sample loading The covariance is ; The prediction method for predicting the parameters to be optimized in a model-integrated active disturbance rejection controller is as follows: , , , and ,in For covariance The covariance matrix is ​​composed of elements. covariance The covariance vector matrix composed of elements; , , , and These are the model gain estimates. Bandwidth of the extended state observer Expected tracking factor Model 0 parameters and the first model parameters The predicted value; These are the model gain estimates. Bandwidth of the extended state observer Expected tracking factor Model 0 parameters and the first model parameters The column vector formed by these.

Images