Dynamic control method and system for coal-fired generating unit based on mechanism data hybrid driving

Abstract

The application discloses a dynamic control method and system for a coal-fired unit based on mechanism data hybrid driving, relates to the technical field of industrial process control, and comprises the following steps: based on the thermodynamic principle and a deep belief network, a mechanism-data hybrid prediction model of a subcritical unit steam pocket pressure, superheated steam pressure and power generation load is established; taking the steady-state operation working condition points under different load sections as reference points, linearization processing is performed on the mechanism-data hybrid prediction model, and a linear state space model group containing system states, control inputs and control outputs is obtained; based on the linear state space model group, dual-mode economic model prediction control including an economic optimization mode and a stable control mode is performed, so that the collaborative control of guaranteeing system stability and optimizing economic benefits at the same time is realized. The application effectively suppresses system state oscillation and improves the dynamic stability and economic operation performance of the unit under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, and in particular to a dynamic control method and system for coal-fired power units based on mechanism data hybrid drive, for load control and economic optimization of coal-fired power units. Background Technology

[0002] With the continuous development of new power systems, the proportion of renewable energy installed capacity is constantly increasing, and its output exhibits significant volatility and intermittency. To maintain the safe and stable operation of the power grid, thermal power units still need to undertake the main peak-shaving and frequency regulation tasks, and frequently participate in automatic generation control, responding quickly to changes in power commands over a wide load range. Due to the increased proportion of renewable energy in new power systems, its output volatility and intermittency place higher demands on the wide-load response, stability, and dynamic economy of thermal power units. The coordinated control system (CCS) is the core of thermal power unit control, and model predictive control (MPC) is widely used. However, traditional MPC does not fully consider economic indicators, making it difficult to achieve overall unit economic optimization. Economic model predictive control (EMPC) can balance dynamic performance and economy, but its effectiveness depends on the predictive model. Existing single and hybrid modeling methods have shortcomings and cannot meet the needs of wide-load operation. At the same time, the closed-loop stability of EMPC is difficult to guarantee after introducing economic targets, and existing stability assurance methods have problems such as conservatism or insufficient flexibility, limiting the improvement of the dynamic economic performance of the unit.

[0003] In summary, existing control methods cannot meet all the diverse needs, and there is an urgent need to propose improved solutions to adapt to the operational requirements of thermal power units under the new power system. Summary of the Invention

[0004] To address the problems existing in the background technology, the purpose of this invention is to provide a dynamic control technology for coal-fired power units based on mechanistic data hybrid drive. This technology aims to solve the problems that existing modeling methods struggle to accurately characterize the time-varying nonlinear characteristics of key system parameters while ensuring physical consistency for the operation requirements of subcritical boiler-turbine units under wide load and frequent adjustment conditions. At the same time, existing economic model predictive control and its stability-assured control methods often exhibit strong conservatism when dealing with complex operating conditions, resulting in insufficient dynamic economy of the control strategy during large-scale load dynamic adjustment.

[0005] To achieve the above technical objectives, this invention discloses a dynamic control method for coal-fired power units based on mechanism-data hybrid drive, comprising the following steps: Based on thermodynamic principles and deep belief networks, a mechanism-data hybrid prediction model for subcritical unit drum pressure, superheated steam pressure, and power generation load is established. Using steady-state operating conditions under different load ranges as benchmarks, the mechanism-data hybrid prediction model is linearized to obtain a set of linear state-space models that include system state, control input, and control output. Based on a set of linear state-space models, a dual-mode economic model predictive control, including an economic optimization mode and a stability control mode, is implemented to achieve coordinated control that simultaneously ensures system stability and optimizes economic benefits.

[0006] Preferably, when constructing the mechanism-data hybrid prediction model, the boiler-turbine unit is divided into a coal feeding system, an evaporation system, a superheating system, and a turbine system. Mechanism models of each subsystem are established based on the thermodynamic principles of mass conservation and energy conservation, respectively. The overall mechanism model of the unit is obtained through model coupling and simplification. While keeping the thermodynamic mechanism structure unchanged, the key nonlinear parameters in the overall mechanism model of the unit are fitted and updated online using a deep belief network to obtain the mechanism-data hybrid prediction model.

[0007] Preferably, when fitting and updating key nonlinear parameters online, the key nonlinear parameters include combustion coefficient, saturated water flow rate, superheated steam flow rate, and turbine power generation coefficient, which are used to characterize the differentiable nonlinear relationship between system variables and key parameters during load changes.

[0008] Preferably, when obtaining the linear state-space model set, based on the mechanism-data hybrid prediction model, the steady-state operating condition points under different load segments of the unit are selected as linearization reference points, and the hybrid prediction model is locally linearized to obtain the linear state-space model set, which is used to describe the dynamic characteristics of the unit near the operating condition.

[0009] Preferably, when obtaining different load segments of the generator set, the load segments are divided into: load segments based on 35% to 45% of the rated maximum generating load of the subcritical generator set; load segments based on 45% to 55% of the rated maximum generating load of the subcritical generator set; load segments based on 55% to 70% of the rated maximum generating load of the subcritical generator set; load segments based on 70% to 85% of the rated maximum generating load of the subcritical generator set; and load segments based on 75% to 100% of the rated maximum generating load of the subcritical generator set.

[0010] Preferably, when performing dual-mode economic model predictive control, dual-mode economic model predictive control is performed through a dual-mode ADMM-EMPC controller. In economic mode, the controller optimizes the economic efficiency of unit operation by minimizing the economic performance function. In stable mode, the auxiliary controller uses dynamically updated feedback gain to identify the system convergence path and introduces the economic performance index into the stability optimization objective function, thereby further optimizing the operating economic efficiency while ensuring system stability.

[0011] Preferably, when performing dual-mode economic model predictive control, the alternating direction multiplier method is used to dynamically update the feedback gain in the stable control mode to identify the convergence path. In this method, the stability constraints are reconstructed using the alternating direction multiplier method. By splitting variables and projecting semidefinite cones, the semidefinite programming problem is decomposed into a series of iteratively solvable subproblems. At the same time, an augmented Lagrange penalty term is introduced to penalize constraint violations and guide the iterative solution to gradually return to the feasible region.

[0012] Based on the same inventive concept, this invention also provides a collaborative control system for dynamic optimization of the economic performance of coal-fired power units, comprising: The basic model building module is used to establish a mechanism-data hybrid prediction model for subcritical unit drum pressure, superheated steam pressure, and power generation load based on thermodynamic principles and deep belief networks. The state-space model construction module is used to linearize the mechanism-data hybrid prediction model based on the steady-state operating conditions under different load segments, and obtain a set of linear state-space models that include system state, control input and control output. The collaborative control module is used to perform dual-mode economic model predictive control based on a linear state-space model group, including economic optimization mode and stability control mode, so as to achieve collaborative control that simultaneously ensures system stability and optimizes economic benefits.

[0013] The present invention discloses the following technical effects: This invention can effectively suppress system state oscillations while ensuring the stable operation of the boiler-turbine unit, achieve rapid and accurate load tracking, significantly reduce coal consumption and valve throttling losses, and improve the dynamic stability and economic operation performance of the unit under complex operating conditions. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is the control principle diagram described in this invention.

[0016] Figure 2 This is the flowchart of the dual-mode economic model predictive control described in this invention.

[0017] Figure 3 This is a prediction effect diagram based on the deep belief network-assisted model described in this invention.

[0018] Figure 4 This is a prediction effect diagram of the mechanism-data hybrid driven model described in this invention.

[0019] Figure 5 This is a diagram illustrating the load control effect of the unit as described in this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] like Figure 1 As shown, this invention provides a dynamic control method for coal-fired power units based on mechanism-data hybrid drive, comprising the following steps: Step S100: Establish a mechanism-data hybrid prediction model. Based on thermodynamic principles and deep belief networks, establish a mechanism-data hybrid prediction model for subcritical unit drum pressure, superheated steam pressure, and power generation load. In one embodiment, step S100 proposes using a deep belief network to fit and update key nonlinear parameters in the unit's mass conservation and energy conservation mechanism model. While maintaining the thermodynamic mechanism structure unchanged, the constructed mechanism-data hybrid model uses the deep belief network to uncover the differentiable nonlinear mapping relationship between system variables and key parameters during load changes. This effectively improves the model's prediction accuracy and nonlinear characteristic representation capability for the unit's dynamic behavior, significantly reduces model structural complexity, overcomes the problem of linearization in traditional hybrid models, and facilitates integration into Engineering Economic Model Predictive Control (EMPC) strategies.

[0022] In one embodiment, step S101 in step S100: establish a thermodynamic mechanism model of the generator set.

[0023] For example, to facilitate modeling and analysis, the complex thermal power generation system is divided into four subsystems: pulverizing system, evaporation system, superheating system and turbine system, and simplified assumptions are made for the research object.

[0024] For example, the simplified assumptions are as follows: 1) Based on the dominant characteristics of the internal thermodynamic process of the boiler, the boiler system is divided into two parts: the evaporation system and the superheating system; 2) It is assumed that the flow rate of the working fluid in the gas phase zone is mainly determined by the pressure difference between the two ends of the section, and the influence of changes in physical parameters such as gas density on the flow rate is ignored; 3) The metal temperature of the boiler heating surface changes slowly with the working fluid temperature. Without considering heat transfer, it can be approximated as a function of the working fluid temperature; 4) Based on actual operation observations, the change in the steam drum water level is extremely small, and it is regarded as a constant in the model; 5) The high-pressure cylinder, intermediate-pressure cylinder, low-pressure cylinder and regenerative system of the steam turbine are regarded as a whole; 6) Based on the analysis of the actual operation of the unit and historical operation data, the fluctuations of feedwater enthalpy and main steam enthalpy are very small, and they can be approximated by constants respectively.

[0025] For example, a pulverizing system consists of a coal bunker, feeder, coal mill, classifier, pulverizer hopper, and conveying pipelines. Its core function is to grind raw coal into pulverized coal that meets the combustion requirements of the boiler. This process often involves delays due to the need to transport and accumulate the pulverized coal, primarily occurring during the coal feeding time of the feeder. Residence time of coal mill And the time it takes for the wind to deliver powder It can be described as... , ; The mass balance equation with the coal mill as the equilibrium point is as follows: In the formula, This refers to the coal feed rate of the coal feeder. The flow rate of pulverized coal fed into the coal mill. The flow rate of pulverized coal output from the coal mill. The flow rate of pulverized coal entering the boiler can be expressed as follows, based on the characteristics of the coal mill: Solving the four formulas together, we get: ,in, For the total delay time, , , , and These are the pulverizing inertia time, pulverizer output coefficient, coal abrasion resistance coefficient, coal moisture content coefficient, and coal powder fineness coefficient, respectively.

[0026] For example, the evaporation system includes an economizer, water-cooled walls, a lower space of the steam drum, downcomers, headers, etc. It is used to convert preheated feedwater into saturated steam, and mass balance equations and energy balance equations are established with the steam drum as the equilibrium point: , , where v, , , , , and These represent volume, density, pressure, flow rate, internal energy, enthalpy, and heat, respectively. , and Represents the metal's mass, specific heat capacity, and temperature. (Subscript) , , , and These represent saturated steam, saturated water, feedwater, steam drum, and evaporation system, respectively. Using equations... Assumptions 3) and 4) The above equation can be simplified and transformed into: , ,in: ; .

[0027] For example, the superheating system includes a low-temperature superheater, a screen-type superheater, and a high-temperature superheater. It is used to heat saturated steam to superheated steam for delivery to the steam turbine. Mass balance equations and energy balance equations are established with the high-temperature superheater as the equilibrium point: , The variables have the same meaning as in the evaporation system, and the subscripts are... , Representing the main steam flow and the superheated system respectively; using the equation Assumptions 3) and 4) The above equation can be simplified and transformed into: , ,in ; ; ; .

[0028] For example, the total heat absorbed by the evaporation system and the superheating system With the concentration of pulverized coal entering the furnace related, Fuel coefficient: .

[0029] For example, the steam turbine system includes high-, medium-, and low-pressure cylinders and a regenerative system, a condenser, a feedwater heater, a main shaft, and a generator, and is the core device for energy conversion in a thermal power unit. Considering the steam turbine as a whole, the steam turbine power... It can be simplified to: in, This is the power generation coefficient.

[0030] In one embodiment, step S102 in step S100 constructs a deep belief network-assisted model.

[0031] For example, to simplify the model, a deep belief network (DPN) is used to establish auxiliary sub-models for fuel coefficient, saturated steam flow rate, turbine coefficient, and superheated steam flow rate, activating the model by selecting the Sigmoid function. The DPN can automatically extract deep nonlinear features from unit operating data and obtain clear input-output function expressions through layer-by-layer pre-training. This mapping is layer-by-layer differentiable, allowing direct Jacobian linearization of the network model at the operating point, thus facilitating the construction of linear prediction models within the EMPC framework. Its function can be expressed as: ,in For activation function, The weights of each hidden layer, For the bias of each hidden layer, This represents the number of hidden layers.

[0032] For example, the input variables of the data-driven model are selected based on the physical coupling relationship of the unit. First, the fuel coefficient has a non-linear relationship with the amount of coal entering the furnace; therefore, the coal feed rate is selected as the input of the Deep Belief Network 1 model. Second, based on assumption 2), the saturated steam flow rate is driven by the pressure difference between the drum pressure and the superheated steam; therefore, the drum pressure and the superheated steam pressure are selected as the input of the Deep Belief Network 2 model. Third, the turbine coefficients depend on the superheated steam flow rate and steam parameters. Given that the turbine valve opening, feedwater flow rate, and coal feed rate are key factors in regulating the steam flow rate, steam parameters, and total heat input, respectively, they are determined as the input of the Deep Belief Network 3 model. Finally, the superheated steam flow rate is determined by the valve opening and upstream pressure; therefore, the turbine valve opening and the superheated steam pressure are selected as the input of the Deep Belief Network 4 model. When the unit operates in steady state, the parameters inside the unit do not change with time, i.e., the differential term is zero. Representing the value under steady-state conditions, the equations above can be used to establish the equations for four auxiliary models: , , , .

[0033] In one embodiment, step S103 in step S100 involves solving the model.

[0034] For example, , addition and union , Further organization yields: , , ,in: ; ; ; ; ; with k1, k2, Substituting these values, we obtain the complete nonlinear state-space model. Let the state vector... Input vector The following complete simplified model can be obtained: In this equation, a deep belief network is introduced to describe the key nonlinear relationships. This not only effectively uncovers the nonlinear characteristics of the system, but also ensures that the equation has good differentiability with respect to the state variables and control inputs, thus facilitating subsequent model linearization and controller design.

[0035] Step S200 linearizes the model constructed in step S100 at steady-state operating points under different load segments and establishes a linear state-space equation set. In one embodiment, step S200 is based on the mechanism-data hybrid prediction model constructed in step S100. The steady-state operating condition points under different load segments of the unit are selected as linearization reference points. The hybrid prediction model is then locally linearized to obtain a linear state-space model containing system state variables, control input variables, and output variables, which is used to describe the dynamic characteristics of the unit near the operating condition.

[0036] In one embodiment, in step S200, the load segments are pre-divided according to the power generation load of the subcritical unit, and the steady-state operating point is selected from the historical operating conditions of the unit in different load segments.

[0037] For example, the specific load segment division includes: setting a load segment based on 35% to 45% of the rated maximum generating load of the subcritical generator set; setting a load segment based on 45% to 55% of the rated maximum generating load of the subcritical generator set; setting a load segment based on 55% to 70% of the rated maximum generating load of the subcritical generator set; setting a load segment based on 70% to 85% of the rated maximum generating load of the subcritical generator set; and setting a load segment based on 75% to 100% of the rated maximum generating load of the subcritical generator set.

[0038] For example, the steady-state operating points for different load segments are selected as shown in Table 1.

[0039] Table 1 Steady-state operating conditions Steady-state operating points were selected at unit loads of 240MW, 300MW, 400MW, 480MW, and 540MW. Linearizing the simplified model generated in step S103 yields the following equation: ,in , , , , , .

[0040] For example, discretization is performed under the zero-order hold (ZOH) assumption, where the control input is within the sampling interval. The internal remains constant. Considering the implementation delay of the digital control system, assume time... The output is determined by the control input applied within the interval. Therefore, the state and output update equations are as follows: .

[0041] For example, without loss of generality, the steady-state operating point is shifted to zero, and... , Substituting the values, the equation can be simplified to: .

[0042] For example, this model is a three-input, three-output process dynamic model. The input variables are coal feed rate, turbine valve opening, and feedwater flow rate; the output variables are drum pressure, superheated steam pressure, and unit load; and the intermediate variable is the amount of coal entering the furnace. The model structure is as follows: .

[0043] Step S300: Design of a dual-mode economic model predictive controller based on ADMM.

[0044] In one embodiment, in step S300, based on the linear state-space model obtained in step S200, a predictive control optimization problem including unit operation constraints and economic performance indicators is constructed, and a dual-mode predictive controller is designed. In the economic mode, unit load tracking is achieved, and coal consumption and turbine valve opening loss are optimized. In the steady mode, Lyapunov stability is applied, and the offline feedback gain is updated online through ADMM. The two economic indicators of coal consumption and turbine valve opening loss are incorporated into the control objective to reduce the conservatism of the control strategy and improve dynamic economy. By solving the optimization problem online, the optimal control quantity for future time moments is output and applied to the boiler-turbine unit.

[0045] In one embodiment, step S301 in step S300 designs a dual-mode economic model predictive control strategy based on the model obtained in step S200.

[0046] For example, such as Figure 2As shown, this invention presents the control flow of the proposed dual-mode ADMM-EMPC controller. The main objective of this method is to reconcile the trade-off between operational economy and closed-loop stability under wide load variations. The proposed dual-mode ADMM-EMPC switches between two types of controllers based on the system state: one for economic optimization and the other for stability recovery. Specifically, in economic mode, the controller optimizes unit operation economy by minimizing the economic performance function; in stable mode, the auxiliary controller uses dynamically updated feedback gain. This allows for the identification of more flexible system convergence paths, and the introduction of economic performance indicators into the stability optimization objective function, thereby further optimizing operational economic efficiency while ensuring system stability.

[0047] For example, the dual-mode ADMM-EMPC switches the controller based on the current state and runtime of the system, achieving dynamic switching between economic mode and stable mode.

[0048] For example, when the system state is in a predefined stable subset Within the specified time frame, and the runtime did not exceed the threshold. At this time, the system operates in Mode 1, i.e., economic mode. In this mode, the controller optimizes to improve the economic efficiency of unit operation. When the system state deviates from the feasible region... Or the running time exceeds the threshold When the system switches to mode 2, i.e., stable mode, the auxiliary controller is activated to restore system stability. To quantitatively characterize the economic operating performance of coal-fired power generating units, this invention constructs an economic performance function. This function, serving as a core indicator for evaluating and optimizing unit operating efficiency, is specifically defined as a weighted sum of load tracking error, coal consumption, and turbine valve throttling losses: In the formula, Set the unit load value. The tracking error representing the unit's output load. The coal consumption of the unit is representative. This represents the throttling loss of the unit's main steam valve; the larger the opening, the smaller the loss. These are the weighting coefficients for each sub-economic indicator.

[0049] For example, when the system switches from mode 1 to mode 2, the controller's primary task is to restore system stability and ensure that the system state converges to an equilibrium point. The proposed dual-mode ADMM-EMPC control strategy is as follows: ,in, Indicates the length of the prediction time domain; and They represent the times respectively. For the future The state deviation vector and control input deviation vector at each moment; and These are the state transition matrix and input matrix of the linearized system, respectively. and These are the output matrix and the direct transfer matrix, respectively. This represents the system state vector measured at the current moment. This represents the steady-state control input vector. The deviation output of the auxiliary controller is represented in the form of linear state feedback. This represents the switching time threshold between mode 1 and mode 2. (Symbol) This represents the set difference operation, i.e. and . The objective function for stability optimization under mode 2 is represented.

[0050] In one embodiment, step S302 in step S300 designs the auxiliary controller in stable mode based on ADMM.

[0051] For example, the core innovation of the stable mode lies in the design of the auxiliary controller based on ADMM. The control law in mode 2 adopts a state feedback form. Unlike traditional dual-mode economic model predictive control methods that rely on fixed feedback gain. Unlike auxiliary controllers, fixed-gain designs often result in highly conservative control strategies, causing the system to recover stability along a pre-set and rigid convergence trajectory. This lack of adaptability to the current operating state, coupled with neglecting economic performance factors in stable mode, makes it difficult to fully exploit dynamic economic optimization potential while ensuring stability. Therefore, this invention introduces ADMM to adjust the feedback gain. Dynamic updates are performed, and economic performance indicators are explicitly incorporated into the stability optimization objective function. This allows the auxiliary controller to identify more flexible convergence paths based on the current system state. While strictly satisfying Lyapunov stability constraints and ensuring system stability under worst-case conditions, it also effectively reduces economic costs, thus balancing system stability and dynamic economic performance.

[0052] For example, stability constraints are constructed using the form of linear matrix inequalities (LMI). Traditional feedback gain design methods based on LMI can rigorously guarantee the stability of the closed-loop system through convex optimization techniques. Therefore, variable transformation is introduced. and ,in Represents the state feedback gain matrix. Let be a scalar performance index. Through the above variable substitution, the original stability constraint can be transformed into a linear matrix inequality form: ,in, This represents the system state vector measured at the current sampling time. LMI1 is used to ensure that the current system state is within the range specified by LMI1. Within the defined invariant ellipsoid; LMI2 imposes Lyapunov stability constraints to ensure that the system satisfies This ensures the asymptotic stability of the closed-loop system; LMI3 is used to constrain the amplitude of the control input, where .

[0053] For example, however, the above methods typically involve computationally intensive semidefinite programming (SDP) problems and rely on numerical solvers to obtain a fixed feedback gain. This approach is insufficient to meet the requirements for online dynamic updates. To achieve dynamic updates of the feedback gain, the Alternating Direction Multiplier Method (ADMM) is introduced to reconstruct the stability constraints. Through variable splitting and positive semidefinite cone projection, the originally complex positive semidefinite programming problem is decomposed into a series of iteratively solvable subproblems. Simultaneously, an augmented Lagrange penalty term is introduced to penalize constraint violations, guiding the iterative solution to gradually return to the feasible region. Based on this, an optimization objective function in the stable mode is constructed. Its definition is as follows: ,in It is the upper bound of the Lyapunov function, reducing This can tighten the stability region as much as possible and improve control accuracy. It is a stability block matrix constructed from LMI2. These are introduced auxiliary variables that are forced to satisfy certain conditions in each iteration. This makes LMI2 approach To ensure stable contraction of the system, (Lagrange multipliers for dual variables) measure the degree of inconsistency between variables; This is the consistency penalty coefficient, which adjusts the convergence speed and numerical stability. , , The penalty weights for each item. , Here is the constraint matrix for LMI1. for The eigenvalue, which primarily determines the degree of punishment for violating the positive semi-qualitative property, drives... It converges to the feasible region that satisfies LMI1. , Here is the constraint matrix for LMI3. for The eigenvalues ​​are used to constrain the rate of change of the input variables. The economic costs incurred in the current state, of which By adjusting the weights This can reduce coal consumption and valve opening loss while ensuring stability, thereby balancing economy and dynamic performance.

[0054] For example, such as Figure 3 The image shows the prediction performance of the auxiliary model based on a deep belief network. The data sampling period is 15 seconds, and the data period is from July 10, 2024 to July 27, 2024. For the above data, the system was first stopped and outliers were removed. Then, based on the output features, a sliding discriminant method was used to select 1000 sets of steady-state data that met the conditions for training the auxiliary model. The root mean square error (RMSE) of training was used as the evaluation criterion, and the model identification accuracy of the extreme learning machine and the traditional polynomial form was compared. The results are as follows. Figure 3 As shown in the table, the accuracy of the model assisted by the deep belief network is significantly higher, and it can deeply learn the nonlinear characteristics of the unit. The specific recognition results are shown in Table 2.

[0055] Table 2 Comparison of recognition performance of auxiliary models For example, such as Figure 4 The diagram shows the comparison results of the mechanism-data hybrid model. From a total of 100,000 data sets, 20,000 sets with large load variations were selected for training, and another 3,000 sets were selected as the test set to verify the model's performance. First, considering the model's highly nonlinear and multi-extreme distribution characteristics, a gradient-free swarm intelligence algorithm—Differential Evolution (DE)—was used for global parameter optimization. This algorithm, through population initialization, mutation, crossover, and selection operations, can effectively search within the domain to obtain a globally optimal or near-globally optimal solution. Then, using the optimal solution obtained from differential evolution as the initial point, a nonlinear least squares method was introduced to further refine the parameter fitting, improving the fitting accuracy and approximating the final optimal solution while maintaining the physical feasibility constraints of the parameters. The population size was 50, and the crossover probability was 0.9. A total of 100 iterations were performed. The identified dynamic parameters are: , , , , , , The specific prediction errors are shown in the table below. It can be seen that this hybrid prediction model has good following characteristics under continuous large-scale load increases and decreases in the unit, and can accurately reflect the dynamic characteristics of the unit. The maximum average relative error is 0.68%, indicating that the model has high accuracy.

[0056] Table 3 Error of Model Output Parameters on Test Set For example, such as Figure 5 The figure shows the effect of wide-range variable load control. To verify the effectiveness of the proposed controller, simulation experiments were conducted under two wide-range variable load conditions. The main evaluation focused on the tracking performance of the dual-mode ADMM-EMPC under different system operating conditions and its economic efficiency during dynamic operation, comparing it with traditional MPC and dual-mode EMPC. Considering the actual control requirements and economic optimization objectives of the power plant operation, the weighting coefficients of the economic performance function (Equation 30) were set as follows: , , The control input constraints are , , The consistency penalty coefficient of the auxiliary controller is set to , , , The weight matrices q and r of a traditional MPC controller are selected as follows: , .

[0057] For example, in load rise control, the initial upper bound of the stable region is calculated as follows: The threshold for economic zones is set as follows: The dynamic responses of the system output and control input are as follows: Figure 5 a and Figure 5 As shown in b, from the output response perspective, the traditional MPC has the fastest rise time, but is accompanied by significant initial oscillations, especially noticeable in pressure responses. In contrast, the dual-mode ADMM-EMPC strategy exhibits superior stability and smoothness; although its rise time is slightly longer, the dynamically updated feedback gain K effectively suppresses system oscillations. Regarding control input response, all three controllers can quickly adjust the coal feed rate when receiving a load increase command. Water supply flow rate and turbine valve opening However, there are significant differences in control actions: both the dual-mode EMPC and the dual-mode ADMM-EMPC make more aggressive adjustments in the initial stage to maximize economic benefits during the transient phase. More importantly, from an economic perspective, the dual-mode ADMM-EMPC achieves the lowest coal consumption and the smallest valve throttling loss in steady state, verifying that the ADMM-based auxiliary controller can more fully tap the economic optimization potential of the unit while ensuring system stability.

[0058] For example, in load reduction control, the unit load setpoint decreases from 504MW to 407MW. The responses of the system output and control input are as follows: Figure 5 c and Figure 5 As shown in d. From the perspective of output performance, as... Figure 5 As shown in Figure c, the dual-mode ADMM-EMPC strategy maintains excellent tracking capability during load reduction. Although the traditional MPC has the fastest load reduction rate, it is accompanied by a more significant pressure drop, while the dual-mode ADMM-EMPC can achieve a smooth transition of unit load and steam pressure. Regarding control input response, such as... Figure 5 As shown in Figure d, all three controllers can drive the coal feed rate, water flow rate, and turbine valve opening to new steady-state values. Consistent with load increase control, the dual-mode ADMM-EMPC still achieves the lowest coal consumption and minimum valve throttling loss in steady state, further verifying that this control strategy can effectively optimize economic performance while ensuring system stability.

[0059] For example, in order to quantitatively analyze the overall economic performance of each controller, the system's total economic cost function is... Defined as the cumulative sum of costs at each stage of the simulation process: ,in Let the system simulation time be denoted as and the economic indicators in load increase control be denoted as . The economic indicators in load reduction control are As shown in Table 4.

[0060] Table 4 Comparison of Overall Economic Performance of Load Control As can be clearly seen from Table 4, compared with traditional MPC and ordinary EMPC, the ADMM-EMPC controller maintains system stability while exhibiting better overall economic performance of the generator set during dynamic operation due to the additional introduction of economic indicator functions in the auxiliary controller.

[0061] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0062] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0063] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A dynamic control method for coal-fired power units based on mechanism-data hybrid driving, characterized in that, Includes the following steps: Based on thermodynamic principles and deep belief networks, a mechanism-data hybrid prediction model for subcritical unit drum pressure, superheated steam pressure, and power generation load is established. Using steady-state operating conditions under different load ranges as benchmarks, the mechanism-data hybrid prediction model is linearized to obtain a set of linear state-space models that include system state, control input, and control output. Based on the aforementioned linear state-space model set, a dual-mode economic model predictive control, including an economic optimization mode and a stability control mode, is implemented to achieve coordinated control that simultaneously ensures system stability and optimizes economic benefits.

2. The dynamic control method for coal-fired power units based on mechanism data hybrid drive according to claim 1, characterized in that: When constructing the mechanism-data hybrid prediction model, the boiler-turbine unit is divided into a coal feeding system, an evaporation system, a superheating system, and a turbine system. Mechanism models of each subsystem are established based on the thermodynamic principles of mass conservation and energy conservation, respectively. The overall mechanism model of the unit is obtained through model coupling and simplification. While keeping the thermodynamic mechanism structure unchanged, the key nonlinear parameters in the overall mechanism model of the unit are fitted and updated online using a deep belief network to obtain the mechanism-data hybrid prediction model.

3. The dynamic control method for coal-fired power units based on mechanism data hybrid drive according to claim 2, characterized in that: When fitting and updating key nonlinear parameters online, the key nonlinear parameters include combustion coefficient, saturated water flow rate, superheated steam flow rate, and turbine power generation coefficient, which are used to characterize the differentiable nonlinear relationship between system variables and the key parameters during load changes.

4. The dynamic control method for coal-fired power units based on mechanism data hybrid drive according to claim 1, characterized in that: When obtaining the linear state-space model set, based on the mechanism-data hybrid prediction model, the steady-state operating condition points under different load segments of the unit are selected as linearization reference points, and the hybrid prediction model is locally linearized to obtain the linear state-space model set, which is used to describe the dynamic characteristics of the unit near the operating condition.

5. The dynamic control method for coal-fired power units based on mechanism data hybrid drive according to claim 4, characterized in that: When obtaining different load segments of the generating unit, the load segments are divided into: load segments based on 35% to 45% of the rated maximum generating load of the subcritical generating unit; load segments based on 45% to 55% of the rated maximum generating load of the subcritical generating unit; load segments based on 55% to 70% of the rated maximum generating load of the subcritical generating unit; load segments based on 70% to 85% of the rated maximum generating load of the subcritical generating unit; and load segments based on 75% to 100% of the rated maximum generating load of the subcritical generating unit.

6. The dynamic control method for coal-fired power units based on mechanism-data hybrid drive according to claim 1, characterized in that: When performing dual-mode economic model predictive control, the dual-mode ADMM-EMPC controller is used to perform dual-mode economic model predictive control. In economic mode, the controller optimizes the economic efficiency of unit operation by minimizing the economic performance function. In stable mode, the auxiliary controller adopts dynamically updated feedback gain to identify the system convergence path and introduces the economic performance index into the stability optimization objective function, thereby further optimizing the operating economic efficiency while ensuring system stability.

7. The dynamic control method for coal-fired power units based on mechanism data hybrid drive according to claim 1, characterized in that: When performing predictive control using a dual-mode economic model, the feedback gain is dynamically updated using the alternating direction multiplier method under the stable control mode to identify the convergence path. Specifically, the stability constraints are reconstructed using the alternating direction multiplier method. By splitting variables and projecting semidefinite cones, the semidefinite programming problem is decomposed into a series of iteratively solvable subproblems. At the same time, an augmented Lagrange penalty term is introduced to penalize constraint violations, guiding the iterative solution to gradually return to the feasible region.

8. A collaborative control system for dynamic optimization of the economic performance of a coal-fired power unit, used to execute the dynamic control method for a coal-fired power unit based on mechanism data hybrid drive as described in claim 1, characterized in that, include: The basic model building module is used to establish a mechanism-data hybrid prediction model for subcritical unit drum pressure, superheated steam pressure, and power generation load based on thermodynamic principles and deep belief networks. The state-space model construction module is used to linearize the mechanism-data hybrid prediction model based on the steady-state operating conditions under different load segments, so as to obtain a set of linear state-space models that include system state, control input and control output. The collaborative control module is used to perform dual-mode economic model predictive control, including economic optimization mode and stable control mode, based on the linear state-space model group, so as to achieve collaborative control that simultaneously ensures system stability and optimizes economic benefits.

Images