A Multivariable Adaptive Regulation System and Control Method for Steam Turbines under Wide Load and Variable Operating Conditions

By constructing a full-link adaptive control architecture, online model updates and constraint follow-up driven by operating conditions are realized, solving the problems of multivariable coupling mismatch and fixed model inaccuracy of steam turbines under wide load and variable operating conditions. This improves the anti-disturbance capability and response speed of the control system and extends the service life of the control valve.

CN123109796APending Publication Date: 2026-11-17安徽争宇峰节能科技有限公司
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611263046.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

The existing turbine DEH regulating system suffers from problems such as multivariable coupling mismatch, fixed model inaccuracy, inability to adapt constraints, and insufficient disturbance resistance under wide load and variable operating conditions. This leads to slow load response, increased speed fluctuations, increased main steam pressure overshoot, and frequent valve oscillations, affecting equipment safety and control quality.

Method used

A full-link adaptive control architecture is constructed by employing a working condition pattern recognition unit, a multivariable online identification unit, a dynamic decoupling model predictive control layer, a valve position nonlinear inverse compensation execution layer, and a working condition adaptive constraint management module. This architecture enables working condition-driven online model updates, constraint follow-up, and execution compensation, and solves the aforementioned problems through multivariable collaborative optimization control.

Benefits of technology

It maintains excellent control quality across the full load range, solves the problems of fixed model mismatch and constraint inability to adapt, improves disturbance rejection capability, ensures stability margin in deep peak shaving low load section and response speed in rapid load change section, and extends the service life of the control valve.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123109796A_ABST
    Figure CN123109796A_ABST
Patent Text Reader

Abstract

This invention discloses a multivariable adaptive regulation system and control method for a steam turbine operating under wide load and variable conditions, relating to the field of steam turbine control technology. The system includes: an operating condition pattern recognition unit, a multivariable online identification unit, a dynamic decoupling model predictive control layer, a valve position nonlinear inverse compensation execution layer, and an operating condition adaptive constraint management module. The method includes: real-time acquisition of operating parameters, online identification of operating condition types, online model identification and parameter set updating, construction of a dynamic predictive model, adaptive determination of constraint boundaries and weights, rolling optimization and valve command generation, and valve position nonlinear inverse transformation correction and execution. This invention achieves a complete closed loop from operating condition perception to execution feedback by constructing a five-in-one full-link adaptive architecture integrating operating condition recognition, online identification, predictive control, nonlinear compensation, and constraint management. Unlike existing technologies, it maintains excellent control quality across the entire load operating range.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of steam turbine control technology, and in particular to a multivariable adaptive adjustment system and control method for a steam turbine operating under wide load and variable conditions. Background Technology

[0002] With the continuous growth of new energy installed capacity, deep peak shaving and rapid load response of thermal power units have become key supports for the safe and stable operation of the power grid. As the core device for power and speed regulation, the digital electro-hydraulic control system (DEH) of the steam turbine directly affects the unit's load-changing capacity and operational safety. Existing DEH control systems generally adopt fixed-parameter PID control, single-loop decoupling control, or condition-segmented setpoint control architectures, with control parameters tuned offline, exhibiting good regulation quality only near rated operating conditions. When the unit enters the deep peak shaving low-load section, performs rapid load increases and decreases, switches between pure condensing and heating operation modes, or experiences significant fluctuations in main steam parameters, the system's dynamic performance deteriorates significantly, manifesting as slow load response, increased speed fluctuations, and increased main steam pressure overshoot.

[0003] There are strong coupling relationships among multiple variables such as power, speed, main steam pressure, valve opening, and heat load. Adjusting the opening of a certain valve will simultaneously affect multiple controlled variables. The traditional feedforward plus single PID structure cannot effectively suppress multivariate coupling disturbances. The controlled variables are mutually constrained, making it difficult to simultaneously meet multiple control indicators. At the same time, the variable operating condition model adopts a fixed structure and does not consider slow time-varying factors such as valve flow characteristic drift, efficiency changes caused by scaling or wear of flow passages, and time-delay nonlinearity of the thermodynamic system during long-term turbine operation. The fixed model may be applicable in the early stage of commissioning, but the model mismatch will intensify after several years of operation, and the control quality will gradually decline. When the load changes over a wide range, the coordinated action of multiple valves is prone to frequent reciprocating oscillations. Repeated valve action not only aggravates equipment fatigue wear but also causes continuous fluctuations in main steam pressure and load. The system cannot meet the unified requirements of response speed and stability margin.

[0004] The existing adaptive mechanism is too simplistic, only performing limited parameter segment switching between different load segments. It does not have a true online identification capability, cannot predictively compensate for main steam pressure disturbances, heating steam extraction disturbances, etc., and does not have the ability to automatically adjust constraint boundaries according to operating conditions.

[0005] In summary, existing technologies suffer from several shortcomings under wide-load variable operating conditions, including multivariable coupling mismatch, model inaccuracy, inability to adapt constraints, and insufficient disturbance rejection capabilities. There is an urgent need for a multivariable turbine control system and method capable of covering the entire load range and possessing online identification and adaptive optimization capabilities. Therefore, this invention proposes a multivariable adaptive control system and method for steam turbines under wide-load variable operating conditions to address the problems existing in the prior art. Summary of the Invention

[0006] To address the aforementioned problems, the present invention aims to propose a multivariable adaptive regulation system and control method for steam turbines under wide load and variable operating conditions. This system establishes a full-link adaptive control architecture that integrates operating condition-driven operation, online model updates, constraint follow-up, and execution compensation. It achieves multivariable collaborative optimization control across the entire load operating range, solving the problems of multivariable coupling mismatch, poor adaptability of fixed models, inability to adjust constraint boundaries according to operating conditions, and insufficient anti-disturbance capability in existing steam turbine DEH regulation systems under wide load and variable operating conditions.

[0007] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a multivariable adaptive control system for a steam turbine under wide load and variable operating conditions, comprising an operating condition mode recognition unit, a multivariable online identification unit, a dynamic decoupling model predictive control layer, a valve position nonlinear inverse compensation execution layer, and an operating condition adaptive constraint management module, wherein:

[0008] The operating condition mode recognition unit is used to acquire real-time operating data of the steam turbine unit, identify the operating condition type of the current operating condition based on the real-time operating data, and generate an operating condition type label. As the "perception layer" of the system, this unit continuously collects steam turbine unit operating data and determines the current operating condition type, providing an operating condition label basis for all subsequent adaptive actions.

[0009] The multivariable online identification unit is used to identify the multivariable coupling coefficient matrix of the steam turbine unit under the current operating conditions based on real-time operating data and operating condition type labels, and generate a dynamically updated identification parameter set. As the "model update layer" of the system, this unit continuously identifies the coupling relationship between variables during the normal operation of the unit and generates a dynamic parameter set that changes with the operating conditions.

[0010] The dynamic decoupling model predictive control layer is used to retrieve the corresponding target identification parameters from the dynamically updated identification parameter set according to the working condition type label, construct a multivariate dynamic prediction model based on the target identification parameters, and perform multi-objective rolling optimization solution on the multivariate dynamic prediction model in the rolling time domain, outputting multiple control commands. This layer serves as the "decision core layer" of the system and performs rolling time domain optimization based on the real-time model provided by the identification unit.

[0011] The valve position nonlinear inverse compensation execution layer is used to receive control commands from multiple control valves, perform nonlinear inverse transformation correction on the control commands from multiple control valves according to the inverse mapping function of the flow characteristics of each control valve, generate valve position setpoints for each control valve, and output the valve position setpoints for each control valve to the servo actuator of the corresponding control valve. This layer, as the "execution compensation layer" of the system, is responsible for converting the ideal flow command output by the control layer into the actual valve position opening command.

[0012] The working condition adaptive constraint management module is used to dynamically adjust the dynamic constraint boundary of the multivariate dynamic prediction model in the rolling optimization solution process according to the working condition type label. This module serves as the system's "safety boundary layer" and dynamically adjusts various hard and soft constraints in the optimization solution process according to the current working condition.

[0013] A further improvement is that the working condition mode recognition unit includes:

[0014] The data caching submodule is used to cache a preset number of consecutive control cycle sampling data in real-time running data;

[0015] The feature extraction submodule is used to extract the load change rate, the absolute value of the main steam pressure deviation, the rate of change of the extraction steam flow, and the absolute value of the speed deviation from the sampling data of a continuous preset number of control cycles, and form a multi-dimensional operating condition feature vector.

[0016] The working condition determination submodule is used to input a multi-dimensional working condition feature vector into a preset working condition classifier, and the preset working condition classifier outputs a working condition type label.

[0017] The operating condition type label is selected from one of the following operating condition types: rated steady-state operating condition, deep peak shaving low load steady-state operating condition, rapid load increase dynamic operating condition, rapid load decrease dynamic operating condition, pure condensing operating condition, heating extraction steam operating condition, and main steam pressure large fluctuation disturbance operating condition.

[0018] A further improvement is that the multivariate online identification unit includes:

[0019] The data synchronization submodule is used to perform timestamp alignment on real-time running data and generate a multivariable time-series synchronization data sequence.

[0020] The model structure selection submodule is used to select the model order and time delay parameters for the current identification cycle based on the working condition type label;

[0021] The recursive parameter estimation submodule is used to identify the multivariate coupling coefficient matrix under the current operating conditions online based on the multivariate time-series synchronous data sequence, model order, and time delay parameters using a recursive regression method with a forgetting factor.

[0022] The parameter set storage submodule is used to classify and store the multivariable coupling coefficient matrix according to the operating condition type label, and generate a dynamically updated identification parameter set that includes steady-state operating condition parameter set, dynamic operating condition parameter set for load increase and decrease, and heating switching transition operating condition parameter set.

[0023] The row variables of the multivariate coupling coefficient matrix are the valve opening, extraction butterfly valve opening, main steam temperature and pressure parameters, and reheat parameters, while the column variables are the unit's actual power output, speed, pressure values ​​at each stage, and extraction steam flow rate for heating.

[0024] A further improvement is that the dynamic decoupling model prediction control layer includes:

[0025] The model parameter loading submodule is used to read target identification parameters from the dynamically updated identification parameter set according to the working condition type label, and assign the target identification parameters to the coefficient matrices of each state equation of the multivariate dynamic prediction model.

[0026] The prediction time domain rolling submodule is used to recursively calculate the predicted output sequence of the controlled variable in future control cycles based on a multivariate dynamic prediction model within a preset prediction time domain, starting from the current control cycle.

[0027] The cost function construction submodule is used to determine the target weight vector based on the working condition type label, and to construct the cost function for the current rolling optimization cycle based on the target weight vector and the dynamic constraint boundary.

[0028] The optimization solution submodule is used to perform quadratic programming to solve the cost function within the feasible region defined by the dynamic constraint boundary, and generate the optimal control increment sequence that minimizes the cost function in the control time domain.

[0029] The instruction output submodule is used to extract the control increment of the current control cycle from the optimal control increment sequence, and generate a multi-channel control instruction by superimposing the output instruction of the previous control cycle.

[0030] A further improvement is that the adaptive constraint management module for operating conditions includes:

[0031] The constraint boundary baseline library submodule is used to store the constraint boundary baseline values ​​corresponding to each operating condition type label. The constraint boundary baseline values ​​include the upper limit value of the valve position change rate, the upper limit value of the main steam pressure deviation, the upper limit value of the speed deviation, and the valve position command oscillation suppression constraint value.

[0032] The real-time boundary generation submodule is used to read the target constraint boundary baseline value corresponding to the current working condition type label, correct the target constraint boundary baseline value online according to the deviation direction between the measured value and the set value of the current controlled variable, and generate the dynamic constraint boundary of the current control cycle.

[0033] The boundary gradual transition submodule is used to gradually transition the constraint boundary of the working condition before the switch to the constraint boundary of the working condition after the switch within a preset gradual transition time after the working condition type label is switched, with a preset interpolation time constant, and generate dynamic constraint boundary during the transition period.

[0034] The boundary output submodule is used to output the dynamic constraint boundary or transition period dynamic constraint boundary to the dynamic decoupling model prediction control layer.

[0035] A further improvement is that the valve position nonlinear inverse compensation execution layer includes:

[0036] The flow characteristic storage submodule is used to store the valve position command-actual flow mapping points of each control valve at different historical sampling times, and to construct the flow characteristic data table of each control valve.

[0037] The inverse mapping interpolation submodule is used to retrieve the adjacent mapping points corresponding to the target flow from the flow characteristic data table according to the multi-channel control valve command, and perform interpolation calculations on the adjacent mapping points to generate the valve position setting values ​​of each control valve.

[0038] The online update submodule for flow characteristics is used to acquire the actual valve position feedback signal and the corresponding actual steam flow measurement value of each valve at a preset acquisition time, store the actual valve position feedback signal and the actual steam flow measurement value as new mapping points in the flow characteristic data table, and remove the earliest mapping point in the flow characteristic data table at a preset update cycle.

[0039] The inverse transformation output submodule is used to send the valve position setting values ​​of each regulating valve to the servo actuator of each regulating valve at a set period.

[0040] A further improvement is that the dynamic decoupling model predictive control layer is equipped with a disturbance feedforward compensation submodule, which is used to identify the current disturbance type based on real-time operating data, retrieve the corresponding feedforward compensation coefficient from the preset feedforward compensation coefficient table according to the disturbance type, generate the feedforward compensation amount based on the feedforward compensation coefficient, and superimpose the feedforward compensation amount onto the multi-channel regulating valve control command. The disturbance types include load command step disturbance type, main steam pressure fluctuation disturbance type, heating extraction steam flow rate change disturbance type, and grid frequency deviation disturbance type.

[0041] A control method for a multivariable adaptive regulation system of a steam turbine under wide load and variable operating conditions includes the following steps:

[0042] Step S1: Real-time acquisition of operating parameters

[0043] The operating parameters of the steam turbine unit are collected in real time through the operating condition mode recognition unit and the multivariable online identification unit. The operating parameters include the opening degree of the regulating valve, the opening degree of the extraction steam butterfly valve, the main steam temperature and pressure value, the reheat parameters, the actual power output of the unit, the speed, the pressure value of each stage, and the heat extraction steam flow rate.

[0044] Step S2: Online identification of working condition type

[0045] The operating parameters are timestamped and aligned by the operating condition mode recognition unit to generate a multivariate time-series synchronization data sequence. The operating condition type of the current control cycle is identified based on the load command change rate and the measured power change rate in the multivariate time-series synchronization data sequence, and an operating condition type label is generated.

[0046] Step S3: Online model identification and parameter set update

[0047] The multivariate online identification unit selects the model order and time delay parameters for the current identification period based on the working condition type label, and uses the recursive regression method with forgetting factor to identify the multivariate time-series synchronous data sequence online, generating a multivariate coupling coefficient matrix under the current operating condition, and storing the multivariate coupling coefficient matrix as a dynamically updated identification parameter set according to the working condition type.

[0048] Step S4: Construction of Dynamic Prediction Model

[0049] The dynamic decoupling model predictive control layer retrieves target identification parameters from the dynamically updated identification parameter set based on the working condition type label, and constructs a multivariate dynamic prediction model based on the target identification parameters;

[0050] Step S5: Adaptive Determination of Constraint Boundaries and Weights

[0051] The adaptive constraint management module determines the dynamic constraint boundary of the current control cycle based on the operating condition type label, and the dynamic decoupling model prediction control layer determines the target weight vector based on the operating condition type label.

[0052] Step S6: Scrolling optimization and gate adjustment instruction generation

[0053] The dynamic decoupling model predictive control layer takes the current control cycle as the starting time and recursively calculates the predicted output sequence of the controlled variable in the future control cycle based on the multivariate dynamic prediction model within the preset prediction time domain. It constructs a cost function based on the target weight vector and dynamic constraint boundary, solves the cost function within the feasible domain defined by the dynamic constraint boundary, and generates the optimal control increment sequence that minimizes the cost function. It extracts the control increment of the current control cycle from the optimal control increment sequence, and generates a multi-channel control command by superimposing the output command of the previous control cycle.

[0054] Step S7: Valve position nonlinear inverse transformation correction and execution

[0055] The valve position nonlinear inverse compensation execution layer obtains the flow characteristic mapping relationship of each regulating valve, performs nonlinear inverse transformation correction on the multi-channel regulating valve control command according to the flow characteristic mapping relationship, generates the valve position setting value of each regulating valve, and outputs the valve position setting value of each regulating valve to the servo actuator of each regulating valve.

[0056] The beneficial effects of this invention are as follows: By constructing a five-in-one full-link adaptive architecture integrating operating condition identification, online identification, predictive control, nonlinear compensation, and constraint management, this invention achieves a complete closed loop from operating condition perception to execution feedback. Unlike existing single-loop or simple segmented control, it maintains excellent control quality across the entire load range. Through a multi-variable online identification unit that updates the coupling coefficient matrix in real time and stores it according to operating conditions, it solves the technical problem of mismatch in fixed transfer function models over a wide load range, ensuring that the internal model always remains consistent with the actual dynamic characteristics of the unit, providing an accurate model foundation for high-quality control. The invention employs a model predictive control strategy driven by operating condition type labels, where the cost function weights and dynamic constraint boundaries are adaptively adjusted according to the operating condition. This resolves the technical contradiction that the same set of control parameters cannot simultaneously address high and low load control quality, ensuring stability margin in deep peak shaving low-load sections and response speed in rapid load changing sections. A valve position nonlinearity inverse compensation execution layer is introduced. Through online updates of the flow characteristic mapping relationship, the control command of the regulating valve is inversely transformed and corrected, eliminating the nonlinear distortion caused by flow characteristic drift after long-term valve operation. This significantly improves the linearity and stability of flow regulation in low-load areas. A disturbance feedforward compensation submodule is set up, providing differentiated compensation based on the disturbance type. For large disturbances, the response gain is automatically reduced to prevent overshoot; for small disturbances, the response speed is improved by tracking the disturbance. This effectively suppresses the impact of typical disturbances such as main steam pressure fluctuations and heating steam extraction changes on the controlled variable. Through the boundary gradual transition function of the operating condition adaptive constraint management module, a smooth transition of constraint boundaries is achieved during operating condition switching. Combined with the rolling optimization of multivariable predictive control, this fundamentally eliminates the disturbance impact and regulating valve oscillation phenomena commonly seen during segmented switching of multiple controllers, extending the service life of the regulating valve. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the framework structure of the wide-load variable operating condition steam turbine multivariable adaptive adjustment system of the present invention;

[0058] Figure 2 This is a schematic diagram of the control method of the multivariable adaptive adjustment system for a steam turbine under wide load and variable operating conditions according to the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] It should be noted that the technical means not described in detail in the following embodiments are all conventional means in the art, are not the key points of the invention, and will not be elaborated upon.

[0061] Example 1

[0062] according to Figure 1 As shown, this embodiment provides a multi-variable adaptive adjustment system for a steam turbine under wide load and variable operating conditions. It is deployed inside the controller of the steam turbine's digital electro-hydraulic control system and works in conjunction with field sensors, transmitters, and servo actuators.

[0063] The system consists of five core modules: a working condition pattern recognition unit, a multivariable online identification unit, a dynamic decoupling model predictive control layer, a valve position nonlinear inverse compensation execution layer, and a working condition adaptive constraint management module. These five modules work together to construct a complete adaptive control closed loop.

[0064] The operating condition mode recognition unit's input terminal connects to the unit's sensor and transmitter groups. The hardware includes a speed probe, power transmitter, pressure / temperature transmitters at various stages, valve position displacement sensors, and extraction steam flow meters, enabling real-time acquisition of unit operating data. The acquired data covers key parameters such as control valve opening, extraction steam butterfly valve opening, main steam temperature and pressure, reheat parameters, actual generated power, speed, pressure at various stages, and heating extraction steam flow rate. After preprocessing and feature extraction of the raw data, the unit outputs accurate operating condition type labels.

[0065] This unit comprises three sub-modules: data caching, feature extraction, and operating condition determination. The division of labor is clear, and data is processed step by step.

[0066] The data caching submodule has a built-in circular buffer that uses a first-in, first-out (FIFO) management logic to continuously cache sampled data for 200 consecutive control cycles. Each time a new cycle of data is updated, the oldest set of historical data in the buffer is automatically removed to ensure data timeliness and continuity.

[0067] The feature extraction submodule reads continuously sampled data from the buffer, calculates and extracts four core feature quantities, and combines them to form a multi-dimensional operating condition feature vector. The four feature quantities include load change rate, absolute value of main steam pressure deviation, extraction steam flow rate change rate, and absolute value of speed deviation. The load change rate is calculated by performing a first-order difference operation on the actual power generation time-series data of the unit, followed by smoothing with a 0.5Hz first-order low-pass filter, effectively eliminating high-frequency interference noise.

[0068] The operating condition determination submodule incorporates a pre-trained support vector machine multi-classifier. The training data covers the entire range of rated load from 30% to 100% of the unit, enabling it to identify operating conditions under full load. The submodule inputs multi-dimensional operating condition feature vectors into the classifier and outputs a unique operating condition type label, specifically including seven types of operating conditions: rated steady-state operating condition, deep peak shaving low-load steady-state operating condition, rapid load increase dynamic operating condition, rapid load decrease dynamic operating condition, pure condensing operation condition, heating extraction steam operation condition, and main steam pressure large fluctuation disturbance condition.

[0069] The input terminals of the multivariable online identification unit are connected to the operating condition pattern recognition unit and the field sensor group, respectively. It synchronously receives real-time operating data and operating condition tags, identifies the multivariable coupling characteristics of the unit online according to the current operating condition, dynamically updates the multivariable coupling coefficient matrix, and forms a dynamically identified parameter set that is classified and stored, providing accurate parameter support for the upper-level control model.

[0070] The unit comprises four sub-modules: data synchronization, model structure selection, recursive parameter estimation, and parameter set storage.

[0071] The data synchronization submodule performs timestamp alignment on the raw sampled data from multiple channels to unify the data time base. Alignment method: Using the start time of the control cycle as the reference point, the sampled value from each channel closest to the reference point is selected. Accurate reference time data is calculated through linear interpolation, generating a multivariate data vector for time synchronization and eliminating timing deviations in the multi-channel data.

[0072] The model structure selects sub-modules, and every 10 control cycles (single identification cycle), matches model parameters based on the current operating condition label, and configures the model order and time delay differently:

[0073] 1. Rated steady-state and deep peak-shaving low-load steady-state operating conditions, using a second-order model with a time delay of 2 control cycles;

[0074] 2. Rapid load increase / decrease dynamic operating conditions, using a third-order model with a time delay of 1 control cycle;

[0075] 3. The transition process between pure condensing and heating steam extraction modes adopts a third-order model with a time delay of 3 control cycles.

[0076] The recursive parameter estimation submodule, based on synchronous time-series data and matched model parameters, uses recursive least squares with a forgetting factor to perform online identification and solve for the multivariate coupling coefficient matrix. The formula for (recursive parameter identification with forgetting factor) is as follows:

[0077]

[0078] In the formula: for Estimates of parameters to be identified in the time-matrix model (multivariate coupling coefficient matrix and parameter vector). for Time-based historical parameter estimates for The time-mapping algorithm gain vector is used to correct parameter bias. for The unit is controlled to output measured values ​​(power, speed, extraction steam flow, etc.) at any given time. for The time-series data regression vector is composed of the unit control input time-series data. For regression vectors The transpose of the matrix, for The time-varying covariance matrix is ​​used to characterize the convergence properties of parameter identification error. for Time-based historical covariance matrix It is a forgetting factor used to weaken the weight of old historical data and improve the ability to track new data.

[0079] The row variables of the coupling coefficient matrix are the valve opening, extraction butterfly valve opening, main steam temperature and pressure, and reheat parameters; the column variables are the actual power output of the unit, speed, pressure at each stage, and heating extraction steam flow rate, accurately representing the multivariate coupling correlation characteristics.

[0080] The parameter set storage submodule stores the identified coefficient matrix according to the operating conditions, namely, steady-state operating condition parameter set, dynamic operating condition parameter set for load increase and decrease, and heating switching transition operating condition parameter set, so as to realize dynamic parameter updates and on-demand retrieval.

[0081] The dynamic decoupled model predictive control layer accesses the output data of the multivariable online identification unit, the operating condition mode recognition unit, and the operating condition adaptive constraint management module. Based on real-time operating condition parameters, a dynamic prediction model is constructed. Through rolling time-domain multi-objective optimization, high-precision multi-channel control commands are output. At the same time, disturbance feedforward compensation is superimposed to improve anti-interference capability.

[0082] This layer contains six sub-modules: model parameter loading, prediction time-domain rolling, cost function construction, optimization solution, instruction output, and perturbation feedforward compensation.

[0083] In each control cycle, the model parameter loading submodule retrieves the corresponding identification parameters based on the operating condition label and assigns them to the discrete-time state-space prediction model (multivariate state-space prediction model). The formula is as follows:

[0084]

[0085] In the formula: for The system state vector at any given time includes intermediate state variables such as unit pressure, flow rate, and valve position. for The system's predicted state vector for the next cycle at time 1; for The system control input vector includes control variables such as valve opening, extraction butterfly valve opening, main steam temperature and pressure, and reheat parameters. for The system outputs controlled vectors at all times, including controlled variables such as actual unit power generation, speed, pressure at each stage, and steam extraction flow rate for heating. This is the system state matrix, representing the temporal coupling relationships of the internal states of the unit. The control input matrix characterizes the relationship between the control input and the system state. The state-output matrix represents the mapping relationship between the system state and the controlled output. The direct transfer matrix characterizes the direct coupling effect of the control input on the controlled output. , , , All coefficient matrices are assigned values ​​based on the online identification results of the current operating conditions.

[0086] The prediction time domain rolling submodule is set to 20 control cycles in the prediction time domain and 5 control cycles in the control time domain. Starting from the current moment, it recursively calculates the future controlled variable prediction sequence based on the state space model.

[0087] The cost function construction submodule configures differentiated weight vectors according to different operating conditions, constructs a quadratic cost function, and balances the deviation of the controlled variable with the fluctuation of the control increment. (The formula for the rolling optimization cost function is as follows:)

[0088]

[0089] In the formula: The rolling optimization cost function value is used to determine the optimal control effect; a smaller value indicates better control performance. For prediction in the time domain, this system uses a fixed 20 control cycles. To control the time domain, this system is fixed at 5 control cycles. for Time prediction Predicted value of the controlled output at any given time. for The controlled output is always set to the rated value. Weighted matrix The second norm, Weighted matrix The second norm, for The input increment is constantly controlled, that is, the change in the opening of the regulating valve and butterfly valve in adjacent cycles. This is a diagonal matrix of controlled variable deviation weights, containing deviation weights for speed, pressure, power, and extraction steam flow rate, which adaptively adjusts according to operating conditions. The incremental weight diagonal matrix is ​​used to control frequent fluctuations in control actions and is adaptively adjusted according to operating conditions.

[0090] The differentiated weighting configurations for each operating condition are as follows:

[0091] 1. Deep peak shaving under low load steady state: speed deviation 0.8, main steam pressure deviation 0.6, power deviation 0.3, control increment 0.2;

[0092] 2. Rapid load increase / decrease dynamics: power deviation 0.8, speed and main steam pressure deviations 0.3 each, control increment 0.1;

[0093] 3. Rated steady state: speed deviation 0.6, main steam pressure deviation 0.5, power deviation 0.4, control increment 0.3;

[0094] 4. Heating steam extraction operation: extraction steam flow deviation 0.7, main steam pressure deviation 0.5, power deviation 0.3, control increment 0.2;

[0095] 5. Operating condition with large fluctuations in main steam pressure: main steam pressure deviation 0.9, speed and power deviations 0.2 each, control increment 0.4.

[0096] The optimization solution submodule uses the interior-point method to perform quadratic programming on the cost function, solving for the optimal control increment sequence within the dynamically constrained feasible region. The instruction output submodule extracts the optimal control increment for the current period, superimposes it with historical output instructions, and generates the basic control instructions.

[0097] The disturbance feedforward compensation submodule identifies four types of disturbances in real time (load step, main steam pressure fluctuation, extraction steam flow change, and grid frequency deviation), retrieves the matching compensation coefficients to calculate the compensation amount, and superimposes it onto the basic control commands. Specific compensation parameters are: load step disturbance coefficient 0.6, lasting 10 cycles; main steam pressure fluctuation disturbance coefficient 0.8, lasting 15 cycles; extraction steam flow disturbance coefficient 0.7, lasting 12 cycles; and grid frequency disturbance coefficient 0.9, lasting 5 cycles.

[0098] The valve position nonlinearity inverse compensation actuator receives the upper-level control valve control command, combines it with the actual flow characteristics of each control valve, completes the nonlinearity inverse transformation correction, outputs a precise valve position setpoint, drives the servo actuator to move, and solves the control deviation problem caused by the nonlinearity of the control valve flow.

[0099] This layer contains four sub-modules: traffic characteristic storage, inverse mapping interpolation, online update of traffic characteristics, and inverse transformation output.

[0100] The flow characteristic storage submodule establishes an independent data table for each control valve, stores multiple sets of two-dimensional mapping coordinates of "valve position command - actual flow", and records the historical operating characteristic data of the unit.

[0101] The inverse mapping interpolation submodule completes nonlinear inverse correction through linear interpolation, accurately matching the valve position corresponding to the target flow rate. The formula for (linear interpolation inverse mapping of control valve position) is as follows:

[0102]

[0103] In the formula: The final target valve position setting value after calibration. To control the target desired steam flow rate corresponding to the control command, , The flow characteristic data table contains two adjacent sets of measured flow rates that enclose the target flow rate. For low-side flow, For high-side flow, , For the corresponding , Historical valve position command values ​​for measured flow. When the target expected flow exceeds the data's mapping range, the valve position is calculated using proportional extrapolation, while simultaneously forcibly limiting the valve position value to the upper and lower limits of the valve's physical opening to prevent over-range operation.

[0104] The online flow characteristic update submodule collects valve position feedback and corresponding steam flow in real time under steady-state conditions and a load change rate of <1% of rated load / minute, and updates the data table with new mapping points. The oldest data is removed every 1440 collection cycles, maintaining the number of mapping points in a single table to no more than 200, ensuring that the characteristic data is valid in real time.

[0105] The inverse converter output submodule sends the corrected valve position setpoint to the servo actuator at a 20ms cycle to form a closed-loop control.

[0106] The adaptive constraint management module dynamically and adaptively adjusts the constraint boundaries of the rolling optimization based on real-time operating condition labels, avoiding constraint jumps when operating conditions change, ensuring a smooth and shock-free control process, and sending constraint parameters to the optimization solution submodule in real time.

[0107] The module consists of four sub-modules: constraint boundary baseline library, real-time boundary generation, boundary gradual transition, and boundary output.

[0108] The constraint boundary baseline library stores the baseline parameters corresponding to each operating condition, covering the upper limit of the valve position change rate, the upper limit of the main steam pressure deviation, the upper limit of the speed deviation, and the valve position oscillation suppression constraint value.

[0109] The real-time boundary generation submodule combines the actual deviation of the controlled variable to correct the baseline: if the measured value is higher than the set value, a negative correction is made; if it is lower than the set value, a positive correction is made. The correction magnitude is proportional to the absolute value of the deviation, and the maximum correction amount does not exceed 20% of the baseline value, generating the dynamic constraint boundary for the current cycle.

[0110] The boundary gradual transition submodule monitors the working condition switching action. After the working condition switch, it starts a gradual transition of 60 control cycles. With an interpolation time constant of 0.05 and a first-order exponential smoothing method, it realizes the continuous gradual change of the constraint boundary without instantaneous jump. After the transition period, it fully matches the new working condition baseline constraint.

[0111] The five units operate on a unified control cycle of 20ms and exchange data in real time through the controller's internal data bus, forming a complete adaptive closed-loop control logic of "condition perception → model update → optimization decision → nonlinear compensation → constraint management".

[0112] Example 2

[0113] according to Figure 2 As shown, this embodiment provides a control method for a multivariable adaptive regulation system of a steam turbine under wide load and variable operating conditions. It employs the system provided in Embodiment 1 and the overall process consists of seven steps, achieving full-process adaptive operation, online model updating, dynamic constraint adjustment, and precise nonlinear compensation. The steps include:

[0114] Step S1: Real-time acquisition of operating parameters

[0115] The system collects comprehensive operating parameters of the unit through various sensors and transmitters at differentiated sampling frequencies. These parameters include valve opening, extraction butterfly valve opening, main steam temperature and pressure, reheat parameters, actual power generation, speed, pressure at each stage, and extraction steam flow rate for heating. The sampling frequencies for each device are: speed probe 1000Hz; power, pressure, and valve position sensors 100Hz; temperature and extraction steam flow sensors 50Hz. All signals are converted from analog to digital and then synchronously transmitted to the operating condition identification unit and data synchronization submodule via the controller data bus.

[0116] Step S2: Online identification of working condition type

[0117] First, the acquired multi-channel data is timestamped and linearly interpolated to generate a time-series synchronization data vector. Based on the synchronization data, the load command change rate and the measured power change rate are calculated, and then smoothed by a 0.5Hz first-order low-pass filter to eliminate interference.

[0118] The operating condition type is determined based on preset classification logic. Specific determination rules are as follows:

[0119] 1. Steady-state operating conditions: The absolute value of the load command change rate is <1% of the rated load / minute, and the absolute value of the measured power change rate is <0.5% of the rated load / minute; the load is ≥90% of the rated load as the rated steady-state operating condition, and 30% to 50% of the rated load as the deep peak-shaving low-load steady-state operating condition.

[0120] 2. Dynamic operating conditions: A positive load command change rate with an absolute value > 2% of rated load / minute indicates a rapid load increase condition; a negative load command change rate with an absolute value > 2% of rated load / minute indicates a rapid load decrease condition.

[0121] 3. Operating mode and conditions: When the extraction steam flow rate is less than 5% of the rated extraction steam flow rate, it is a pure condensing operating condition; when it is greater than 10% of the rated extraction steam flow rate, it is a heating extraction steam operating condition.

[0122] 4. Disturbance Condition: If the absolute value of the main steam pressure deviation is greater than 3% of the rated pressure and continues for more than 5 control cycles, it is judged as a disturbance condition with large pressure fluctuation.

[0123] Each control cycle outputs a working condition type label in real time and sends it to each core unit synchronously.

[0124] Step S3: Online model identification and parameter set update

[0125] The matching rules are consistent with those in Example 1, based on the order and time delay parameters of the matching model according to the real-time operating condition labels. A recursive least squares method with a variable forgetting factor is used to identify synchronous time-series data online, establish a multivariate autoregressive exogenous model, and solve for the multivariate coupling coefficient matrix.

[0126] Forgetting Factor Adaptive Configuration: Steady-State Conditions =0.995, ensuring identification stability; dynamic operating conditions =0.97, improving parameter tracking speed; Disturbance conditions =0.98, balancing stability and speed.

[0127] The parameter set is updated using a sliding window weighted average, with the window covering 10 identification periods. The weight of the most recent 4 periods is 0.8, and the weight of the oldest 6 periods is 0.2. After the update, the parameters are stored according to the working conditions to form a dynamic parameter set.

[0128] Step S4: Construction of Dynamic Prediction Model

[0129] Each control cycle retrieves the identification parameters corresponding to the current operating condition, assigns them to the discrete state space prediction model, updates the model coefficient matrix in real time, and constructs a multivariate dynamic prediction model that adapts to the current unit operating state, accurately characterizing the multivariate coupled dynamic characteristics of the unit under varying operating conditions.

[0130] Step S5: Adaptive Determination of Constraint Boundaries and Weights

[0131] Based on real-time operating condition labels, dual adaptive adjustment is implemented: First, constraint boundary adaptation reads the baseline constraint values ​​of the operating conditions and dynamically corrects them by combining the sign and magnitude of the controlled variable deviation. During operating condition switching, a 60-cycle exponential smooth transition is executed to eliminate constraint jumps. Second, weight adaptation configures the cost function according to preset operating condition matching rules. , The weight matrix allows for differentiated control priorities for various scenarios, including steady-state, dynamic, heating, and disturbance scenarios.

[0132] Step S6: Scrolling optimization and gate adjustment instruction generation

[0133] Based on a dynamic prediction model, the predicted sequence of the controlled variable is recursively calculated within a 20-cycle prediction time domain. A quadratic cost function is constructed by combining adaptive weights and dynamic constraint boundaries. The optimal control increment sequence within a 5-cycle control time domain is solved using interior-point quadratic programming.

[0134] Extract the optimal control increment for the current cycle and overlay it with historical output commands to generate basic control commands. After identifying the type of disturbance to the unit, match the corresponding feedforward compensation coefficient and compensation duration, calculate the compensation amount, and overlay it onto the control commands to improve the system's disturbance rejection capability.

[0135] Step S7: Valve position nonlinear inverse transformation correction and execution

[0136] The flow characteristic mapping data of each regulating valve is retrieved, and the nonlinear correction of the regulating valve control command is performed through the linear interpolation inverse mapping formula to eliminate the control error caused by the nonlinear characteristics of the regulating valve flow.

[0137] The calibrated valve position setpoint is sent to the servo actuator every 20ms to drive the control valve. During the unit's steady-state low-fluctuation period, the control valve flow characteristic data table is automatically updated to discard outdated data and ensure mapping accuracy. At the same time, the actual opening degree is fed back in real time through the valve position displacement sensor, forming a high-precision closed-loop control.

[0138] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multivariable adaptive control system for a steam turbine operating under wide load and variable conditions, characterized in that, include: The operating condition mode recognition unit is used to acquire real-time operating data of the steam turbine unit, identify the operating condition type of the current operating condition, and generate an operating condition type label. The multivariable online identification unit is used to identify the multivariable coupling coefficient matrix of the steam turbine unit under the current operating conditions based on real-time operating data and operating condition type labels, and generate a dynamically updated identification parameter set. The dynamic decoupling model predictive control layer is used to retrieve the corresponding target identification parameters from the dynamically updated identification parameter set according to the working condition type label, construct a multivariate dynamic prediction model based on the target parameters, and perform multi-objective rolling optimization solution on the multivariate dynamic prediction model in the rolling time domain, and output multi-channel valve control commands. The valve position nonlinear inverse compensation execution layer is used to receive multi-channel control commands, perform nonlinear inverse transformation correction on the multi-channel control commands according to the inverse mapping function of the flow characteristics of each control valve, generate valve position setpoints for each control valve, and output the valve position setpoints for each control valve to the servo actuator of the corresponding control valve. The working condition adaptive constraint management module is used to dynamically adjust the dynamic constraint boundaries of the multivariate dynamic prediction model in the rolling optimization solution process based on the working condition type label.

2. The wide-load variable operating condition steam turbine multivariable adaptive control system according to claim 1, characterized in that: The operating condition mode recognition unit includes: The data caching submodule is used to cache a preset number of consecutive control cycle sampling data in real-time running data; The feature extraction submodule is used to extract the load change rate, the absolute value of the main steam pressure deviation, the rate of change of the extraction steam flow, and the absolute value of the speed deviation from the sampling data of a continuous preset number of control cycles, and form a multi-dimensional operating condition feature vector. The operating condition determination submodule is used to input a multi-dimensional operating condition feature vector into a preset operating condition classifier. The preset operating condition classifier outputs an operating condition type label, which is selected from one of the following operating condition types: rated steady-state operating condition, deep peak shaving low load steady-state operating condition, rapid load increase dynamic operating condition, rapid load decrease dynamic operating condition, pure condensing operation condition, heating extraction steam operation condition, and main steam pressure large fluctuation disturbance operating condition.

3. The wide-load variable operating condition steam turbine multivariable adaptive control system according to claim 1, characterized in that: The multivariate online identification unit includes: The data synchronization submodule is used to perform timestamp alignment on real-time running data and generate a multivariable time-series synchronization data sequence. The model structure selection submodule is used to select the model order and time delay parameters for the current identification cycle based on the working condition type label; The recursive parameter estimation submodule is used to identify the multivariate coupling coefficient matrix under the current operating conditions online based on the multivariate time-series synchronous data sequence, model order, and time delay parameters using a recursive regression method with a forgetting factor. The parameter set storage submodule is used to classify and store the multivariable coupling coefficient matrix according to the operating condition type label, and generate a dynamically updated identification parameter set that includes steady-state operating condition parameter set, dynamic operating condition parameter set for load increase and decrease, and heating switching transition operating condition parameter set. The row variables of the multivariate coupling coefficient matrix are the valve opening, extraction butterfly valve opening, main steam temperature and pressure parameters, and reheat parameters, while the column variables are the unit's actual power output, speed, pressure values ​​at each stage, and extraction steam flow rate for heating.

4. The multivariable adaptive control system for a steam turbine under wide load and variable operating conditions according to claim 1, characterized in that: The dynamic decoupling model prediction control layer includes: The model parameter loading submodule is used to read target identification parameters from the dynamically updated identification parameter set according to the working condition type label, and assign the target identification parameters to the coefficient matrices of each state equation of the multivariate dynamic prediction model. The prediction time domain rolling submodule is used to recursively calculate the predicted output sequence of the controlled variable in future control cycles based on a multivariate dynamic prediction model within a preset prediction time domain, starting from the current control cycle. The cost function construction submodule is used to determine the target weight vector based on the working condition type label, and to construct the cost function for the current rolling optimization cycle based on the target weight vector and the dynamic constraint boundary. The optimization solution submodule is used to perform quadratic programming to solve the cost function within the feasible region defined by the dynamic constraint boundary, and generate the optimal control increment sequence that minimizes the cost function in the control time domain. The instruction output submodule is used to extract the control increment of the current control cycle from the optimal control increment sequence, and generate a multi-channel control instruction by superimposing the output instruction of the previous control cycle.

5. The wide-load variable operating condition steam turbine multivariable adaptive control system according to claim 1, characterized in that: The adaptive constraint management module for operating conditions includes: The constraint boundary baseline library submodule is used to store the constraint boundary baseline values ​​corresponding to each operating condition type label, including the upper limit value of the valve position change rate, the upper limit value of the main steam pressure deviation, the upper limit value of the speed deviation, and the constraint value of the valve position command oscillation suppression. The real-time boundary generation submodule is used to read the target constraint boundary baseline value corresponding to the current working condition type label, correct the target constraint boundary baseline value online according to the deviation direction between the measured value and the set value of the current controlled variable, and generate the dynamic constraint boundary of the current control cycle. The boundary gradual transition submodule is used to gradually transition the constraint boundary of the working condition before the switch to the constraint boundary of the working condition after the switch within a preset gradual transition time after the working condition type label is switched, with a preset interpolation time constant, and generate dynamic constraint boundary during the transition period. The boundary output submodule is used to output the dynamic constraint boundary or transition period dynamic constraint boundary to the dynamic decoupling model prediction control layer.

6. The wide-load variable operating condition steam turbine multivariable adaptive control system according to claim 1, characterized in that: The valve position nonlinear inverse compensation execution layer includes: The flow characteristic storage submodule is used to store the valve position command-actual flow mapping points of each control valve at different historical sampling times, and to construct the flow characteristic data table of each control valve. The inverse mapping interpolation submodule is used to retrieve the adjacent mapping points corresponding to the target flow from the flow characteristic data table according to the multi-channel control valve command, and perform interpolation calculations on the adjacent mapping points to generate the valve position setting values ​​of each control valve. The online update submodule for flow characteristics is used to acquire the actual valve position feedback signal and the corresponding actual steam flow measurement value of each valve at a preset acquisition time, store the actual valve position feedback signal and the actual steam flow measurement value as new mapping points in the flow characteristic data table, and remove the earliest mapping point in the flow characteristic data table at a preset update cycle. The inverse transformation output submodule is used to send the valve position setting values ​​of each regulating valve to the servo actuator of each regulating valve at a set period.

7. The wide-load variable operating condition steam turbine multivariable adaptive control system according to claim 1, characterized in that: The dynamic decoupling model predictive control layer is equipped with a disturbance feedforward compensation submodule, which is used to identify the current disturbance type based on real-time operating data, retrieve the corresponding feedforward compensation coefficient from the preset feedforward compensation coefficient table according to the disturbance type, generate the feedforward compensation amount based on the feedforward compensation coefficient, and superimpose the feedforward compensation amount onto the multi-channel regulating valve control command. The disturbance types include load command step disturbance type, main steam pressure fluctuation disturbance type, heating extraction steam flow rate change disturbance type, and grid frequency deviation disturbance type.

8. A control method for a multivariable adaptive regulation system of a steam turbine under wide load and variable operating conditions according to any one of claims 1-7, characterized in that, Includes the following steps: Step S1: Real-time acquisition of turbine unit operating parameters, including control valve opening, extraction butterfly valve opening, main steam temperature and pressure, reheat parameters, actual unit power generation, speed, pressure values ​​at each stage, and heat extraction steam flow rate, through the operating condition mode recognition unit and multivariable online identification unit; Step S2: The operating parameters are timestamped and aligned by the operating condition mode recognition unit to generate a multivariable time-series synchronization data sequence. The operating condition type of the current control cycle is identified based on the load command change rate and the measured power change rate in the multivariable time-series synchronization data sequence, and an operating condition type label is generated. Step S3: The multivariate online identification unit selects the model order and time delay parameters for the current identification period according to the working condition type label, and uses the recursive regression method with forgetting factor to identify the multivariate time series synchronous data sequence online, generating the multivariate coupling coefficient matrix under the current operating condition, and storing the multivariate coupling coefficient matrix as a dynamically updated identification parameter set according to the working condition type. Step S4: The dynamic decoupling model prediction control layer retrieves the target identification parameters from the dynamically updated identification parameter set according to the working condition type label, and constructs a multivariate dynamic prediction model based on the target identification parameters; Step S5: The adaptive constraint management module determines the dynamic constraint boundary of the current control cycle based on the operating condition type label, and the dynamic decoupling model prediction control layer determines the target weight vector based on the operating condition type label; Step S6: The dynamic decoupling model predictive control layer takes the current control cycle as the starting time and recursively calculates the predicted output sequence of the controlled variable in the future control cycle based on the multivariate dynamic prediction model within the preset prediction time domain. It constructs a cost function based on the target weight vector and dynamic constraint boundary, solves the cost function within the feasible domain defined by the dynamic constraint boundary, generates the optimal control increment sequence that minimizes the cost function, extracts the control increment of the current control cycle from the optimal control increment sequence, and generates a multi-channel control command by superimposing the output command of the previous control cycle. Step S7: The valve position nonlinear inverse compensation execution layer obtains the flow characteristic mapping relationship of each regulating valve, performs nonlinear inverse transformation correction on the multi-channel regulating valve control command according to the flow characteristic mapping relationship, generates the valve position setting value of each regulating valve, and outputs the valve position setting value of each regulating valve to the servo actuator of each regulating valve.