Thermal power plant auxiliary system variable working condition self-adaptive robust control method and system

Abstract

This invention discloses an adaptive robust control method and system for auxiliary equipment systems in thermal power plants under varying operating conditions, specifically relating to the field of auxiliary equipment system control technology in thermal power plants. The method includes the following steps: S1, constructing a dual-network architecture; S2, using the prediction model network to perform multi-step prediction of the controlled variables of the auxiliary equipment system, obtaining a predicted state sequence; S3, inputting the current operating condition characteristics, the controlled variable deviation, and the predicted state sequence into the controller network, and the controller network outputting the original control quantity. This invention achieves accurate prediction and adaptive control of the dynamic characteristics of the auxiliary equipment system; ensures the stability of the reinforcement learning controller output, overcoming the insufficient security of traditional black-box models; utilizes a covariance matrix adaptive evolution strategy to perform offline optimization and incremental updates of controller parameters, achieving continuous self-optimization of the controller; and simultaneously realizes equipment degradation perception and active compensation, ensuring continuous operation of the system under sensor failure conditions.

Description

Technical Field

[0001] This invention relates to the field of control technology for auxiliary equipment systems in thermal power plants, and more specifically, to an adaptive robust control method and system for auxiliary equipment systems in thermal power plants under varying operating conditions. Background Technology

[0002] Auxiliary systems in thermal power plants include key equipment such as coal mills, feedwater pumps, induced draft fans, and forced draft fans. Their operating characteristics are typically nonlinear, strongly coupled, and have large time lags. As the power grid's requirements for deep peak shaving and rapid load changes in thermal power units continue to increase, auxiliary systems need to maintain stable operation under frequent changing operating conditions. Traditional proportional-integral-derivative (PID) control strategies use fixed parameters, which are difficult to adapt to model parameter perturbations caused by changes in operating conditions, resulting in a significant decline in control quality. Although existing model predictive control methods can handle some nonlinearities, their predictive models are usually based on fixed models identified offline. When operating conditions change, there is a model mismatch problem, which leads to reduced control accuracy, frequent actuator movements, and accelerated equipment wear.

[0003] Existing technologies have developed solutions that combine neural networks with model predictive control, but the following problems still exist: First, the online correction capability of the model under varying operating conditions is insufficient, making it difficult to cope with rapid load changes; second, the output of intelligent control algorithms such as reinforcement learning lacks stability assurance, posing safety hazards; third, the controller parameters are fixed and cannot be continuously optimized as equipment ages and coal quality changes; and fourth, there is a lack of fault tolerance mechanisms for actuator degradation and sensor failures, so the system reliability needs to be improved.

[0004] Therefore, how to construct an auxiliary machine control method that can adaptively adjust under varying operating conditions, possess robust stability, and continuously evolve has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] To overcome the above-mentioned deficiencies of the prior art, the present invention provides an adaptive robust control method and system for auxiliary equipment systems in thermal power plants under varying operating conditions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an adaptive robust control method for auxiliary equipment systems in thermal power plants under varying operating conditions, comprising the following steps: S1. Construct a dual-network architecture, which includes a prediction model network based on gated recurrent units and a controller network based on deep deterministic policy gradients. S2. Use the prediction model network to perform multi-step prediction of the controlled variables of the auxiliary system to obtain the predicted state sequence; S3. Input the current operating condition characteristics, the controlled variable deviation, and the predicted state sequence into the controller network, and the controller network outputs the original control quantity. S4. Input the original control quantity into the correction layer based on the Lyapunov stability criterion. When the original control quantity causes the system energy function to increase, the correction layer outputs the minimum adjustment amount to correct the original control quantity and generate the final control quantity. S5. Output the final control quantity to the actuator of the auxiliary system; S6. Construct a cloud-based evolution engine. The cloud-based evolution engine uses historical operating data to perform offline optimization of the parameters of the controller network and sends the optimized parameters to the edge controller. The edge controller updates the parameters of the current controller network.

[0007] As a further improvement to the technical solution of the present invention, S2 further includes: the prediction model network adopts a sliding window online fine-tuning mechanism, taking the control quantity and state quantity at the current moment as input and the predicted value of the controlled quantity at the next moment as output; at each preset fine-tuning period, the network parameters of the prediction model network are updated online using the latest running data within the sliding time window.

[0008] As a further improvement to the technical solution of the present invention, during the training process of the controller network in S3, a reward function is constructed, the reward function including: The tracking error penalty term is used to penalize the deviation between the controlled variable and the set value; A control quantity constraint penalty term is applied when the original control quantity output by the controller network exceeds the physical limit of the actuator or the rate of change exceeds a set threshold, resulting in a negative reward. The safety boundary penalty term is used when the predicted value of the controlled variable in the predicted state sequence approaches a preset safety threshold, and the reward function value decreases exponentially.

[0009] As a further improvement to the technical solution of the present invention, the cloud evolution engine in S6 adopts a covariance matrix adaptive evolution strategy for offline optimization; when the edge controller updates the parameters of the current controller network, it adopts an incremental update method and stores the parameters before the update in the parameter backup area.

[0010] As a further improvement to the technical solution of the present invention, the step after S6 includes a parameter rollback step: after the edge controller performs parameter update, the control performance index of the variable operating condition process is continuously monitored for a preset number of times; when the control performance index deteriorates beyond a preset threshold, the parameters before the update are automatically retrieved from the parameter backup area for rollback, and alarm information is generated and uploaded to the cloud.

[0011] As a further improvement to the technical solution of the present invention, the input of the prediction model network in S1 also includes an actuator health index; the actuator health index is obtained by injecting a micro-amplitude excitation signal into the actuator, collecting the deviation between the actual response and the command response of the actuator, using the recursive least squares method to identify the dead zone width and response time constant of the actuator online, and obtaining the health index by weighted fusion based on the identification results.

[0012] As a further improvement to the technical solution of the present invention, the original control quantity output by the controller network in S3 passes through a dead-zone inverse compensator before the input correction layer; the compensation intensity of the dead-zone inverse compensator is negatively correlated with the health index of the actuator, and the compensation intensity increases when the health index decreases.

[0013] As a further improvement to the technical solution of the present invention, a sensor fault-tolerant step is also included: S7. Construct a soft measurement model based on extreme learning machine, using easily measurable variables as input, and outputting estimated values ​​of key controlled variables; S8. When the deviation between the physical sensor signal and the estimated value output by the soft measurement model exceeds a preset threshold, a sensor fault is determined, and the physical sensor signal on which the current operating condition characteristics and the controlled variable deviation input in S3 depend is switched to the estimated value output by the soft measurement model.

[0014] An adaptive robust control system for auxiliary equipment systems in thermal power plants under varying operating conditions, specifically comprising: The data acquisition unit is used to collect current operating condition characteristics, controlled variable deviations, and system status. The prediction module has its input end connected to the data acquisition unit. The prediction module is equipped with a prediction model network based on a gated loop unit, which is used to perform multi-step prediction of the controlled variables of the auxiliary system to obtain the prediction state sequence. The control quantity generation module has a first input terminal connected to the data acquisition unit and a second input terminal connected to the output terminal of the prediction module. The control quantity generation module is equipped with a controller network based on deep deterministic policy gradient, which is used to input the current operating condition characteristics, controlled variable deviation and the predicted state sequence into the controller network to generate the original control quantity. The stability correction module has its input end connected to the output end of the control quantity generation module. The stability correction module has a correction layer based on the Lyapunov stability criterion, which is used to input the original control quantity into the correction layer. When the original control quantity causes the system energy function to increase, the module outputs a minimum adjustment amount to correct the original control quantity and generate the final control quantity. An output module, whose input is connected to the output of the stability correction module, and whose output is used to connect to the actuator of the auxiliary system, for outputting the final control quantity to the actuator; The cloud-based collaborative evolution module has its input end connected to the output end of the control quantity generation module to acquire the operating data of the controller network. Its output end is connected to the parameter update end of the controller network in the control quantity generation module to perform offline optimization of the parameters of the controller network using historical operating data, and send the optimized parameters to the edge controller, which then updates the parameters of the current controller network.

[0015] As a further improvement to the technical solution of the present invention, the cloud-based collaborative evolution module includes: The cloud-based evolution engine is deployed on a cloud platform. Its input end is connected to a historical database to read historical running data. The cloud-based evolution engine adopts a covariance matrix adaptive evolution strategy for offline optimization, and its output end is used to send parameter increments. An edge update unit, deployed on the edge controller, has its input end connected to the output end of the cloud evolution engine and its output end connected to the parameter update end of the controller network in the control quantity generation module. It is used to receive the parameter increment and incrementally update the parameters of the current controller network. The parameter backup area is located inside the edge update unit and is used to store the parameters before the update. The performance monitoring subunit has its input end connected to the output end of the control quantity generation module and the output end of the output module, and is used to monitor control performance indicators. Its output end is connected to the rollback trigger end of the edge update unit, and is used to trigger parameter rollback when the indicators deteriorate.

[0016] The beneficial effects of this invention are: This invention achieves accurate prediction and adaptive control of the dynamic characteristics of auxiliary systems by constructing a dual-network architecture consisting of a prediction model network based on gated recurrent units and a controller network based on deep deterministic policy gradients. By introducing a correction layer based on the Lyapunov stability criterion, the stability of the reinforcement learning controller output is ensured, overcoming the insufficient security of traditional black-box models. Through a cloud-based collaborative evolution mechanism, the controller parameters are optimized offline and incrementally updated using a covariance matrix adaptive evolution strategy, achieving continuous self-optimization of the controller. Furthermore, through actuator health fusion and dead-zone inverse compensation, equipment degradation perception and proactive compensation are achieved. Finally, a sensor fault-tolerant mechanism ensures continuous operation of the system even in the event of sensor failure. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention.

[0018] Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0019] 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.

[0020] This embodiment uses the outlet air temperature control of a coal mill in a thermal power plant as a specific application scenario for illustration. However, those skilled in the art should understand that the technical solution of this invention is also applicable to other auxiliary systems in thermal power plants such as feedwater pumps, induced draft fans, and forced draft fans.

[0021] As attached Figure 1 The adaptive robust control method for auxiliary equipment systems in thermal power plants under varying operating conditions, as shown, includes the following steps: S1. Construct a dual-network architecture, which includes a prediction model network based on gated recurrent units and a controller network based on deep deterministic policy gradients. (1) The prediction model network adopts a gated recurrent unit network, and its network structure is as follows: Input layer: Dimension m, input features include control quantities at the current moment (such as cold air damper opening, hot air damper opening), state quantities (such as coal mill current, primary air pressure, coal mill inlet air temperature), and actuator health indicators; Hidden layer: A two-layer GRU structure is used, with 128 hidden units in each layer and the activation function is the tanh function; Output layer: Dimension 1, outputs the predicted value of the controlled variable at the next moment (coal mill outlet air temperature). Training method: Offline pre-training was performed using historical operating data of thermal power plants (sampling period of 1 second). The loss function adopted was the mean square error function, the optimizer was the Adam algorithm, and the learning rate was set to 0.001.

[0022] (2) The controller network adopts the deep deterministic policy gradient algorithm, and its network structure is as follows: Actor Network: The input is a state vector, and the output is the control increment. The state vector includes the controlled variable deviation, the rate of change of the deviation, and the predicted state sequence output by the prediction model network. The actor network contains three fully connected layers with 256, 128, and 64 neurons, respectively. The activation function is the modified linear unit function, and the activation function of the output layer is the hyperbolic tangent function. The output range is mapped to the range of control variable changes that the actuator can execute. The Critics Network takes a state vector and an action vector as input and outputs a Q-value. It consists of three fully connected layers with 256, 128, and 64 neurons, respectively, and uses a modified linear unit function as the activation function. Training method: An experience replay mechanism is adopted, with a replay pool capacity of 10,000 samples. The target network is updated using a soft update method, with the update coefficient τ set to 0.001 and the discount factor γ set to 0.99.

[0023] S2. Use a predictive model network to perform multi-step prediction of the controlled variables of the auxiliary system to obtain the predicted state sequence. Specifically, a rolling prediction method is adopted, taking the state variable at the current time t as input to predict the controlled variable at time t+1; then taking the predicted value at time t+1 as input to predict the controlled variable at time t+2, and so on, for a total of N prediction steps, where N takes the value of 5 to 10, to form a predicted state sequence. As a preferred implementation, the prediction model network adopts a sliding window online fine-tuning mechanism. Specifically, the prediction model network takes the control quantity and state quantity at the current moment as input and outputs the predicted value of the controlled quantity at the next moment. At each preset fine-tuning period (e.g., 10 minutes), the network parameters of the prediction model network are updated online using the latest running data within the sliding time window (e.g., the previous 30 minutes). The online fine-tuning adopts the mini-batch gradient descent algorithm, with the batch size set to 32 and the learning rate set to 0.0001.

[0024] S3. The controller network outputs the raw control quantity, inputting the current operating condition characteristics, the controlled variable deviation, and the predicted state sequence into the controller network, and the controller network outputs the raw control quantity. The current operating conditions include unit load commands, load change rate, number of coal mills in operation, etc., which are represented by multi-dimensional feature vectors. During the training of the controller network, a reward function is constructed, which consists of the following three components: The first term is the tracking error penalty term R1, which is used to penalize the deviation between the controlled variable and the setpoint. Its expression is: R1 = -k1×e 2 ; Where e is the deviation between the controlled variable and the set value, and k1 is the weighting coefficient, which ranges from 0.1 to 1.0; The second term is the control quantity constraint penalty term R2. When the original control quantity output by the controller network exceeds the physical limit of the actuator or the rate of change exceeds a set threshold, a negative reward is given. Its expression is: R2 = -k2; Wherein, k2 is the penalty value, which ranges from 10 to 100. Specifically, a penalty is triggered when the valve opening command exceeds the range of 0% to 100%, or when the valve opening change rate exceeds 5% per second. The third term is the safety boundary penalty term R3. When the predicted value of the controlled variable in the predicted state sequence approaches the preset safety threshold, the reward function value decreases exponentially, and its expression is: R3=-exp(λ×(TT limit )); Where T is the predicted value of the controlled variable (such as the predicted value of the coal mill outlet air temperature), T limit λ is the safety threshold (e.g., the safety threshold for coal mill outlet air temperature is 85℃), and λ is the sensitivity coefficient, with a value range of 1.0 to 2.0. The expression for the total reward function R is: R = R1 + R2 + R3; The training objective of the controller network is to maximize the cumulative reward.

[0025] S4, Lyapunov stability correction: The original control quantity is input into the correction layer based on the Lyapunov stability criterion. When the original control quantity causes the system energy function to increase, the correction layer outputs the minimum adjustment amount to correct the original control quantity and generate the final control quantity. Specifically, the Lyapunov energy function V(x) of the system is constructed, where x is the system state vector, and the energy function V(x) is defined as a quadratic form of the system state: V(x) = x T Px; Where P is a positive definite matrix, obtained offline by solving linear matrix inequalities.

[0026] The judgment logic of the correction layer is as follows: - Calculate the current state x k The energy function value V(x) under k ); - Calculate the applied original control quantity u RL The predicted state x after pred And the corresponding energy function value V(x) pred ); - If V(x) pred )-V(x k If )≥0, it indicates that the original control quantity leads to an increase in the energy function, and the system may become unstable. In this case, the correction layer outputs the minimum adjustment amount Δu, so that the corrected control quantity u is applied. corr =u RL After +Δu, V(x) is satisfied. pred' )-V(x k ) < 0; - The minimum adjustment amount Δu is obtained by solving the following constrained optimization problem: ; The constraint condition is satisfied: V(f(x) k ,u RL +Δu))-V(x k ) < 0, and physical constraints of the actuator (such as valve opening limit, change rate limit); If the original control quantity already satisfies the stability condition, i.e., V(f(x) k ,u RL ))-V(x k If ) < 0, then the correction layer directly outputs the original control quantity.

[0027] S5. The final control quantity is output to the actuator of the auxiliary system. For the control of the coal mill outlet air temperature, the final control quantity is converted into the opening command of the cold air damper and the hot air damper, and output to the electric actuator through a 4-20 mA current signal.

[0028] S6. Cloud-based collaborative evolution: Build a cloud-based evolution engine. The cloud-based evolution engine uses historical operating data to optimize the parameters of the controller network offline and sends the optimized parameters to the edge controller. The edge controller updates the parameters of the current controller network. As a preferred implementation, the cloud-based evolutionary engine employs a covariance matrix adaptive evolutionary strategy for offline optimization. The core parameters of the covariance matrix adaptive evolutionary strategy are set as follows: (1) The formula for calculating the population size λ is: λ = 4 + floor(3 × ln(n)) Where n is the dimension of the parameters to be optimized. The actual optimization is a subset of key parameters, including the weights of the last layer of the actor network, the weights of the last layer of the critic network, and the weight coefficients of the reward function, totaling about 50 parameters. (2) The initial step size σ is set to 0.5; (3) Termination condition: The number of iterations reaches 1000, or the optimal value changes by less than 10 for 50 consecutive iterations. -6 .

[0029] The objective function is the comprehensive performance index J, and its expression is: J=w1×t settle +w2×overshoot+w3×e ss +w4×N action ; Among them, t settle For variable operating condition response time, overshoot is the overshoot amount, e ss For steady-state error, N actionThe number of actions performed by the actuator is represented by w1, w2, w3, and w4, which are the corresponding weighting coefficients.

[0030] When the edge controller updates the parameters of the current controller network, it adopts an incremental update method and stores the parameters before the update to the parameter backup area. The cloud evolution engine performs optimization during periods when the unit is under low load (such as nighttime load below 30% of rated load) or during shutdown periods to avoid affecting the normal operation of the unit.

[0031] As a preferred implementation, a parameter rollback step is also included after S6: after the edge controller performs parameter updates, the control performance indicators of the variable operating condition process are continuously monitored for a preset number of times; when the control performance indicators deteriorate beyond a preset threshold (such as the response time increasing by more than 20%, or the overshoot increasing by more than 5%), the parameters before the update are automatically retrieved from the parameter backup area for rollback, and alarm information is generated and uploaded to the cloud.

[0032] In a preferred embodiment, in S1, the input to the prediction model network also includes an actuator health index, which is obtained through the following method: A small excitation signal is injected into the actuator. The small excitation signal is a sine wave signal with a frequency of 0.1 Hz and an amplitude of 0.5% of the opening. The deviation between the actual response and the command response of the actuator is collected. The dead zone width and response time constant of the actuator are identified online using the recursive least squares method.

[0033] The formula for calculating recursive least squares is as follows: ; in, Let be the vector of parameters to be identified at time k (including dead zone width d and response time constant τ). φ k Let be the input vector at time k; y k This is the output measurement value at time k; K k Let be the gain vector at time k; P k Let be the covariance matrix at time k; λ RLS The forgetting factor has a value of 0.98. I is the identity matrix; The superscript T denotes the transpose of a vector; The superscript -1 indicates the inverse of the matrix.

[0034] The health index H is obtained by weighted fusion of the identification results, and its calculation formula is as follows: H = α × (1 - d / d) max )+β×(1-τ / τmax ); Where d is the identified dead zone width, and τ is the identified response time constant. max τ is the preset dead zone failure threshold. max The preset response time failure threshold is α and β, which are weighting coefficients that satisfy α+β=1.

[0035] In a preferred embodiment, the original control quantity output by the controller network in S3 passes through a dead-zone inverse compensator before the input correction layer. The compensation intensity of the dead-zone inverse compensator is negatively correlated with the health index of the actuator. When the health index decreases, the compensation intensity increases.

[0036] The compensation function expression for the dead-zone inverse compensator is: u comp =u RL +sign(u RL )×(d / 2)×γ(H); Among them, u RL is the original control quantity, d is the identified dead zone width, sign(·) is the sign function, and γ(H) is the compensation intensity coefficient.

[0037] The mapping relationship between the compensation intensity coefficient γ(H) and the health index H is as follows: When H > 0.9, γ(H) = 0.5; When H < 0.5, γ(H) = 1.0; When 0.5 ≤ H ≤ 0.9, γ(H) is determined by linear interpolation, and the calculation formula is as follows: γ(H)=0.5+(1.0-0.5)×(0.9-H) / (0.9-0.5); In a preferred embodiment, this method further includes a sensor fault-tolerant step: S7. Construct a soft measurement model based on extreme learning machine, with easily measurable variables as input and outputting estimated values ​​of key controlled variables. For the coal mill outlet air temperature, the easily measurable variables include: coal mill current, primary air pressure, cold air damper opening, hot air damper opening, and coal feed rate. The extreme learning machine model structure is: 5 nodes in the input layer, 20 nodes in the hidden layer, the activation function is the sigmoid function, and 1 node in the output layer. The output calculation formula for the Extreme Learning Machine is: y ELM =Σⱼβⱼ×g(wⱼᵀ ×x+bⱼ); Where x is the input vector, wⱼ is the input weight, bⱼ is the hidden layer bias, g(·) is the sigmoid activation function, and g(z) = 1 / (1+e -z ), where βⱼ is the output weight.

[0038] S8. Real-time monitoring of the deviation between the physical sensor signal and the estimated value output by the soft measurement model. The deviation calculation formula is as follows: ; Among them, y sensor y represents the measurement value from the physical sensor. ELM This is the estimated value output by the soft measurement model; When Δ exceeds the preset threshold for three consecutive sampling periods (e.g., temperature deviation exceeds 5℃), the sensor is determined to be faulty, and the current operating condition characteristics input in S3 and the physical sensor signal on which the controlled variable deviation depends are switched to the estimated value output by the soft measurement model.

[0039] As attached Figure 2 As shown, this embodiment provides an adaptive robust control system for auxiliary equipment systems in thermal power plants under varying operating conditions, including a data acquisition unit, a prediction module, a control quantity generation module, a stability correction module, an output module, and a cloud-based collaborative evolution module. The data acquisition unit is used to collect current operating condition characteristics, controlled variable deviations, and system status. This unit communicates with field sensors and the DCS system to obtain real-time operating data.

[0040] The prediction module has its input end connected to the data acquisition unit. It is used to receive the control quantity, state quantity and actuator health index of the current moment output by the data acquisition unit. The prediction module has a prediction model network based on gated loop unit, which is used to perform multi-step prediction of the controlled quantity of the auxiliary machine system to obtain the prediction state sequence. The output end of the prediction module outputs the prediction state sequence to the control quantity generation module. The control quantity generation module has a first input terminal connected to the data acquisition unit to receive the current operating condition characteristics and the controlled variable deviation; its second input terminal is connected to the output terminal of the prediction module to receive the predicted state sequence. The control quantity generation module is equipped with a controller network based on deep deterministic policy gradient, which is used to input the current operating condition characteristics, the controlled variable deviation, and the predicted state sequence into the controller network to generate the original control quantity. The output terminal of the control quantity generation module outputs the original control quantity to the stability correction module.

[0041] The stability correction module has its input connected to the output of the control quantity generation module. It is used to receive the original control quantity. The stability correction module has a correction layer based on the Lyapunov stability criterion. The original control quantity is input into the correction layer. When the original control quantity causes the system energy function to increase, the minimum adjustment amount is output to correct the original control quantity and generate the final control quantity. The output of the stability correction module outputs the final control quantity to the output module.

[0042] The output module has its input end connected to the output end of the stability correction module to receive the final control quantity; its output end is used to connect to the actuator of the auxiliary system to convert the final control quantity into a signal recognizable by the actuator (such as a 4-20 mA current signal or Modbus bus command) and output it to the actuator.

[0043] The cloud-based collaborative evolution module has its input end connected to the output end of the control quantity generation module to obtain the operating data of the controller network; its output end is connected to the parameter update end of the controller network in the control quantity generation module to perform offline optimization of the controller network parameters using historical operating data, and send the optimized parameters to the edge controller, which then updates the parameters of the current controller network.

[0044] As a preferred implementation, the cloud-based collaborative evolution module specifically includes: The cloud-based evolution engine is deployed on a cloud platform. Its input end connects to a historical database to read historical running data. The cloud-based evolution engine uses a covariance matrix adaptive evolution strategy for offline optimization. Its output end is used to send parameter increments. The edge update unit, deployed on the edge controller, has its input connected to the output of the cloud evolution engine to receive parameter increments; its output is connected to the parameter update end of the controller network in the control quantity generation module to perform incremental updates on the parameters of the current controller network. The parameter backup area is located inside the edge update unit and is used to store parameters before the update. The performance monitoring subunit has its input connected to the output of the control quantity generation module and the output of the output module, and is used to monitor control performance indicators; its output is connected to the rollback trigger of the edge update unit, and is used to trigger parameter rollback when the indicators deteriorate.

[0045] The system data flow is as follows: 1. Real-time control flow: Data acquisition unit → Prediction module → Control quantity generation module → Stability correction module → Output module → Actuator.

[0046] 2. Cloud-based collaborative flow: The control quantity generation module uploads runtime data → historical database → cloud evolution engine → edge update unit → parameter update of the controller network in the control quantity generation module.

[0047] 3. Performance monitoring and rollback flow: The performance monitoring subunit monitors and controls performance → when the indicators deteriorate, the edge update unit is triggered to roll back → parameters are retrieved from the parameter backup area for recovery.

[0048] The technical solution of this embodiment was verified in the control of the outlet air temperature of a coal mill in a 300MW thermal power unit. The results show that: (1) Under variable load conditions (50% rated load → 75% rated load → 50% rated load change), the outlet air temperature overshoot is reduced from 8.2 degrees Celsius under traditional proportional-integral-derivative control to 3.5 degrees Celsius, and the adjustment time is shortened from 180 seconds to 95 seconds; (2) The number of times the actuator (cold air damper) operates is reduced by about 35%, effectively extending the equipment life; (3) In the case of simulated sensor failure, the system automatically switches to soft measurement mode, the control function remains normal, and the wind temperature fluctuation range is controlled within ±2 degrees Celsius; (4) The evolution of cloud parameters has enabled the controller performance to continue to improve within three months, and the overall performance index J has decreased by about 28%.

[0049] In summary, the present invention provides an adaptive robust control method and system for auxiliary equipment systems in thermal power plants under varying operating conditions. By constructing a dual-network architecture consisting of a predictive model network based on gated cyclic units and a controller network based on deep deterministic policy gradients, it achieves accurate prediction and adaptive control of the dynamic characteristics of the auxiliary equipment system. By introducing a correction layer based on the Lyapunov stability criterion, it ensures the stability of the reinforcement learning controller output, overcoming the insufficient security of traditional black-box models. Through a cloud-based collaborative evolution mechanism, it utilizes an adaptive evolution strategy based on the covariance matrix to perform offline optimization and incremental updates of the controller parameters, achieving continuous self-optimization of the controller. Furthermore, through actuator health fusion and dead-zone inverse compensation, it achieves equipment degradation perception and proactive compensation. Finally, through a sensor fault-tolerant mechanism, it ensures continuous operation of the system under sensor failure conditions.

[0050] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for adaptive robust control of auxiliary equipment systems in thermal power plants under varying operating conditions, characterized in that: Includes the following steps: S1. Construct a dual-network architecture, which includes a prediction model network based on gated recurrent units and a controller network based on deep deterministic policy gradients. S2. Use the prediction model network to perform multi-step prediction of the controlled variables of the auxiliary system to obtain the predicted state sequence; S3. Input the current operating condition characteristics, the controlled variable deviation, and the predicted state sequence into the controller network, and the controller network outputs the original control quantity. S4. Input the original control quantity into the correction layer based on the Lyapunov stability criterion. When the original control quantity causes the system energy function to increase, the correction layer outputs the minimum adjustment amount to correct the original control quantity and generate the final control quantity. S5. Output the final control quantity to the actuator of the auxiliary system; S6. Construct a cloud-based evolution engine. The cloud-based evolution engine uses historical operating data to perform offline optimization of the parameters of the controller network and sends the optimized parameters to the edge controller. The edge controller updates the parameters of the current controller network.

2. The adaptive robust control method for auxiliary equipment systems in thermal power plants under varying operating conditions according to claim 1, characterized in that: S2 further includes: the prediction model network adopts a sliding window online fine-tuning mechanism, taking the control quantity and state quantity at the current moment as input and the predicted value of the controlled quantity at the next moment as output; at each preset fine-tuning period, the network parameters of the prediction model network are updated online using the latest running data within the sliding time window.

3. The adaptive robust control method for auxiliary equipment systems in thermal power plants under varying operating conditions according to claim 1, characterized in that: During the training process of the controller network in S3, a reward function is constructed, which includes: The tracking error penalty term is used to penalize the deviation between the controlled variable and the set value; A control quantity constraint penalty term is applied when the original control quantity output by the controller network exceeds the physical limit of the actuator or the rate of change exceeds a set threshold, resulting in a negative reward. The safety boundary penalty term is used when the predicted value of the controlled variable in the predicted state sequence approaches a preset safety threshold, and the reward function value decreases exponentially.

4. The adaptive robust control method for auxiliary equipment systems in thermal power plants under varying operating conditions according to claim 1, characterized in that: The cloud-based evolution engine in S6 employs a covariance matrix adaptive evolution strategy for offline optimization; when the edge controller updates the parameters of the current controller network, it uses an incremental update method and stores the parameters before the update in the parameter backup area.

5. The adaptive robust control method for auxiliary equipment systems in thermal power plants under varying operating conditions according to claim 1, characterized in that: The step S6 is followed by a parameter rollback step: after the edge controller performs parameter updates, it continuously monitors the control performance indicators of the variable operating condition process for a preset number of times; when the control performance indicators deteriorate beyond a preset threshold, it automatically retrieves the parameters before the update from the parameter backup area for rollback and generates alarm information to upload to the cloud.

6. The adaptive robust control method for auxiliary equipment systems in thermal power plants under varying operating conditions according to claim 1, characterized in that: The input to the prediction model network in S1 also includes an actuator health index; the actuator health index is obtained by injecting a micro-amplitude excitation signal into the actuator, collecting the deviation between the actual response and the command response of the actuator, using the recursive least squares method to identify the dead zone width and response time constant of the actuator online, and obtaining the health index by weighted fusion based on the identification results.

7. The adaptive robust control method for auxiliary equipment systems in thermal power plants under varying operating conditions according to claim 6, characterized in that: The original control quantity output by the controller network in S3 passes through a dead-zone inverse compensator before the input correction layer; the compensation intensity of the dead-zone inverse compensator is negatively correlated with the health index of the actuator, and the compensation intensity increases when the health index decreases.

8. The adaptive robust control method for auxiliary equipment systems in thermal power plants under varying operating conditions according to claim 1, characterized in that, It also includes sensor fault-tolerant steps: S7. Construct a soft measurement model based on extreme learning machine, using easily measurable variables as input, and outputting estimated values ​​of key controlled variables; S8. When the deviation between the physical sensor signal and the estimated value output by the soft measurement model exceeds a preset threshold, a sensor fault is determined, and the physical sensor signal on which the current operating condition characteristics and the controlled variable deviation input in S3 depend is switched to the estimated value output by the soft measurement model.

9. A variable-condition adaptive robust control method for auxiliary equipment systems in thermal power plants according to any one of claims 1-8, characterized in that, It also includes a variable-condition adaptive robust control system for auxiliary equipment systems in thermal power plants, specifically including: The data acquisition unit is used to collect current operating condition characteristics, controlled variable deviations, and system status. The prediction module has its input end connected to the data acquisition unit. The prediction module is equipped with a prediction model network based on a gated loop unit, which is used to perform multi-step prediction of the controlled variables of the auxiliary system to obtain the prediction state sequence. The control quantity generation module has a first input terminal connected to the data acquisition unit and a second input terminal connected to the output terminal of the prediction module. The control quantity generation module is equipped with a controller network based on deep deterministic policy gradient, which is used to input the current operating condition characteristics, controlled variable deviation and the predicted state sequence into the controller network to generate the original control quantity. The stability correction module has its input end connected to the output end of the control quantity generation module. The stability correction module has a correction layer based on the Lyapunov stability criterion, which is used to input the original control quantity into the correction layer. When the original control quantity causes the system energy function to increase, the module outputs a minimum adjustment amount to correct the original control quantity and generate the final control quantity. An output module, whose input is connected to the output of the stability correction module, and whose output is used to connect to the actuator of the auxiliary system, for outputting the final control quantity to the actuator; The cloud-based collaborative evolution module has its input end connected to the output end of the control quantity generation module to acquire the operating data of the controller network. Its output end is connected to the parameter update end of the controller network in the control quantity generation module to perform offline optimization of the parameters of the controller network using historical operating data, and send the optimized parameters to the edge controller, which then updates the parameters of the current controller network.

10. The adaptive robust control system for variable operating conditions of auxiliary equipment systems in thermal power plants according to claim 9, characterized in that: The cloud-based collaborative evolution module includes: The cloud-based evolution engine is deployed on a cloud platform. Its input end is connected to a historical database to read historical running data. The cloud-based evolution engine adopts a covariance matrix adaptive evolution strategy for offline optimization, and its output end is used to send parameter increments. An edge update unit, deployed on the edge controller, has its input end connected to the output end of the cloud evolution engine and its output end connected to the parameter update end of the controller network in the control quantity generation module. It is used to receive the parameter increment and incrementally update the parameters of the current controller network. The parameter backup area is located inside the edge update unit and is used to store the parameters before the update. The performance monitoring subunit has its input end connected to the output end of the control quantity generation module and the output end of the output module, and is used to monitor control performance indicators. Its output end is connected to the rollback trigger end of the edge update unit, and is used to trigger parameter rollback when the indicators deteriorate.

Images