Thermal power plant optimization control system based on digital twinning

Abstract

The invention relates to the technical field of electric power energy engineering and industrial intelligent control, in particular to a thermal power plant optimization control system based on digital twinning, and the system comprises a data mapping module which collects real-time data, and reconstructs the real-time data into a virtual entropy-enhanced potential energy field reflecting a thermal state through a graph neural network; the potential well evolution calculation module is used for setting an optimal well of energy consumption and emission constraints and calculating a geodesic line gradient path evolved to the optimal well; the reverse damping control module is used for superposing virtual thermal inertia damping along a path and resolving a control instruction; the closed-loop execution module issues an instruction to drive the unit state to approach the global optimal trap; according to the method, prediction that a pure data driving model violates physical common sense is avoided, and accurate inference and early warning of the temperature of areas without sensors such as the center of the hearth are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power energy engineering and industrial intelligent control technology, in particular to a thermal power plant optimization control system based on digital twinning. BACKGROUND

[0002] In the current thermal power plant operation control environment, the distributed control system needs to collect a large amount of temperature, pressure and flow rate and other measurement point data in real time. These data usually exhibit discretization and spatial sparsification characteristics, and the operation of physical units is accompanied by huge thermal inertia and complex thermodynamic coupling processes. To achieve precise control of the unit, the existing scheme generally adopts high-dimensional computational fluid dynamics simulation or full-field solution of Navier-Stokes equations, trying to analyze the full-field state and obtain the optimal operating parameters through accurate physical modeling. Although this scheme has theoretical accuracy in offline analysis, due to its extremely high computational complexity and cubic growth with dimension, the calculation period from state perception to instruction generation is often several hours, which cannot meet the real-time response requirement of industrial process control in milliseconds, thus forming a calculation deadlock between global optimization objectives and real-time response capabilities. In addition, a pure data-driven model is prone to generate non-physical predictions that violate the law of conservation of energy, and theoretical optimal instructions lacking inertia constraints will cause high-frequency jitter or response overshoot of the actuator, making it difficult to adapt to the actual response limit of the physical device. Therefore, how to break down the barrier between discrete data and continuous thermal field, reduce the computational complexity of high-dimensional nonlinear systems while ensuring the constraints of physical laws and the feasibility of device execution, and achieve real-time global optimization, has become a technical problem to be solved. SUMMARY

[0003] To solve the above technical problems, the present application provides a thermal power plant optimization control system based on digital twinning. Specifically, the technical scheme of the present application comprises: A data mapping module for collecting real-time data of the distributed control system of the thermal power plant, reconstructing discrete measurement point data into a virtual entropy-increasing potential field reflecting the thermodynamic state of the unit through a graph neural network; A potential well evolution calculation module for setting a global optimal potential well with energy consumption and emission as constraint conditions in the virtual entropy-increasing potential field, and calculating the geodesic gradient path of the current state point evolving to the global optimal potential well; A reverse damping control module for superimposing a virtual friction damping term simulating the thermal inertia of the device based on the geodesic gradient path, and solving the corresponding actuator control instruction; A closed-loop execution module for issuing the actuator control instruction to the actuator to drive the physical unit state to approach the global optimal potential well along the geodesic gradient path.

[0004] Preferably, the data mapping module comprises: a spatial discretization unit for discretizing the boiler space into a grid, using the values of sparse sensors as anchor points; a field reconstruction unit for reconstructing the discrete measurement points into a continuous virtual potential energy surface by graph convolution network or Kriging interpolation to predict the virtual potential energy of unmeasured regions; the virtual potential energy surface is constructed using a hybrid architecture of physical information neural network and sparse field reconstruction, wherein the physical constraints are embedded into the loss function of the physical information neural network.

[0005] Preferably, the spatial transformation performed by the data mapping module follows the following mapping logic: map the control variables of coal feed rate and damper opening to the potential energy injected into the field; map the operating states of the furnace temperature distribution and pressure distribution to the field strength distribution; map the optimization goal of minimum energy consumption and minimum emissions to the minimum action principle of the field, so that the system tends to evolve in the direction of minimum entropy increase or optimal energy dissipation path.

[0006] Preferably, the potential well evolution calculation module comprises: an operator extraction unit for extracting the evolution equation on the low-dimensional manifold using manifold learning, introducing the reduced order approximation operator of Hamilton-Jacobi-Bellman equation, and focusing on the gradient of energy flow; a path optimization unit for defining the current state as a point on the virtual entropy increase potential energy field surface, calculating the gradient descent path that makes the point enter the global optimal potential well at the fastest speed, and converting the solution of high-dimensional nonlinear differential equations into gradient optimization on the potential energy surface.

[0007] Preferably, the reverse damping control module comprises: a damping calculation unit for introducing a reverse entropy damping term, adding a virtual friction force proportional to the rate of change in the mathematical equation, to simulate the thermal inertia of the physical world and prevent regulation oscillation; an instruction decoding unit for inversely decoding the abstract energy flow direction of the calculated gradient path into specific actuator instructions, including the rotational speed of the coal mill and the opening of the secondary air damper.

[0008] Preferably, the system runs on an edge computing gateway configured with a tensor processing unit; when the data mapping module stores the virtual entropy increase potential energy field, a sparse matrix storage format is used to optimize memory consumption; the data mapping module processes Modbus or OPC protocol data streams from the distributed control system, with a sampling frequency not less than 1Hz.

[0009] Preferably, the data mapping module is further used for: Based on the continuity reconstruction of the entropy field, the numerical value of the area without installed sensors is inferred, a virtual sensor emerges, and early warning of local high temperature corrosion is formed; The area without installed sensors includes the furnace center flame temperature area.

[0010] Preferably, the closed-loop execution module further comprises a safety logic control unit: The safety logic control unit is used to monitor the unit load state, and when the unit is in the 30% to 100% load section, the optimization algorithm is enabled; When the non-steady state extreme working condition of the unit in the start-stop machine stage is monitored, the safety logic control unit automatically degrades the system to a safety logic control.

[0011] Compared with the prior art, the present application has the following beneficial effects: 1. The potential well evolution calculation module uses manifold learning to extract a low-dimensional evolution equation and introduces a reduced-order approximation operator of the Hamilton-Jacobi-Bellman equation, which converts the complex solution of traditional high-dimensional nonlinear differential equations into gradient optimization on the potential energy surface. This mechanism significantly reduces the computational complexity, compresses the simulation level calculation which originally takes several hours to milliseconds, and effectively solves the calculation deadlock between the global optimization goal and the real-time response capability; 2. The present application constructs a hybrid architecture that combines physical information neural networks and sparse field reconstruction, and embeds physical constraints such as energy conservation into the neural network loss function. This method not only reconstructs the continuous virtual potential energy surface from discrete measurement data using graph neural networks, but also ensures that the reconstructed field strictly follows the laws of thermodynamics, avoiding predictions that violate physical common sense by purely data-driven models, and achieving accurate inference and early warning of the temperature in the area without installed sensors such as the furnace center; 3. The present application introduces a reverse damping control module, which superimposes a virtual friction damping term simulating the thermal inertia of the device on the geodesic gradient path. By adding a reverse entropy damping proportional to the state change rate, the system can filter out high-frequency jitter instructions that are theoretically optimal but physically unexecutable at the mathematical level, effectively preventing regulation oscillation caused by frequent opening and closing of dampers, and ensuring the smoothness and executability of the control instructions of the actuator within the response limit of the physical device; 4. The present application uses sparse matrix storage format to optimize the three-dimensional data storage of the virtual field, greatly compresses the memory consumption to adapt to the edge computing gateway resources, and uses tensor processing units to accelerate the operation. At the same time, the system is embedded with a dual-mode safety logic, which enables the optimization algorithm only when the unit is in the steady load section, and automatically degrades to safety control in extreme working conditions such as start-stop machine, ensuring efficient operation on the edge side while building a safety bottom line for artificial intelligence intervention in industrial control. BRIEF DESCRIPTION OF DRAWINGS

[0012] The application will be further explained in connection with the accompanying drawings and embodiments: Figure 1 is a structural diagram of the system of the application. DETAILED DESCRIPTION

[0013] To make the objectives, technical solutions and advantages of the application clearer, the application will be further described in detail below with specific embodiments. EMBODIMENT

[0014] Please refer to Figure 1 A thermal power plant optimization control system based on digital twinning, comprising: A data mapping module, configured to collect real-time data of a distributed control system of the thermal power plant, and reconstruct discrete measurement point data into a virtual entropy-increasing potential field reflecting a thermodynamic state of a unit through a graph neural network; A potential well evolution calculation module, configured to set a global optimal potential well with energy consumption and emission as constraint conditions in the virtual entropy-increasing potential field, and calculate a geodesic gradient path of evolution of a current state point to the global optimal potential well; A reverse damping control module, configured to superimpose a virtual friction damping term simulating thermal inertia of a device based on the geodesic gradient path, and solve corresponding actuator control instructions; A loop execution module, configured to issue the actuator control instructions to actuators, and drive a physical unit state to approach the global optimal potential well along the geodesic gradient path.

[0015] This embodiment details the overall architecture and operation mechanism of the thermal power plant optimization control system based on digital twinning, which aims to solve the calculation deadlock contradiction between global optimization objectives and real-time response capability in traditional thermal power plant control. The system starts the data mapping module to collect real-time data of a distributed control system (DCS) of the thermal power plant; the data mapping module defines a state vector , wherein represents temperature, pressure and flow measurement point data, respectively; Since the DCS data usually presents discrete characteristics, this module uses a graph neural network technology to construct an adjacency matrix characterizing the spatial topology between measurement points, and the specific rules for constructing the adjacency matrix are as follows: calculate the Euclidean distance of measurement points in three-dimensional space , if is less than a default thermal mechanical connection threshold , the value of the thermal length of the physical device features where the sensor is located, for example, 1.5 times the length of the water wall pipe, then it is determined that node and node exist a connection edge, and , otherwise ; define the adjacency matrix after adding self-loop , whose corresponding degree matrix , by formula:

[0016] perform feature aggregation to define the initial feature matrix Each row of the matrix corresponds to a sensor node, and its feature vector contains normalized measurement values, temperature, pressure, and three-dimensional spatial coordinates of the node ; in the formula is the re-normalized Laplacian operator; in the formula is the re-normalized adjacency matrix after adding self-loop, is the degree matrix of , is the degree matrix of is the learnable weight matrix of the layer, is a nonlinear activation function that reconstructs discrete measurement point data into a continuous virtual entropy-increasing potential field scalar function ; the function value represents the current consumption of the system's scattered energy field level; The potential well evolution calculation module performs operations in the virtual entropy-increasing potential field. This module sets the global most dominant well , which satisfies and the Hessian matrix is positive definite; then calculate the geodesic line gradient path of the current state point evolving to the global most dominant well, which follows the geodesic line equation on the Riemannian manifold , in the formula is the second Christoffel symbol, and the metric tensor constructed from the potential virtual energy field is derived, where the metric tensor is defined as the conformal deformation of the Euclidean metric , so that the regional spatial distance of the potential curve is stretched, so that the natural endpoint of the geodesic line in the high potential area represents the connection coefficient of the curved space. Assuming that the connection tensor of the manifold space is a unit matrix and ignoring the second-order curvature effect, the second-order dynamic problem is reduced and approximated to the negative gradient direction along the potential surface , where is the gradient evolution step size coefficient; On this basis, the reverse damping control module based on the geodesic line gradient path superimposes the virtual friction damping term of the simulation device thermal inertia to solve the corresponding actuator control instruction ; this step introduces the reverse mechanism to calculate the ideal driving force of the state space, which is:

[0017] where, is the default virtual velocity-force conversion gain matrix, whose physical unit is set as or equivalent generalized force / generalized velocity unit, which is equivalent to the damping coefficient in physical dimension, but acts as a proportional gain in control logic for converting gradient velocity into driving force dimension; or directly using force balance equation, where is the guard force generated by the potential field, which is mapped into control increment by using matrix transformation of matrix at this time, and the calculation formula is where is the generalized inverse matrix, where is the damping coefficient matrix determined according to the thermal time constant of the device, which is used to prevent the instruction change rate from exceeding the physical device response limit; The closed-loop execution module issues the actuator control instruction to the coal mill, damper baffle and other actuators to drive the physical unit state to approach the global optimal advantage well along the geodesic gradient path.

[0018] Embodiment 2: The data mapping module comprises: a spatial discretization unit for discretizing the boiler space into a grid, using the values of the sparse sensors as anchor points; a field reconstruction unit for predicting the virtual potential energy of the unmeasured area by a graph convolution network or Kriging interpolation, and reconstructing the discrete measurement points into a continuous virtual potential energy surface; the virtual potential energy surface is constructed by using a hybrid architecture of a physical information neural network and a sparse field reconstruction, wherein the physical constraint is embedded into the loss function of the physical information neural network; This embodiment specifically implements the data mapping module in Embodiment 1, and focuses on solving the engineering problem of reconstructing a continuous field from sparse data; The spatial discretization unit discretizes the boiler physical space into a three-dimensional grid, and defines a set of grid nodes ; using the specific values of the sparse installed sensors, such as the furnace wall temperature measurement points, as anchor points, i.e. for the nodes where the constraint is forced; The field reconstruction unit predicts the virtual potential energy of the unmeasured area by a graph convolution network; in this embodiment, the virtual potential energy surface is constructed by using a hybrid architecture of a physical information neural network PINN and a sparse field reconstruction; the core lies in that the training loss function is defined as:

[0019] where is the weight coefficient of the physical constraint term, the data error term , is the number of sensors; the physical constraint term is defined as the residual norm of the energy conservation partial differential equation, and specifically is:

[0020] in For density, For specific heat capacity, Three-dimensional fluid velocity vector field The velocity field The average flow field derived from the CFD cold simulation database or real-time air volume measurement points is used as a known input during the thermal field reconstruction process. Thermal conductivity, For the combustion heat source term; among which, the combustion heat source term Coal feed There is a mapping relationship between them. In the formula For combustion efficiency, The lower heating value of coal. This is a spatial distribution function preset based on the burner's geometric position, specifically using the coordinates of the burner nozzle center. The three-dimensional Gaussian distribution function of the expected value:

[0021] in, Combustion efficiency, dimensionless. The lower heating value of coal, in units:

[0022] : This refers to the mass flow rate of coal feed, in units of: ; This is the normalized combustion space probability density function, whose physical unit is the reciprocal of volume. To meet the full score requirement This ensures the heat source item The dimension of power density This is consistent with the dimensions of the terms on the left-hand side of the energy conservation equation; To simulate the heat release characteristics that decay from the center of a flame to the surrounding areas; By minimizing this loss function, the network weights are updated, ensuring that the reconstructed virtual field still strictly follows the laws of thermodynamics in areas where sensors are not installed. This effectively avoids non-physical predictions that might be generated by a purely data-driven model, which violate the law of conservation of energy.

[0023] Example 3: The spatial transformations performed by the data mapping module follow the following mapping logic: The control variables of coal supply and damper opening are mapped to the potential energy injected into the field; the operating state of the furnace temperature distribution and pressure distribution is mapped to the field intensity distribution; the optimization goal of minimum energy consumption and minimum emission is mapped to the minimum action principle of the field, so that the system tends to evolve in the direction of minimum entropy action or optimal energy dissipation path.

[0024] This embodiment details the mapping logic followed by the spatial transformation performed by the data mapping module, which is the key interface for converting physical engineering problems into mathematical extremum problems; The system performs input mapping to map the control variables of coal supply and damper opening to the potential energy increment injected into the field , and the mapping function is set as:

[0025] wherein, is the system sampling period or discrete time step; : is the conversion coefficient of coal calorific value, with a unit of , used to convert coal supply into energy input; is the conversion coefficient of enthalpy carried by wind, with a unit of or , used to unify the dimensions of control variables; wherein is the spatial distribution weight coefficient, representing the influence of control variable change on local potential energy gradient; the above-mentioned spatial distribution weight coefficients and are obtained as follows: based on the CFD fluid simulation data of the boiler in cold and hot states, the sensitivity response field of coal supply and damper opening change on the full-field vector and temperature scalar is given, and the sensitivity response field is smoothed by a three-dimensional Gaussian kernel function, and the specific smoothing calculation formula is:

[0026] wherein the kernel function , is a smoothing factor representing the thermal influence range, taking a value of 0.05-0.1 times the characteristic width of the furnace, and the normalized processing is performed on the smoothed scalar field value, so that , ensuring the energy conservation of control variables, thereby constructing a three-dimensional weight matrix covering the full-furnace grid; The system first performs dimensionless processing on physical quantities: defining the normalized temperature and the normalized pressure , wherein is the extreme value under the design condition; the system performs state mapping to map the furnace temperature distribution and the pressure distribution The operating state of the system is mapped to the field intensity distribution , define where the modulus of the field intensity represents the instability or the degree of change of the thermodynamic state; The system performs a mapping of the optimization objectives of energy consumption and emissions , and defines a weighted sum as a Lagrangian quantity , where is a preset non-negative weight coefficient, wherein the energy consumption objective is defined as , and the emission objective is defined as , where is calculated by a thermodynamic generation mechanism function of the local temperature ; the optimization objectives of the minimum energy consumption and the minimum emissions are mapped to the generalized least action principle of the field, that is, to find a state evolution path , so that the action functional takes the minimum value ; This means that in the virtual field, the system always tends to find a path that minimizes the integral action along the path, guiding the system to evolve in the direction of the minimum consumption of the dissipated energy field or the optimal energy dissipation path.

[0027] Embodiment 4: The potential well evolution calculation module comprises: An operator extraction unit is configured to extract an evolution equation on a low-dimensional manifold using manifold learning, and introduce a reduced-order approximation operator of the Hamilton-Jacobi-Bellman equation, which is specifically constructed as a multi-layer perceptron network , which is used to fit the gradient field of the value function ; the multi-layer perceptron network is configured to have an input dimension equal to the manifold dimension , and an output dimension of the gradient vector , and internally contains 3 hidden layers, with the number of nodes in each layer being 128, 64, and 32, respectively, and using a Swish activation function to improve the convergence speed; in the offline stage, the network parameters are pre-trained by minimizing the Bellman residual loss function , so that the approximate gradient is directly output by the network forward propagation in the online calculation, avoiding real-time numerical solution of the partial differential equation and focusing on the gradient of energy flow; A path optimization unit is configured to define the current state as a point on the virtual entropy-increasing potential field surface, and calculate a gradient descent path that makes the point enter the global optimal potential well at the fastest speed. The gradient descent path calculated here is essentially based on the value function extracted in this embodiment, and according to the Bellman optimality principle, the gradient descent path is calculated in the negative gradient direction of the value function This path is mathematically equivalent to a global optimization function that satisfies the principle of least action, thereby ensuring that local focus point calculation can achieve near-global function optimization solution; solving high-dimensional nonlinear differential equations is converted to gradient optimization on the potential energy surface.

[0028] This embodiment concretizes the potential well evolution calculation module in embodiment 1, and focuses on how to quickly solve the optimal path; This module uses manifold learning techniques such as Isomap or LLE to map high-dimensional state space to low-dimensional manifold , where ; In order to avoid the huge overhead of solving the Navier-Stokes equation in the whole field, this unit introduces the reduced-order approximation operator of Hamilton-Jacobi-Bellman (HJB) equation; the value function satisfies the HJB equation:

[0029] where is the reduced-order system dynamics equation, is the transient value; instead of solving the complete analytical solution, the operator extraction unit uses a neural network to approximate , that is, the gradient of the energy flow; On this basis, the path optimization unit defines the current state as the starting point on the virtual entropy potential energy field surface, and calculates the gradient descent path that makes the point enter the global optimal potential well at the fastest speed; the specific algorithm uses a discretized gradient update rule: , where is the learning rate; This process realizes the conversion of solving high-dimensional nonlinear differential equations in traditional methods to gradient optimization on the potential energy surface, reduces the computational complexity from the CFD simulation level to the gradient calculation level , so that the optimization calculation that originally takes several hours can be completed in milliseconds.

[0030] Embodiment 5: The reverse damping control module includes: A damping calculation unit is used to introduce a reverse entropy damping term, increase a virtual friction force proportional to the rate of change in the mathematical equation, simulate the thermal inertia of the physical world, and prevent regulation oscillation; An instruction decoding unit is used to inversely solve the abstract energy flow direction of the calculated gradient path into specific actuator instructions, and the actuator instructions include the mill speed and the secondary air door opening degree.

[0031] This embodiment elaborates on the reverse damping control module in Embodiment 1, focusing on solving the stability and executability of the control. The damping calculation unit introduces a reverse entropy damping term. The ideal driving force obtained by gradient descent Based on this, add a factor related to the rate of change of state. The virtual friction force is proportional to the equation, and the formula is: ;in This is the coefficient matrix of the virtual equations, whose physical dimensions are set as generalized force / generalized velocity to ensure the consistency of the equation dimensions. The force defined here is not mechanical force, but rather an abstract force driving the evolution of the state vector in the generalized state space, and its dimensions are adaptive to the state variables. Physical units; diagonal elements of a matrix The calculation formula is ,in The preset virtual inertial mass factor is set according to the total installed capacity of the system; for example, a 600MW module is set to [value missing]. Magnitude; For the first The physical devices corresponding to each control loop have the same thermal inertia time; that is... And it includes the efficiency of dimension conversion; here The unit of in the generalized coordinate system is represented as the conjugate momentum unit of the state quantity, ensuring that Strict dimensional consistency of all terms on the right-hand side of the equation; The physics of this setup follows a critical principle: the same applies to thermal inertia time. Devices with smaller, more sensitive responses are prone to overshoot and thus impose a larger virtual response. To prevent oscillations; for large inertial devices, damping should be reduced to avoid an excessively slow system response; when the system state changes too rapidly, i.e. When the pressure is high, it generates reverse resistance, thereby preventing oscillations such as furnace pressure instability caused by frequent and large opening and closing of the damper. The instruction decoding unit will calculate the corrected driving force This abstract energy flow direction can be deduced into a specific executor command vector. ; In this embodiment, the Jacobian matrix Defined as a virtual potential energy field Provide control The partial derivative matrix, i.e. This matrix is ​​calculated online in real time using a neural network model built in the data mapping module through an automatic differentiation mechanism; the Jacobian matrix is ​​based on this real-time calculation. The least squares method is used for control commands, and the formula is as follows: ;in is a regularization factor, and the value range is to , used to ensure numerical stability when the Jacobian matrix approaches a singular value or a pathological value, to prevent the divergence of control commands, so as to obtain the set value of the specific actuator such as the mill speed and the secondary air damper opening degree; The reverse entropy damping mechanism ingeniously introduces the inertia concept existing in the physical entity into the digital twin algorithm, and filters out the high-frequency jitter instructions that are theoretically optimal but physically unexecutable at the pure mathematical level.

[0032] Embodiment 6: The system runs on an edge computing gateway configured with a tensor processing unit; When the data mapping module stores the virtual entropy potential energy field, a sparse matrix storage format is used to optimize memory consumption; The data mapping module processes Modbus or OPC protocol data streams from the distributed control system, and the sampling frequency is not less than 1 Hz; This embodiment specifically describes the hardware deployment and data protocol of the system of embodiment 1; In terms of hardware environment, the system runs on an edge computing gateway configured with a tensor processing unit (TPU); TPU is selected to solve the large number of tensor multiplication operations involved in graph neural network and field gradient calculation; For memory optimization, the data mapping module uses the compressed sparse row (CSR) storage format when storing the virtual entropy potential energy field, which is a three-dimensional matrix ; Specifically, instead of directly storing the dense matrix , three one-dimensional arrays are stored: a non-zero value array , a column index array , and a row offset array ; In addition, the thermodynamic field of the furnace space has obvious local correlation characteristics, and the arrangement of physical sensors in the three-dimensional space has natural sparsity, with high-frequency transient information points only in the combustion core area and the fluid boundary layer. Therefore, by using a threshold to interrupt weakly related areas, only the boundary and specific flow line areas have high gradient information, and sparse storage compresses the memory occupancy space to less than 10% of the original, making it suitable for the limited resources of the edge gateway. In terms of communication protocol, the data mapping module establishes a Socket connection through an Ethernet interface, parses Modbus TCP messages or subscribes to OPC UA nodes, processes real-time data streams, and sets the sampling frequency to , i.e. the time interval , to meet the real-time requirements of power plant process control.

[0033] Embodiment 7: The data mapping module is also used to: Based on the continuity reconstruction of the entropy field, the values of the uninstalled sensor regions are inferred, and the virtual sensor emerges to provide early warning for local high-temperature corrosion. The uninstalled sensor regions include the furnace center flame temperature region.

[0034] This embodiment describes the additional functions of the data mapping module as a soft measurement tool. Based on the continuity reconstruction mechanism of the entropy field, the values of the uninstalled sensor regions are inferred. Specifically, let the furnace space coordinates be , and the known sensor position set be . Since the potential energy field is a continuous function constructed by a physical information neural network (PINN), it satisfies the continuity condition over the entire definition domain . Therefore, for any uninstalled sensor coordinate point , such as the furnace center flame temperature region, the module directly calculates through forward propagation:

[0035] to obtain the inferred value and form the virtual sensor emergence. The system sets a high-temperature corrosion warning threshold , and when is monitored, a warning signal is triggered. This feature achieves the effect of adding sensors at zero cost. Through the furnace center temperature field deduced by the algorithm, the system can early detect local high-temperature corrosion or coking trends and provide warnings.

[0036] Embodiment 8: The closed-loop execution module also includes a safety logic control unit: The safety logic control unit is used to monitor the unit load state. When the unit is in the 30% to 100% load segment, the optimization algorithm is enabled. When the unit is in the non-steady-state extreme working condition of the start-stop machine stage, the safety logic control unit automatically degrades the system to a safety logic control.

[0037] This embodiment perfects the safety mechanism of the closed-loop execution module; The safety logic control unit embedded in the module executes the following dual-mode switching logic: define the unit load as , and the load change rate as . The system sets the lower limit of the steady-state load as , the upper limit as , and the load change rate safety threshold as . In each control period, the unit judges the condition . If is true, the system enables the optimization algorithm based on digital twinning, and outputs control instructions. If If false, that is, when the unit is in the start-stop phase, tripping or non-steady-state extreme working condition of load fluctuation, the safety logic control unit automatically degrades the system to safety logic control, switches the switch , directly outputs conservative instructions through hard-wired logic or PID controller , and prioritizes to ensure that the unit does not stall and does not overpressure; this strategy builds a line of defense for the system and clearly defines the boundary of artificial intelligence intervention in industrial control.

[0038] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A digital twin-based optimized control system for thermal power plants, characterized in that: include: The data mapping module is used to collect real-time data from the distributed control system of thermal power plants and reconstruct discrete measurement point data into a virtual entropy-increasing potential energy field that reflects the thermodynamic state of the unit through a graph neural network. The potential well evolution calculation module is used to set a global optimal well with energy consumption and emission as constraints in the virtual entropy-increasing potential energy field, and to calculate the geodesic gradient path of the current state point evolving towards the global optimal well. The reverse damping control module is used to calculate the corresponding actuator control command based on the geodesic gradient path, superimposed with a virtual friction damping term simulating the thermal inertia of the device. The closed-loop execution module is used to send the control commands of the execution mechanism to the execution mechanism, driving the physical unit state to approach the global optimal well along the geodesic gradient path.

2. The optimized control system for thermal power plants based on digital twins according to claim 1, characterized in that: The data mapping module includes: Spatial discrete element, used to discretize the boiler space into a grid, using the values ​​of sparse sensors as anchor points; The field reconstruction unit is used to predict the virtual potential energy of the unmeasured region through graph convolutional networks or kriging interpolation, and reconstruct discrete measurement points into a continuous virtual potential energy surface. The virtual potential energy surface is constructed using a hybrid architecture of physical information neural network and sparse field reconstruction, wherein physical constraints are embedded in the loss function of the physical information neural network.

3. The optimized control system for thermal power plants based on digital twins according to claim 1, characterized in that: The spatial transformation performed by the data mapping module follows the following mapping logic: The control variables of coal feed rate and damper opening are mapped to potential energy injected into the field; The operating state of furnace temperature and pressure distribution is mapped to field strength distribution; The optimization objectives of minimizing energy consumption and emissions are mapped to the principle of minimum action of the field, causing the system to tend to evolve in the direction of minimizing entropy increase or the optimal energy dissipation path.

4. The optimized control system for thermal power plants based on digital twins according to claim 1, characterized in that: The potential well evolution calculation module includes: The operator extraction unit is used to extract evolution equations on low-dimensional manifolds using manifold learning, introduces a reduced-order approximation operator for the Hamilton-Jacobi-Bellman equation, and focuses on the gradient of energy flow. The path optimization unit is used to define the current state as a point on the virtual entropy-increasing potential energy field surface, calculate the gradient descent path that allows the point to enter the global optimal well at the fastest speed, and transform the solution of the high-dimensional nonlinear differential equation system into gradient optimization on the potential energy surface.

5. The optimized control system for thermal power plants based on digital twins according to claim 1, characterized in that: The reverse damping control module includes: The damping calculation unit is used to introduce a reverse entropy damping term, adding a virtual friction force proportional to the rate of change to the mathematical equation to simulate the thermal inertia of the physical world and prevent regulation oscillations. The instruction decoding unit is used to convert the calculated gradient path, which is an abstract energy flow direction, into specific actuator instructions, including the mill speed and the secondary damper opening.

6. The optimized control system for thermal power plants based on digital twins according to claim 1, characterized in that: The system runs on an edge computing gateway equipped with a tensor processing unit. When the data mapping module stores the virtual entropy-increasing potential energy field, it adopts a sparse matrix storage format to optimize memory consumption. The data mapping module processes Modbus or OPC protocol data streams from the distributed control system, with a sampling frequency of not less than 1Hz.

7. The optimized control system for thermal power plants based on digital twins according to claim 1, characterized in that: The data mapping module is also used for: Based on the continuous reconstruction of the entropy field, the values ​​in the area where no sensors are installed are inferred, forming a virtual sensor emergence, which can be used to provide early warning of local high temperature corrosion. The area where no sensors are installed includes the flame temperature zone at the center of the furnace.

8. The optimized control system for thermal power plants based on digital twins according to claim 1, characterized in that: The closed-loop execution module also includes a security logic control unit: The safety logic control unit is used to monitor the unit load status, and when the unit is in the 30% to 100% load range, the optimization algorithm is activated. When the unit is detected to be in an unsteady extreme operating condition during the start-up and shutdown phase, the safety logic control unit automatically downgrades the system to safety logic control.

Images