Thermal power generating unit cooperative control method based on digital twinning

Abstract

The invention relates to the technical field of automatic control, and discloses a thermal power generating unit cooperative control method based on digital twinning, which comprises the following steps: firstly, off-line executing benchmark construction, and obtaining a benchmark digital twinning model and a benchmark cooperative control parameter set which represent the ideal state of a unit; a twinborn deviation vector is obtained by comparing real-time output of a physical unit with theoretical output of a model, a deviation characteristic vector of quantitative characteristic deviation is obtained through processing of a deconstruction algorithm, then historical deviation characteristics are analyzed by adopting a time sequence prediction model, a prediction output vector representing a future evolution trend is calculated, and then a prediction result is obtained. And calculating a prospective control parameter correction vector based on the prediction vector, and combining the prospective control parameter correction vector with a reference parameter set to generate a dynamic cooperative control parameter set. According to the method, the slow time-varying drift of the dynamic characteristics of the unit can be actively compensated, the continuous optimization of a control strategy is realized, and the stability and the operation economy of a control system are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic control, in particular to a thermal power unit cooperative control method based on digital twinning. BACKGROUND

[0002] Thermal power units are the core infrastructure for ensuring power supply, and their stability and economy of operation rely on advanced distributed control systems (DCS). In the DCS, the coordinated control system (CCS) is responsible for establishing dynamic balance between the boiler and the steam turbine to respond to the load instructions of the power grid while maintaining the stability of key parameters such as main steam pressure and temperature.

[0003] However, as a complex nonlinear time-varying system, the dynamic characteristics of thermal power units will drift over time. Factors such as equipment aging, component wear, heat surface coking and contamination can cause the actual operating characteristics of the unit to deviate from the original design benchmark. Traditional control systems are usually based on the design model at the initial stage of unit commissioning or fixed parameter configuration, and it is difficult to accurately and timely quantify the changes in dynamic characteristics. The evaluation of control performance relies on lagging economic and performance indicators, such as the results of automatic generation control (AGC) evaluation or the records of key parameter exceedances, which lack the ability to immediately perceive changes in unit status, resulting in less objective and timely basis for control adjustments.

[0004] Further, when the performance degradation caused by factors such as coking and wear accumulates to a certain extent and leads to a decline in control quality, maintenance personnel only make parameter settings through offline analysis and repeated trial and error. This corrective measure is passive and lacks foresight, unable to predict the trend of performance deterioration and intervene in advance, leaving the system stability at risk. The entire setting process is time-consuming and labor-intensive, and it is difficult to ensure that the adjusted parameters can adapt to future changing operating conditions.

[0005] In addition, existing technologies generally lack a complete closed-loop adaptive control mechanism. Offline setting methods that rely on human experience are not only inefficient, but also have blind spots in the adjustment process, making it difficult to ensure the optimality of the parameters and forming a continuous and automatic optimization loop. Therefore, the unit is difficult to maintain in the most efficient and stable operating range at all times, especially when dealing with complex operating conditions such as frequent load changes and fuel property fluctuations, the control system's adaptability is insufficient, affecting the economy and flexibility of the unit. Therefore, developing a control technology that can actively perceive, predict and continuously optimize is of great significance for improving the overall operating efficiency of thermal power units. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a thermal power unit cooperative control method based on digital twinning, which solves the problems of inaccurate model, poor multi-system coordination, slow response to dynamic working conditions, difficult fault diagnosis and tracing, and difficulty in achieving global optimization in the traditional control method.

[0007] To solve the above technical problems, the present application provides a thermal power unit cooperative control method based on digital twinning. This method first performs a benchmark construction process in the offline stage to obtain a benchmark digital twinning model and a benchmark cooperative control parameter set. The benchmark digital twinning model is used to represent the dynamic response characteristics of the physical thermal power unit in the ideal healthy state, and the benchmark cooperative control parameter set is a set of fixed control parameters obtained by offline simulation optimization based on the model.

[0008] In the online operation stage of the unit, the thermal power unit cooperative control method based on digital twinning performs online deviation feature deconstruction processing, which obtains the control input and actual output of the physical thermal power unit in real time, feeds the real-time control input into the benchmark digital twinning model, and calculates the theoretical output. The difference between the actual output of the physical unit and the theoretical output constitutes a twinning deviation vector. Then, the twinning deviation vector is processed by a deconstruction algorithm to extract and quantify the deviation feature vector that can represent the degree of deviation of the current actual dynamic characteristics of the physical thermal power unit from the ideal state.

[0009] Then, the thermal power unit cooperative control method based on digital twinning performs deviation feature evolution trend prediction processing. This processing uses a time series prediction model to analyze a sequence composed of deviation feature vectors at consecutive historical time points to calculate a predicted output vector. The predicted output vector represents the possible state of the deviation feature vector after a certain preset time step in the future, thereby realizing forward-looking prediction of the change trend of the unit characteristics.

[0010] Subsequently, the thermal power unit cooperative control method based on digital twinning performs forward-looking control parameter reconstruction processing. This processing uses a preset parameter reconstruction mapping relationship to convert the predicted output vector obtained in the foregoing step into a forward-looking control parameter correction vector. The correction vector clearly indicates the specific adjustment amount that needs to be made to the benchmark control parameters to cope with the changes in the unit characteristics in the future. Then, the forward-looking control parameter correction vector is combined with the benchmark cooperative control parameter set to generate a dynamic cooperative control parameter set that can adapt to the current and future state of the unit.

[0011] Finally, the dynamic cooperative control parameter set is used to perform online control to generate and apply real-time control input to the physical thermal power unit, thereby completing a closed loop of adaptive control. This process is continuously iterated to ensure that the control strategy always matches the actual dynamic characteristics of the unit.

[0012] As a preferred technical solution, in the online deviation feature decomposition process, a discrete-time state space model describing the evolution process of the deviation feature is established, the model takes the discrete-time twin deviation vector as the observation output and the deviation feature vector as the internal state to be estimated, the decomposition algorithm adopts the Kalman filtering algorithm, and the posterior optimal estimation value of the deviation feature vector at the current time is obtained by recursively solving the state space model.

[0013] As a preferred technical solution, the Kalman filtering algorithm includes a prediction step and an update step, the prediction step calculates the prior estimation value at the current time according to the posterior optimal estimation value at the previous time, and the update step corrects the prior estimation value at the current time using the discrete-time twin deviation vector at the current time to obtain the posterior optimal estimation value at the current time.

[0014] As a preferred technical solution, the time series prediction model is a long short-term memory network, and when performing the evolution trend prediction process, an input sequence composed of the posterior optimal estimation values of the historical deviation feature vectors of a continuous preset length is fed into the long short-term memory network, and the network calculates the prediction output vector through its output layer. The long short-term memory network is constructed through an offline training step, and during the training process, the network parameters are adjusted by minimizing the mean square error loss function.

[0015] As a preferred technical solution, the parameter reconstruction mapping relationship is a linear transformation, and the prediction output vector is mapped to the prospective control parameter correction vector through the multiplication operation of the parameter reconstruction matrix and the prediction output vector.

[0016] As a preferred technical solution, the process of generating the dynamic cooperative control parameter set is realized through a first-order low-pass filter update algorithm, which fuses and calculates the prospective control parameter correction vector, the reference cooperative control parameter set, and the control parameter vector currently in effect to generate a new dynamic cooperative control parameter set in a smooth transition manner, thereby ensuring the stability of parameter switching. The smoothing degree of the update can be set by adjusting the parameter update weight matrix which is a diagonal matrix.

[0017] As a preferred technical solution, the reference digital twin model is represented in the form of a linear time-invariant state space equation, and the reference cooperative control parameter set is obtained through offline simulation optimization, and the optimization objective is to minimize a preset performance index function which comprehensively considers the tracking error and the consumption of control input in the control process.

[0018] The technical scheme provided by the application quantifies unit characteristic deviation by constructing a digital twin model, and predicts the deviation evolution trend by using a time series prediction model, thereby realizing the foresight and self-adaptive reconstruction of the control parameters.

[0019] The application provides a fire power unit cooperative control method based on digital twinning.

[0020] 1、The application can accurately quantify the dynamic characteristic drift of the physical fire power unit caused by long-term operation by constructing a reference digital twin model and online decomposing the twin deviation, thereby providing an explicit and objective basis for subsequent control adjustment, and improving the accuracy and timeliness of state perception compared with relying on experience judgment or lagging performance indicators.

[0021] 2、The application introduces a prediction link of deviation characteristic evolution trend, so that the adjustment of the control parameters is foresighted.

[0022] 3、The application combines parameter reconstruction and online implementation, and forms a complete closed-loop adaptive control structure. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 It is a general flowchart of the fire power unit cooperative control method based on digital twinning of the application;

[0024] Figure 2 It is a flowchart of the offline reference construction process in the embodiment of the application;

[0025] Figure 3 It is a block diagram of the online deviation characteristic decomposition process in the embodiment of the application;

[0026] Figure 4 It is a schematic diagram of the training data construction of the time series prediction model in the embodiment of the application;

[0027] Figure 5 It is a block diagram of the online process flow in the embodiment of the application;

[0028] Figure 6A block diagram for online reconstruction and implementation of control parameters in the embodiment of the present application. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the protection scope of the present application.

[0030] Referring to the drawings in the specification of the present application, Figure 1 The present application provides a digital-twin-based coordinated control method for a thermal power unit, comprising the following steps:

[0031] Step S1: performing offline reference construction processing to obtain a reference digital-twin model representing an ideal health state of the thermal power unit, and a reference coordinated control parameter set corresponding to the digital-twin-based coordinated control method reference digital-twin model.

[0032] The digital-twin-based coordinated control method reference construction processing is an offline step. The digital-twin-based coordinated control method reference digital-twin model is established by performing characteristic disturbance tests on a physical thermal power unit in a preset health state, collecting its dynamic response data and performing model identification. The digital-twin-based coordinated control method reference coordinated control parameter set is a set of fixed optimal parameters obtained by offline tuning of the digital-twin-based coordinated control method reference digital-twin model for a preset coordinated control strategy (including energy balance control, decoupling control, feedforward control).

[0033] Step S2: performing online deviation feature decomposition processing. This processing obtains real-time control input and real-time output of the physical thermal power unit, and uses the digital-twin-based coordinated control method reference digital-twin model to calculate a twin deviation vector, and then processes the digital-twin-based coordinated control method twin deviation vector through a preset decomposition algorithm to obtain a deviation feature vector representing the current characteristic deviation degree of the physical thermal power unit.

[0034] Specifically, the real-time control input applied to the physical thermal power unit is input synchronously to the benchmark digital-twin-based thermal power unit collaborative control method digital twin model to obtain a twin theory output. Then, a difference between the twin theory output of the digital-twin-based thermal power unit collaborative control method and the real-time output of the collected physical thermal power unit is calculated to generate a digital-twin-based thermal power unit collaborative control method twin deviation vector. Subsequently, a digital-twin-based thermal power unit collaborative control method deconstruction algorithm is performed on the digital-twin-based thermal power unit collaborative control method twin deviation vector to extract and quantify the characteristic information associated with the specific dynamic characteristic changes (e.g., changes in heat transfer efficiency, changes in rotational inertia, changes in process coupling degree) of the physical thermal power unit contained therein, thereby obtaining a digital-twin-based thermal power unit collaborative control method deviation characteristic vector.

[0035] Step S3: Perform deviation characteristic evolution trend prediction processing. Based on the sequence of digital-twin-based thermal power unit collaborative control method deviation characteristic vectors at historical time points, the prospective deviation characteristic vector representing the state of the digital-twin-based thermal power unit collaborative control method deviation characteristic vector after a preset time step in the future is calculated through a time series prediction model.

[0036] This step predicts the future trend of characteristic drift caused by slow time-varying factors such as equipment wear and coke accumulation by modeling the change law of the digital-twin-based thermal power unit collaborative control method deviation characteristic vector over time. The digital-twin-based thermal power unit collaborative control method time series prediction model is based on offline training or online adaptive updating of historical data.

[0037] Step S4: Perform prospective control parameter reconstruction processing. This processing is based on the digital-twin-based thermal power unit collaborative control method deviation characteristic vector and the digital-twin-based thermal power unit collaborative control method prospective deviation characteristic vector, and through a preset parameter reconstruction mapping relationship, a parameter correction vector is calculated. Then, the digital-twin-based thermal power unit collaborative control method parameter correction vector is combined with the digital-twin-based thermal power unit collaborative control method benchmark collaborative control parameter set to generate a dynamic collaborative control parameter set.

[0038] Specifically, the parameter reconstruction mapping relationship of the digital-twin-based thermal power unit cooperative control method calculates the specific correction value of each parameter in the benchmark cooperative control parameter set according to the current characteristic deviation degree and the future deviation trend, and forms a parameter correction vector of the digital-twin-based thermal power unit cooperative control method. Subsequently, the dynamic cooperative control parameter set of the digital-twin-based thermal power unit cooperative control method is obtained by algebraically superimposing the parameter correction vector of the digital-twin-based thermal power unit cooperative control method and the benchmark cooperative control parameter set of the digital-twin-based thermal power unit cooperative control method. Finally, the dynamic cooperative control parameter set is used by the cooperative control system of the thermal power unit to perform online control, so as to generate real-time control inputs applied to the physical thermal power unit.

[0039] Step S5: online updating and implementing step of control strategy. This step is the final execution link of the adaptive control method proposed in the present application, which directly acts on the distributed control system (DCS) of the physical thermal power unit, and realizes the closed-loop iterative optimization of the control strategy. The purpose of this step is to integrate the dynamic cooperative control parameter set of the digital-twin-based thermal power unit cooperative control method generated by step S4 into the currently running control parameters in a smooth, stable and safe manner, and finally download to the control system of the physical thermal power unit.

[0040] Specifically, before downloading and executing the dynamic cooperative control parameter set of the digital-twin-based thermal power unit cooperative control method, the system also includes a safety check module responsible for checking whether each parameter value in the new parameter set is within its preset safety constraint range. Only after all parameters pass the check, the update operation is authorized to be executed.

[0041] Subsequently, the dynamic cooperative control parameter set of the digital-twin-based thermal power unit cooperative control method that passes the safety check is downloaded to the distributed control system of the physical thermal power unit through a dedicated data interface. The downloading operation is designed as an atomic operation to ensure that all parameters in the parameter vector are updated synchronously in the same control period, thereby avoiding system transient oscillation caused by inconsistent parameters.

[0042] At the next control time, the new dynamic cooperative control parameter set of the digital-twin-based thermal power unit cooperative control method officially takes effect, replacing the original parameter set to guide the operation control of the thermal power unit. At this point, a complete forward-looking, adaptive closed-loop control parameter optimization is completed. The online updating and implementing step will continue to be executed iteratively at each control period as time goes on, so as to continuously adapt the cooperative control strategy of the thermal power unit to the changes in its dynamic characteristics.

[0043] Referring to the accompanying Figure 1 and the accompanying Figure 2The method of the present application first performs a step of establishing a reference digital twin model in step S1. The purpose of this step is to obtain a mathematical model that can accurately describe the dynamic response characteristics of the physical thermal power unit in an ideal healthy state.

[0044] Specifically, this step first needs to determine a preset healthy state of the physical thermal power unit. In one specific embodiment, the preset healthy state based on the digital-twin-based thermal power unit collaborative control method refers to the state of the boiler heating surface being clean after the thermal power unit completes a planned maintenance, and the unit is stably operated at a preset reference load operating point. The reference load operating point based on the digital-twin-based thermal power unit collaborative control method is, for example, 75% of the rated load.

[0045] When the thermal power unit is in the preset healthy state based on the digital-twin-based thermal power unit collaborative control method, a unit characteristic disturbance test is performed. In one specific embodiment, the test injects a preset disturbance signal into one or more control input channels (such as coal supply command, water supply flow command, or turbine main steam valve opening command). The disturbance signal based on the digital-twin-based thermal power unit collaborative control method can be a step signal or a pseudo-random binary sequence (PRBS) signal. At the same time of injecting the disturbance signal, the input data sequence of the control input channel and the output data sequence of a plurality of key process output variables based on the digital-twin-based thermal power unit collaborative control method are synchronously collected and recorded. The key process output variables based on the digital-twin-based thermal power unit collaborative control method include, for example, main steam pressure, unit active power, and main steam temperature.

[0046] After completing the data collection, the input data sequence and the output data sequence collected are used to identify the parameters of the reference digital twin model based on the digital-twin-based thermal power unit collaborative control method using a preset system identification algorithm. In one specific embodiment, the system identification algorithm based on the digital-twin-based thermal power unit collaborative control method is a subspace identification method.

[0047] In one specific embodiment, the reference digital twin model based on the digital-twin-based thermal power unit collaborative control method is represented in the form of a linear time-invariant state space equation, and its mathematical expression is as follows:

[0048]

[0049] wherein u(t) represents a control input vector at time t, each component of the control input vector being a signal value of the control input channel subjected to disturbance in the unit characteristic disturbance test based on the digital-twin-based thermal power unit collaborative control method, such as coal supply command, water supply flow command, etc.; x(t) represents a state derivative vector of the digital twin model at time t; x dt(t) represents the state vector of the benchmark digital twin model of the digital-twin-based thermal power unit cooperative control method at time t, each component of the state vector being a state variable representing the dynamic characteristics of the unit, such as the main metal wall temperature, the boiler heat storage, etc.; y dt (t) represents the theoretical output vector of the benchmark digital twin model of the digital-twin-based thermal power unit cooperative control method at time t, each component of the theoretical output vector being a theoretical value of a key process output variable of the digital-twin-based thermal power unit cooperative control method, such as the main steam pressure, the unit active power, etc.; A0 represents the state matrix of the benchmark digital twin model of the digital-twin-based thermal power unit cooperative control method; B0 represents the input matrix of the benchmark digital twin model of the digital-twin-based thermal power unit cooperative control method; C0 represents the output matrix of the benchmark digital twin model of the digital-twin-based thermal power unit cooperative control method; C0 represents the feedforward matrix of the benchmark digital twin model of the digital-twin-based thermal power unit cooperative control method.

[0050] Through the above steps, a set of state space matrices A0, B0, C0, C0 capable of representing the dynamic characteristics of the thermal power unit in an ideal healthy state can be obtained, thereby completing the establishment of the benchmark digital twin model.

[0051] In step S1, after completing the establishment step of the benchmark digital twin model, a construction step of the benchmark cooperative control strategy is performed. The purpose of this step is to design and set a set of cooperative control strategies offline based on the established benchmark digital twin model of the digital-twin-based thermal power unit cooperative control method, so as to obtain a fixed benchmark cooperative control parameter set as the basis for subsequent online adjustment.

[0052] In a specific embodiment, the cooperative control strategy of the digital-twin-based thermal power unit cooperative control method includes a bidirectional decoupling control strategy, a dynamic feedforward control strategy, and a direct energy balance control strategy. The design of these strategies is based on the dynamic characteristics of the benchmark digital twin model of the digital-twin-based thermal power unit cooperative control method (described by the state space matrices A0, B0, C0, C0).

[0053] The bidirectional decoupling control strategy of the digital-twin-based thermal power unit cooperative control method aims to reduce the coupling effect between the boiler dynamics and the turbine dynamics. The dynamic feedforward control strategy of the digital-twin-based thermal power unit cooperative control method aims to pre-adjust the control input using the change of the load instruction. The direct energy balance control strategy of the digital-twin-based thermal power unit cooperative control method aims to calculate the benchmark fuel input amount according to the energy demand corresponding to the load instruction.

[0054] The core output of the construction steps of the benchmark collaborative control strategy is a benchmark collaborative control parameter set, denoted as P0. The benchmark collaborative control parameter set P0 contains all the adjustable parameters required to implement the collaborative control strategy described above, and its mathematical expression is as follows:

[0055] P0={p 0,1 ,p 0,2 ,…,p 0,k};

[0056] In a specific embodiment, the components of the benchmark collaborative control parameter set P0 include, for example, the element values of the decoupling matrix in the two-way decoupling control strategy of the digital-twin-based collaborative control method for thermal power generating units; the feedforward gain and time constant in the dynamic feedforward control strategy of the digital-twin-based collaborative control method for thermal power generating units; and the correction coefficient for heat value calculation in the direct energy balance control strategy of the digital-twin-based collaborative control method for thermal power generating units.

[0057] The specific numerical values of the benchmark collaborative control parameter set P0 of the digital-twin-based collaborative control method for thermal power generating units are obtained through offline simulation optimization on the benchmark digital twin model of the digital-twin-based collaborative control method for thermal power generating units. In a specific embodiment, the optimization process adjusts the numerical values of the benchmark collaborative control parameter set P0 to minimize a preset performance index function J, and an example of the performance index function J of the digital-twin-based collaborative control method for thermal power generating units is as follows:

[0058]

[0059] where T sim represents the simulation duration; e err (t) represents the tracking error vector between the unit output set value and the model output value during the simulation process; Δu(t) represents the change rate vector of the control input during the simulation process; Q err and R u represent the preset positive definite weight matrices for balancing the tracking performance and the control input smoothness; w e and w u represent scalar weight coefficients.

[0060] By executing an algorithm (such as the sequential quadratic programming method) to solve the optimization problem described above, a set of parameter values that optimizes the performance index function J can be obtained, and this set of parameter values constitutes the benchmark collaborative control parameter set P0 of the digital-twin-based collaborative control method for thermal power generating units.

[0061] Referring to the accompanying Figure 1 and the accompanying Figure 3The method of the present application, in step S2, first performs a real-time calculation step of the twin deviation vector. This step is continuously performed during the online operation of the thermal power unit, and its purpose is to quantize in real time the deviation between the actual operating state of the physical thermal power unit and the ideal state described by the reference digital twin model.

[0062] In one specific embodiment, this step first acquires in real time the control input vector u(t) applied to each actuator of the physical thermal power unit at the current time t from the distributed control system (DCS) of the thermal power unit. At the same time, the measured values of the key process output variables measured by the sensors on the physical thermal power unit are collected in real time from the distributed control system of the digital twin-based thermal power unit collaborative control method to form the actual output vector y p (t) of the physical unit.

[0063] Step S2 synchronously inputs the acquired real-time control input vector u(t) into the reference digital twin model of the digital twin-based thermal power unit collaborative control method established in step S1, and the reference digital twin model of the digital twin-based thermal power unit collaborative control method performs numerical integration operation based on the state space matrix (A0, B0, C0, C0) inside it, so as to calculate the theoretical output vector y dt (t) that the unit should have under the same control input vector u(t) when it is in an ideal healthy state. In this process, the collection of the actual output vector y p (t) of the physical unit and the calculation of the theoretical output vector y dt (t) of the reference digital twin model are synchronized based on a unified time stamp to ensure consistency of comparison.

[0064] After obtaining the actual output vector y p (t) of the physical unit and the theoretical output vector y dt (t) of the reference digital twin model at the same time t, vector subtraction operation is performed to generate the continuous-time twin deviation vector e(t), whose calculation formula is as follows:

[0065] e(t) = y p (t) - y d t(t);

[0066] Wherein, y p (t) represents the actual output vector collected from the physical thermal power unit at time t, and each component is the measured value of a key process variable, such as the actual measured value of the main steam pressure and the actual measured value of the unit active power; y dt(t) represents the theoretical output vector of the reference digital twin model calculated by the digital-twin-based coordinated control method for thermal power generating units at time t, the physical meaning of each component of which is the same as that of the actual output vector y p (t) of the digital-twin-based coordinated control method for thermal power generating units, for example, the theoretically calculated value of the main steam pressure and the theoretically calculated value of the unit active power; e(t) represents the continuous-time twin deviation vector calculated at time t, each component of which represents the deviation between the actual value and the theoretical value of the key process output variable.

[0067] By repeatedly performing the above steps, the twin deviation vector flow updated over time can be obtained. The continuous-time twin deviation vector e(t) obtained at any time t will be the direct input data of the subsequent dynamic deconstruction step of the deviation feature at the corresponding time.

[0068] The method of the present application, in step S2, after performing the real-time calculation step of the twin deviation vector, then performs the dynamic deconstruction step of the deviation feature. The purpose of this step is to extract and quantify the feature information with clear physical meaning that can represent the deviation of the specific dynamic characteristics of the physical thermal power generating unit from the continuous-time twin deviation vector e(t) calculated in the previous step.

[0069] The continuous-time twin deviation vector e(t) of the digital-twin-based coordinated control method for thermal power generating units is a mixed signal under the joint action of multiple physical characteristic deviations and the influence of measurement noise. In order to realize the targeted adjustment of the subsequent control parameters, it is necessary to perform the dynamic deconstruction step of the deviation feature to decompose the mixed signal of the digital-twin-based coordinated control method for thermal power generating units into a group of independent deviation feature components.

[0070] In a specific embodiment, the dynamic deconstruction step of the deviation feature is realized by performing a state estimation algorithm based on the Kalman filtering algorithm, which runs in the discrete time domain. Therefore, first, the continuous-time twin deviation vector e(t) of the digital-twin-based coordinated control method for thermal power generating units is sampled at a preset sampling period T s The continuous-time twin deviation vector e(t) of the digital-twin-based coordinated control method for thermal power generating units is sampled to obtain a discrete-time observation sequence. The sampling value obtained at the discrete time k (corresponding to the continuous time t = k·T s ) is denoted as the discrete-time twin deviation vector e k .

[0071] The dynamic deconstruction step of the deviation feature establishes a discrete-time state space model describing the evolution process of the deviation feature, and takes the discrete-time twin deviation vector e k as the observation output of the system. The mathematical expression of this discrete-time state space model is as follows:

[0072]

[0073] where e k represents a discrete-time twin deviation vector sampled from the continuous-time twin deviation vector e(t) at discrete time instant k as the observation input; F k represents a real deviation feature vector which is not directly measurable at discrete time instant k; A f represents a feature transition matrix describing the evolution relationship of the deviation feature from k-1 time instant to k time instant; C f represents an observation matrix describing the linear relationship between the real deviation feature vector F k and the discrete-time twin deviation vector e k ; w k-1 represents a process noise vector at k-1 time instant; v k represents a measurement noise vector at k time instant.

[0074] In one specific embodiment, each component of the real deviation feature vector F k is associated with a specific physical property deviation. For example, the first component thereof can represent the heat transfer efficiency attenuation amount caused by the coking of the boiler heating surface; and the second component thereof can represent the rotational inertia change amount caused by the turbine rotor wear.

[0075] The discrete-time state space model is recursively solved by a Kalman filtering algorithm, which includes a prediction step and an update step at each discrete time instant k.

[0076] The calculation formula of the prediction step is as follows:

[0077]

[0078] The calculation formula of the update step is as follows:

[0079]

[0080] S k|k = (I - K k C f ) S k|k-1 ;

[0081] In the above formulas, A f represents a feature transition matrix describing the evolution relationship of the deviation feature from k-1 time instant to k time instant; C f represents an observation matrix describing the linear relationship between the real deviation feature vector F k and the discrete-time twin deviation vector e k ; Q f represents the covariance matrix of the process noise vector w k-1 ; and R fThis indicates a digital twin-based collaborative control method for thermal power units, measuring the noise vector v. k The covariance matrix; This represents the posterior optimal estimate at time k-1, which is the result calculated at the previous time step. This represents the prior estimate of the state at time k based on the state at time k-1. This represents the observation value e at time k. k After correction, the posterior optimal estimate of the state at time k; S k-1|k-1 S represents the covariance matrix of the posterior best estimate at time k-1, i.e., the covariance result at the previous time step; k|k-1 Represents the prior estimate The covariance matrix of S; k|k This represents the posterior optimal estimate of the collaborative control method for thermal power units based on digital twins. The covariance matrix; K k Let represent the Kalman gain matrix calculated at time k; I represents the identity matrix; (·) T This represents the transpose operation of a matrix or vector; (·) -1 This represents the inverse operation of a matrix.

[0082] By repeating the prediction and update steps at each discrete time k, the posterior optimal estimate of the deviation eigenvector of the digital twin-based thermal power unit cooperative control method at the current time can be obtained.

[0083] See attached document Figure 4 The time series prediction model used in step S3 of this invention is pre-established through offline construction and training steps before the online prospective prediction step is executed. The details of the construction and training steps are described below.

[0084] First, the historical training dataset is collected and constructed. This step requires the thermal power unit to run online for a preset long period (e.g., several weeks or months) after steps S1 and S2 have been deployed, and during this period, continuously record the posterior optimal estimate sequence output by the deviation feature dynamic deconstruction step in step S2. This yields a time series of deviation feature vectors containing N historical samples.

[0085] Next, a sliding window process is performed on the collected historical time series to construct input-output sample pairs for supervised learning. Specifically, at any discrete time k in the sequence, the input sequence X is defined. k and the corresponding target output vector

[0086] A digital twin-based collaborative control method for thermal power units: Input sequence Xk is composed of the posterior optimal estimation values of the consecutive L historical deviation feature vectors from the time k-L+1 to the time k:

[0087]

[0088] wherein, denotes the posterior optimal estimation value at the starting time of the input sequence (i.e. the time k-L+1); denotes the posterior optimal estimation value at the time k-1; denotes the posterior optimal estimation value of the deviation feature vector at the time k.

[0089] The target output vector Y k is the deviation feature vector estimation value at the Hth time after the time k:

[0090]

[0091] wherein, denotes the posterior optimal estimation value at the future time k+H.

[0092] By performing the above operation on the historical time sequence from k=L to k=N-H, a training dataset containing M=N-L-H+1 training sample pairs (X k ,Y k ) can be generated, wherein N represents the total length of the collected historical time sequence.

[0093] In one specific embodiment, the time series prediction model of the digital-twin-based coordinated control method for thermal power generating units adopts a long short-term memory (LSTM) network structure. The network contains one or more LSTM layers and a final output layer, and each LSTM layer is composed of multiple LSTM units. When processing the input sequence X k , the LSTM unit performs the following calculation at each time τ (τ from 1 to L) of the sequence:

[0094] f τ =σ(W f ·[h τ-1 ,x τ ]+b f );

[0095] i τ =σ(W i ·[h τ-1 ,x τ ]+b i );

[0096]

[0097] o τ =σ(Wo ·[h τ-1 ,x τ ]+b o );

[0098] h τ =o τ ⊙tanh(C τ );

[0099] In the above equations, x τ denotes the input vector at the τ-th time step of the input sequence X k h τ-1 denotes the hidden state output vector of the LSTM unit at the (τ-1)-th time step; C τ-1 denotes the cell state vector of the LSTM unit at the (τ-1)-th time step; W f ,W i ,W C ,W o denote weight matrices; b f ,b i ,b C ,b o denote bias vectors; f τ ,i τ ,o τ denote the activation vectors of the forget gate, input gate and output gate at the τ-th time step, respectively; denotes the candidate cell state vector at the τ-th time step; C τ denotes the updated cell state vector at the τ-th time step; h τ denotes the hidden state output vector at the τ-th time step; σ(·) denotes the Sigmoid activation function; tanh(·) denotes the hyperbolic tangent activation function; [·,·] denotes the concatenation operation of two vectors; and denotes the element-wise multiplication of vectors.

[0100] After the LSTM network processes the last time step of the input sequence X k , the final hidden state vector h L is obtained, which is fed into a fully connected output layer to calculate the predicted output vector of the model

[0101]

[0102] where W out and b out are the weight matrix and bias vector of the output layer, respectively.

[0103] Finally, the training step of the model is performed. This step minimizes a pre-defined loss function ​This is used to adjust all weight matrices and bias vectors in the network. In one specific implementation, mean squared error (MSE) is used as the loss function:

[0104]

[0105] Among them, Y k Let represent the target output vector constructed at time k; The model represents the input X. k The resulting predicted output vector; The L2 norm of the vector is the square of the vector, which is the sum of the squares of the vector's components; M represents the total number of sample pairs in the constructed training dataset.

[0106] This minimization process is performed using a gradient descent optimization algorithm (such as the Adam algorithm), with the loss function calculated via backpropagation. Calculate the gradient of each parameter and update the parameters accordingly. Repeat this process until the loss function is calculated. When the value converges to below a preset threshold, the training of the time series prediction model is complete.

[0107] See attached document Figure 1 and attached Figure 5 In step S3 of this invention, after the construction and training steps of the time series prediction model have been performed in advance, the generation step of prospective bias features is performed online. This generation step of prospective bias features is the online execution part of step S3.

[0108] The purpose of this step is to use the offline-trained time series prediction model to generate a predicted value of the deviation feature vector for the next H steps at each discrete time k, based on the current and most recent historical deviation feature vector estimates.

[0109] In one specific implementation, at the current discrete time k, the forward bias feature generation step first obtains the latest posterior optimal estimates of the L consecutive bias feature vectors output by step S2 from the real-time data buffer.

[0110] These estimates constitute the input sequence X used for prediction. k Its mathematical expression is completely consistent with that defined during the training phase:

[0111]

[0112] Subsequently, the forward bias feature generation step uses the constructed input sequence X k The data is fed into the trained temporal prediction model. The temporal prediction model performs one forward propagation calculation, which is exactly the same as described in the training steps, and finally generates the prediction output vector. It represents the prediction of the deviation eigenvector at the future time k+H.

[0113] The predicted output vector generated in this step This will be used as input data for the subsequent optimization generation step of the forward-looking control parameters.

[0114] In step S4 of this invention, a forward-looking correction amount generation step is performed. This step is a crucial step in predicting the connection deviation characteristics and adjusting the actual control parameters.

[0115] The purpose of the forward correction generation step is to transform the predicted output vector generated in step S3 into a forward correction vector. This is transformed into a set of forward-looking control parameter correction vectors with clear physical meanings for the specific control parameters in the current control system of the thermal power unit.

[0116] In one specific implementation, the forward-looking correction is generated through a predefined linear transformation. This linear transformation is achieved by the parameter reconstruction matrix G. r By definition, this parameter reconstruction matrix establishes a mapping relationship from the deviation feature space to the control parameter correction space.

[0117] The forward correction generation step receives the predicted output vector from step S3 at the current time k. The following calculations are performed to generate the forward-looking control parameter correction vector ΔU. k+H :

[0118]

[0119] Wherein, ΔU k+H G represents the forward-looking control parameter correction vector; r Represents the parameter reconstruction matrix; This represents the predicted output vector.

[0120] Parameter Reconstruction Matrix G of Cooperative Control Method for Thermal Power Units Based on Digital Twin r It is a constant matrix that is predetermined before the system goes online and runs. This parameter reconstructs each element g of the matrix. ab All have clear physical meanings, and their numerical values ​​quantify the characteristics of the b-th deviation component. When the b-th deviation characteristic component changes by a unit, the a-th control parameter (ΔU) k+H The amount of compensation correction required for the a-th control parameter.

[0121] Parameter reconstruction matrix G r The determination is based on the results of mechanistic model analysis and system identification of thermal power units. For example, the matrix parameter reconstruction matrix G. rthe first column vector defines the collaborative adjustment amount of the PID parameter of the main steam temperature control loop, the feedforward gain of the feedwater flow, and other control parameters, which is required to completely offset the performance deviation caused by the first deviation feature (e.g., the deterioration of the coking of the boiler heating surface).

[0122] The prospective control parameter correction vector ΔU generated in this step k+H will be used as the core input data of the online reconstruction step of the final execution of the prospective control parameter, to guide the actual adjustment of the control system parameters of the physical thermal power generating unit.

[0123] Referring to the accompanying drawings Figure 1 and the accompanying drawings Figure 6 , in step S5, the online reconstruction step of the prospective control parameter is performed, which is the final execution link of the adaptive control method proposed in the present application and directly acts on the control system of the physical thermal power generating unit.

[0124] The purpose of the online reconstruction step is to integrate the prospective control parameter correction vector ΔU k+H representing the ideal correction direction generated in step S4 into the currently running control parameters in a smooth, stable, and safe manner, thereby generating a set of new control parameters to be put into effect after optimization.

[0125] In a specific embodiment, the online reconstruction step is implemented by a first-order low-pass filter update algorithm. The first-order low-pass filter update algorithm receives ΔU k+H output by step S4 at the current time k, and combines the control parameter vector U k currently in effect at the current time, and a set of reference control parameter vectors U base to calculate the new control parameter vector U k+1 to be put into effect at the next time k+1.

[0126] The calculation formula of the update algorithm is as follows:

[0127] U k+1 = (I-W p )U k +W p (U base +ΔU k+H );

[0128] wherein U k represents the control parameter vector currently in effect in the physical thermal power generating unit control system at the current time k; U k+1 represents the new control parameter vector to be put into effect at the next time k+1; and U baserepresents a set of reference control parameter vectors, which are the design parameter values of the thermal power unit under ideal, unbiased working conditions; ΔU k+H represents a prospective control parameter correction vector; I represents an identity matrix; W p represents a parameter update weight matrix.

[0129] The parameter update weight matrix W of the digital-twin-based cooperative control method of the thermal power unit p is a diagonal matrix, and its determination process is the tuning step of the weight matrix. The parameter update weight matrix is pre-tuned offline before the system goes online. Its mathematical expression is as follows:

[0130] W p = diag (α1, α2, …, αp) ; p ) ;

[0131] Wherein, diag represents a diagonal matrix generation function; p is the total number of control parameters to be reconstructed.

[0132] The diagonal elements α c (c = 1, …, p) of the digital-twin-based cooperative control method of the thermal power unit are filter coefficients between 0 and 1, and the tuning basis of their values is the process characteristics and safety requirements of the cth control parameter corresponding to them. The tuning process is based on expert knowledge and simulation verification, aiming to achieve a balance between response speed and system stability. For control loops with fast response requirements and insensitive to system disturbances, the corresponding coefficient α c takes a larger value;

[0133] For key control loops with stable response requirements and significant impact on system stability (such as main steam temperature), the corresponding coefficient α c takes a smaller value to ensure the smoothness and safety of parameter changes.

[0134] The final new control parameter vector U k+1 generated by the online reconstruction step will be downloaded to the distributed control system of the physical thermal power unit to replace the original control parameter vector U k at k+1 time, thereby completing a complete prospective, adaptive closed-loop control parameter optimization.

[0135] In step S5, the online update and implementation step of the control strategy is performed. The online update and implementation step of the control strategy is the final execution link of the adaptive control method proposed in the present application, which directly acts on the control system of the physical thermal power unit to realize closed-loop iterative optimization of the control strategy.

[0136] The purpose of the online update and implementation step of the control strategy is to implement the prospective control parameter correction vector ΔUk+H , and integrates into the currently running control parameters in a smooth, stable and safe manner, thereby generating a set of new control parameters that will take effect soon after being optimized.

[0137] In a specific embodiment, the online updating and implementing step of the control strategy is implemented by a first-order low-pass filter updating algorithm, which receives ΔU k+H output by step S4 at the current time k k , and combines the control parameter vector U base that is currently taking effect at the current time, to calculate the new control parameter vector U k+1 that will take effect at the next time k+1.

[0138] The implementation of the online updating and implementing step of the control strategy involves downloading the calculated final new control parameter vector U k+1 to the distributed control system of the physical thermal power unit through a dedicated data interface. In a specific embodiment, the downloading operation is designed as an atomic operation to ensure that all parameters in the parameter vector are updated synchronously within the same control period, thereby avoiding system transient oscillation caused by inconsistent parameters.

[0139] Before downloading is executed, the system also includes a safety verification module responsible for verifying whether each parameter value in U k+1 is within its preset safety constraint range, and only after all parameters have passed the verification will the updating operation be authorized to be executed.

[0140] At the next time k+1, the new control parameter vector U k+1 officially takes effect, replacing the original U k to guide the operation control of the thermal power unit. Thus, a complete forward-looking, adaptive closed-loop control parameter optimization is completed. The online updating and implementing step will continue to be executed iteratively at each control period as time k advances.

Claims

1. A method for coordinated control of a thermal power generating unit based on digital twinning, characterized in that, Comprise: S1, performing offline reference construction processing, obtaining reference digital twin model and reference collaborative control parameter set; S2, performing online deviation feature decomposition processing, obtaining real-time control input and real-time output of the physical thermal power unit, calculating the twin deviation vector by using the reference digital twin model, and obtaining the deviation feature vector by processing the twin deviation vector through the decomposition algorithm; S3, performing evolution trend prediction processing of the deviation feature, based on the sequence of deviation feature vectors at historical time, calculating the predicted output vector through the time series prediction model; S4, performing forward-looking control parameter reconstruction processing, based on the predicted output vector, calculating the forward-looking control parameter correction vector through the parameter reconstruction mapping relationship, and combining the forward-looking control parameter correction vector with the reference collaborative control parameter set to generate a dynamic collaborative control parameter set; S5, using the dynamic collaborative control parameter set to perform online control to generate real-time control input applied to the physical thermal power unit.

2. The digital-twin-based coordinated control method of a thermal power generating unit according to claim 1, characterized in that, In step S2, the online deviation feature decomposition processing comprises: A discrete-time state space model describing the evolution process of the deviation feature is established, and the discrete-time twin deviation vector obtained by sampling the twin deviation vector is taken as the system observation output, the discrete-time state space model comprises a state equation and an observation equation, the state equation is used to describe the evolution relationship of the deviation feature vector from the last discrete time to the current discrete time, and the observation equation is used to describe the linear relationship between the deviation feature vector and the discrete-time twin deviation vector; The decomposition algorithm is a Kalman filter algorithm, and the Kalman filter algorithm is used to recursively solve the discrete-time state space model to obtain the posterior optimal estimate value of the deviation feature vector at the current time.

3. The digital-twin-based coordinated control method of a thermal power generating unit according to claim 2, characterized in that, The Kalman filter algorithm comprises a prediction step and an update step: The prediction step is used to calculate the prior estimate value at the current time and the covariance matrix of the prior estimate value according to the posterior optimal estimate value at the last time; The update step is used to calculate the Kalman gain matrix, and correct the prior estimate value by using the Kalman gain matrix and the discrete-time twin deviation vector at the current time to obtain the posterior optimal estimate value and the covariance matrix of the posterior optimal estimate value.

4. The digital-twin-based coordinated control method of a thermal power generating unit according to claim 1, characterized in that, In step S3, the time series prediction model is a long short-term memory network, and the evolution trend prediction processing of the deviation feature comprises: An input sequence composed of historical deviation feature vector posterior optimal estimate values of a continuous preset length is fed into the long short-term memory network, and the long short-term memory network is used to calculate the predicted output vector through an output layer in the long short-term memory network; The calculation of the output layer is the product of the hidden state vector output by the long short-term memory network after processing the last time step of the input sequence and the weight matrix of the output layer, plus the bias vector of the output layer.

5. The digital-twin-based coordinated control method of a thermal power generating unit according to claim 4, characterized in that, The long short-term memory network is constructed through an offline training step, and the offline training step adjusts the network parameters by minimizing the mean square error loss function; The calculation of the mean square error loss function is to sum the square of the norm of the difference between the target output vector and the predicted output vector on all training samples, and then take the average.

6. The digital-twin-based coordinated control method of a thermal power generating unit according to claim 1, characterized in that, In step S4, the parameter reconstruction mapping relationship is a linear transformation, and the execution of the forward-looking control parameter reconstruction processing includes: mapping the predicted output vector into the forward-looking control parameter correction vector through the linear transformation; The calculation of the linear transformation is the product of the parameter reconstruction matrix and the predicted output vector.

7. The digital-twin-based coordinated control method of a thermal power generating unit according to claim 1, characterized in that, In step S4, the generation of the dynamic cooperative control parameter set is to combine the forward-looking control parameter correction vector and the reference cooperative control parameter set through a first-order low-pass filter update algorithm. The calculation of the first-order low-pass filter update algorithm is the product of the result of subtracting the parameter update weight matrix from the unit matrix and the control parameter vector currently in effect, plus the product of the parameter update weight matrix and the sum of the reference cooperative control parameter set and the forward-looking control parameter correction vector.

8. The digital-twin-based coordinated control method of a thermal power generating unit according to claim 7, characterized in that, The parameter update weight matrix is a diagonal matrix, and the diagonal elements of the diagonal matrix are filter coefficients between 0 and 1.

9. The digital-twin-based coordinated control method of a thermal power generating unit according to claim 1, characterized in that, In step S1, the obtained reference digital twin model is in the form of a linear time-invariant state space equation; The linear time-invariant state space equation includes: a state equation for describing that the first-order derivative of a state vector is equal to the product of a state matrix and the state vector, plus the product of an input matrix and a control input vector; an output equation for describing that a theoretical output vector is equal to the product of an output matrix and the state vector, plus the product of a feedforward matrix and the control input vector.

10. The digital-twin-based coordinated control method of a thermal power generating unit according to claim 1, characterized in that, In step S1, the reference cooperative control parameter set is obtained through offline simulation optimization, and the objective of the offline simulation optimization is to minimize a preset performance index function; The calculation of the performance index function is the integral in a preset simulation time length, and the integrand is the sum of two terms, the first term is the product of a scalar weight coefficient and a quadratic form of a tracking error vector, and the second term is the product of another scalar weight coefficient and a quadratic form of a change rate vector of the control input.

Images