Boiler steam production optimization device based on electrical load dynamic matching

Abstract

The present application relates to the technical field of boiler steam production optimization of power plant electricity generation, and particularly relates to a boiler steam production optimization device based on electrical load dynamic matching, which comprises a state sensing module that collects electrical load instructions and boiler operation data to generate a multi-dimensional state matrix; a twin body construction module that constructs a digital twin body model based on the matrix and outputs system parameters; a parallel simulation module that performs multi-strategy deduction in a virtual environment to generate a candidate control strategy set; a rolling optimization module that fuses real-time data and the strategy set to output a fuel air supply instruction sequence through rolling calculation; a feedforward compensation module that dynamically decouples and eliminates the multivariable coupling effect on the instructions; and an execution intervention module that realizes safe execution of the instructions and rapid intervention in risks. The modules are connected through a data flow closed loop to solve the problem of combustion oscillation caused by the mismatch between the dynamic changes of the electrical load and the thermal inertia of the boiler, and to realize dynamic matching of the steam production and the electrical load and safe protection of the boiler.

Description

Technical Field

[0001] This invention relates to the field of boiler steam production optimization technology in power plant power generation, and particularly to a boiler steam production optimization device based on dynamic matching of electrical load. Background Technology

[0002] Gas load represents the real-time changing demand for electricity consumption within the power system, and its fluctuations directly affect the output power of generator sets; boiler steam refers to the high-temperature, high-pressure steam generated by the boiler through fuel combustion, which drives the turbine to rotate and converts it into electrical energy; dynamic matching refers to adjusting the boiler's fuel supply and feedwater flow through an automated control system based on real-time changes in electrical load, in order to control steam production and parameters, so that the steam supply can adapt to changes in power generation demand, thereby maintaining the overall stability and efficiency of power plant operation.

[0003] Existing electrical load and boiler steam dynamic matching optimization technologies have the following technical pain points: the dynamic change rate of electrical load far exceeds the thermal inertia time constant of the boiler system. The steam production optimization algorithm adjusts the combustion process based on real-time electrical signals, but in the transient response, a time mismatch occurs between the algorithm's rapid decision-making and the boiler's slow thermal dynamics, leading to an imbalance between fuel supply and combustion stability, and inducing periodic combustion oscillations. For example, when the power plant participates in grid frequency regulation, the electrical load rises and falls sharply. The algorithm immediately instructs to increase fuel to improve steam production, but the boiler's heat capacity causes a delay in the increase of steam output. The fuel and air mixing becomes unbalanced during the transition, and the pressure and temperature in the combustion chamber fluctuate violently. The oscillation energy is transferred to the membrane water-cooled wall, causing local overheating and alternating thermal stress. Long-term accumulation leads to pipe wall cracks or leaks, endangering the safe operation of the boiler. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a boiler steam production optimization device based on dynamic matching of electrical load. This invention solves the technical problem that the mismatch between the dynamic changes of electrical load and the boiler's thermal inertia time constant causes combustion oscillations in the steam production optimization algorithm during transient processes, which in turn damages the boiler's membrane water-cooled wall.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows:

[0006] The boiler steam production optimization device based on dynamic electrical load matching provided by this invention includes: The status perception module collects electrical load commands, boiler sensor data, fuel parameters, and historical operating data, and outputs a multi-dimensional status matrix. The digital twin construction module connects to the state perception module, receives the multi-dimensional state matrix output by the state perception module, constructs a digital twin model of the boiler based on the multi-dimensional state matrix, and outputs the digital twin model, system state estimate, and dynamic coupling matrix. The parallel simulation module connects the twin construction module and the state perception module. It receives the digital twin model output by the twin construction module and the electrical load command sequence output by the state perception module, performs virtual simulation on the digital twin model, and outputs a set of candidate control strategies. The rolling optimization module connects the parallel simulation module, the state awareness module, and the twin construction module. It receives the candidate control strategy set output by the parallel simulation module, the real-time deviation signal output by the state awareness module, and the system state estimate output by the twin construction module. It calculates the fuel delivery command sequence through rolling optimization. The feedforward compensation module connects the rolling optimization module and the twin construction module. It receives the fuel air supply command sequence output by the rolling optimization module and the dynamic coupling matrix output by the twin construction module, performs decoupling compensation on the fuel air supply command sequence, and outputs the compensated command sequence. The execution intervention module connects the feedforward compensation module and the state perception module. It receives the compensated instruction sequence output by the feedforward compensation module and the safety monitoring signal output by the state perception module, distributes the compensated instruction sequence to the actuator, and performs safety intervention based on the safety monitoring signal.

[0007] Furthermore, in the boiler steam production optimization device based on dynamic matching of electrical load described in this invention, the state perception module includes a data acquisition unit, an image acquisition unit, a component analysis unit, a historical query unit, and a preprocessing unit. The data acquisition unit synchronously acquires the electrical load command curve of the power grid dispatching system, the main steam pressure and flow signal of the boiler, and the thermocouple signals of the wall temperature of each stage of superheater at millisecond intervals, and outputs process variable data. The image acquisition unit continuously acquires a sequence of images of the flame distribution throughout the furnace through the observation hole of the burner using a high-temperature industrial endoscope, and outputs image data. The component analysis unit receives real-time coal composition spectral data provided by the online coal quality analyzer and outputs coal quality data. The historical query unit retrieves waveform segments of historical combustion oscillation events that are similar to the current load, coal quality, and air distribution mode from the relational database and outputs historical waveform data. The preprocessing unit receives process variable data output by the data acquisition unit, image data output by the image acquisition unit, coal quality data output by the component analysis unit, and historical waveform data output by the historical query unit. It performs time-stamp alignment, outlier removal, and standardization on the various types of data collected, and generates a multi-dimensional state matrix with strictly synchronized timestamps, which is then output to the twin construction module.

[0008] Furthermore, in the boiler steam production optimization device based on dynamic matching of electrical load described in this invention, the twin construction module includes a mechanism model identification unit, a visual feature extraction unit, a model fusion unit, and a derived information output unit. The mechanism model identification unit receives the process variables in the multidimensional state matrix output by the state perception module, inputs the process variables into the recursive least squares algorithm with forgetting factor, dynamically adjusts the thermal inertia time constant distribution parameters in the nonlinear distributed parameter mechanism model of the boiler, and outputs the updated mechanism model. The visual feature extraction unit receives the flame image sequence from the multidimensional state matrix output by the state perception module, inputs the flame image sequence into a pre-trained deep convolutional neural network, extracts combustion stability feature vectors representing the center position of flame brightness, pulsation frequency, and fullness, and outputs the feature vectors. The model fusion unit receives the updated mechanism model output by the mechanism model identification unit and the feature vector output by the visual feature extraction unit, injects the feature vector into the updated mechanism model in a weighted summation manner, corrects the combustion reaction rate term, and outputs the fused digital twin model. The derived information output unit receives the fused digital twin model output by the model fusion unit, extracts real-time thermal inertia parameters, system state estimates, and dynamic coupling matrices from the fused digital twin model, and sends them to the parallel simulation module, rolling optimization module, and feedforward compensation module, respectively.

[0009] Furthermore, in the boiler steam production optimization device based on dynamic matching of electrical load described in this invention, the parallel simulation module includes a baseline extrapolation unit, an optimization exploration unit, a robust testing unit, and a confidence assessment unit. The baseline simulation unit receives the digital twin model output by the twin construction module and the electrical load command sequence output by the state perception module. It inputs the electrical load command sequence into the digital twin model to perform open-loop simulation, predicts the baseline response curves of steam production and furnace pressure, and outputs baseline response data. The optimization exploration unit receives the digital twin model output by the twin construction module, runs an oscillation suppression policy network trained with deep deterministic policy gradient in the digital twin model, explores the optimal path to suppress pressure fluctuations by adjusting the virtual fuel command, and outputs the optimized control command trajectory. The robustness testing unit receives the digital twin model output by the twin construction module, injects a step change disturbance of coal quality parameters into the digital twin model, tests the stability boundary of the candidate control strategy under different disturbance amplitudes, and outputs the robustness test results. The confidence assessment unit receives the baseline response data output by the baseline extrapolation unit, the optimized control command trajectory output by the optimization exploration unit, and the robustness test results output by the robustness test unit. It compares the triple extrapolation results, calculates the steam production tracking error, pressure oscillation peak value, and robustness score of each candidate strategy, and generates a set of candidate control strategies with risk assessment labels, which is then sent to the rolling optimization module.

[0010] Furthermore, in the boiler steam production optimization device based on dynamic matching of electrical load described in this invention, the rolling optimization module includes a data integration unit, a problem construction unit, and a rolling solution unit; The data integration unit receives candidate strategy sets from the parallel simulation module, real-time process variable deviation signals from the state awareness module, and system state estimates from the twin construction module, and outputs integrated data. The problem construction unit receives the integrated data output by the data integration unit, constructs an objective function using the sum of squared errors in steam production tracking, the predicted index of combustion oscillation intensity, and the penalty term for the rate of change of actuator action, defines the constraints using the dynamic safe operation envelope provided by the digital twin model, and outputs the optimization problem. The rolling solution unit receives the optimization problem output by the problem construction unit, takes the strategy with the highest evaluation in the candidate strategy set as the initial solution, solves the optimal fuel air delivery command sequence in the future finite time domain in each sampling period, and outputs only the command value at the current moment to the feedforward compensation module.

[0011] Furthermore, in the boiler steam production optimization device based on dynamic matching of electrical load described in this invention, the feedforward compensation module includes an instruction receiving unit, a relationship acquisition unit, a decoupling calculation unit, and a compensation output unit. The instruction receiving unit receives the fuel air delivery instruction sequence output by the rolling optimization module; The relationship acquisition unit receives the dynamic coupling matrix output by the twin construction module; The decoupling calculation unit receives the fuel air supply command sequence output by the command receiving unit and the dynamic coupling matrix output by the relationship acquisition unit, calculates the coupling gain of fuel command changes on furnace pressure based on the dynamic coupling matrix, and generates the corresponding induced draft command feedforward compensation amount. The compensation output unit receives the induced draft command feedforward compensation amount output by the decoupling calculation unit, superimposes the compensation amount onto the fuel supply air command sequence, and outputs the decoupled fuel, supply air, and induced draft coordinated command to the execution intervention module.

[0012] Furthermore, in the boiler steam production optimization device based on dynamic matching of electrical load described in this invention, the execution intervention module includes an instruction distribution unit, a safety monitoring unit, a fuzzy inference unit, a threshold comparison unit, and an intervention execution unit; The instruction distribution unit receives the compensated instruction sequence output by the feedforward compensation module, converts the instruction sequence into fieldbus protocol messages, and sends them to the fuel feeder frequency converter and the blower actuator. The safety monitoring unit receives the safety monitoring signal output by the status perception module and reads the water-cooled wall temperature, the tube wall temperature difference change rate and the furnace pressure oscillation mode signal in parallel. The fuzzy inference unit receives the monitoring signal output by the security monitoring unit, inputs the monitoring signal into the fuzzy logic system, performs inference through the rule base, and outputs a quantified risk value. The threshold comparison unit receives the quantized risk value output by the fuzzy inference unit and the system state estimate provided by the twin construction module, and compares the risk value with the safety threshold in the system state estimate in real time. The intervention execution unit receives the comparison result output by the threshold comparison unit. When the risk value exceeds the intervention threshold, it directly sends a control mode switching command and a gradient limiting signal to the execution mechanism.

[0013] Furthermore, in the boiler steam production optimization device based on dynamic matching of electrical load described in this invention, the confidence evaluation unit of the parallel simulation module feeds back the robustness test results obtained during the deduction process to the twin construction module, and the twin construction module updates the thermal inertia time constant distribution parameters of the mechanism model using the robustness test results; the safety intervention event data recorded by the execution intervention module is fed back to the oscillation suppression strategy network training dataset of the parallel simulation module, and the oscillation suppression strategy network adjusts the network weight parameters and exploration strategy according to the safety intervention event data.

[0014] Furthermore, in the boiler steam production optimization device based on dynamic matching of electrical load described in this invention, the rolling solution unit of the rolling optimization module outputs a sequence of instructions to the feedforward compensation module, which decouples the sequence of instructions. The instruction distribution unit of the execution intervention module sends the decoupled sequence of instructions to the field execution mechanism. The state perception module collects the actual response data of the boiler and generates a new state matrix. The digital twin construction module receives the new state matrix and updates the model parameters. The updated digital twin model provides a new system state estimate and safe operation envelope constraint for the rolling optimization of the next sampling period.

[0015] Furthermore, in the boiler steam production optimization device based on dynamic matching of electrical load described in this invention, the real-time data collected by the state perception module drives the twin construction module to generate a digital twin model; the parallel simulation module performs virtual simulation on the digital twin model to generate a candidate control strategy set; the rolling optimization module performs rolling optimization calculations based on the candidate control strategy set and real-time state data; the feedforward compensation module performs decoupling compensation for optimization instructions; the execution intervention module realizes safe execution of instructions and risk intervention; and the modules are connected through a closed-loop data flow to form a complete control system.

[0016] Beneficial effects of this invention: The boiler steam production optimization device based on dynamic matching of electrical load provided by this invention uses a state perception module to synchronously collect multi-source data to generate a multi-dimensional state matrix. A twin construction module builds a high-fidelity digital twin model based on this matrix. A parallel simulation module performs multi-threaded virtual simulation on the model to generate a set of candidate control strategies. A rolling optimization module integrates real-time data and strategy sets and outputs a fuel supply command sequence through rolling calculation. A feedforward compensation module dynamically decouples the commands to eliminate multi-variable coupling interference. An execution intervention module enables safe execution of commands and rapid intervention against risks. The modules are connected in a closed loop through data flow to form a complete control system. This effectively solves the combustion oscillation problem caused by the mismatch between dynamic changes in electrical load and the boiler's thermal inertia time constant, realizes dynamic matching between boiler steam production and electrical load, and improves system stability and safety. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0018] Figure 1 This is a system architecture diagram of the boiler steam production optimization device based on dynamic matching of electrical load according to the present invention. Detailed Implementation

[0019] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0020] Please see Figure 1The boiler steam production optimization device based on dynamic matching of electrical load provided by the present invention includes: The status perception module collects electrical load commands, boiler sensor data, fuel parameters, and historical operating data, and outputs a multi-dimensional status matrix. The digital twin construction module connects to the state perception module, receives the multi-dimensional state matrix output by the state perception module, constructs a digital twin model of the boiler based on the multi-dimensional state matrix, and outputs the digital twin model, system state estimate, and dynamic coupling matrix. The parallel simulation module connects the twin construction module and the state perception module. It receives the digital twin model output by the twin construction module and the electrical load command sequence output by the state perception module, performs virtual simulation on the digital twin model, and outputs a set of candidate control strategies. The rolling optimization module connects the parallel simulation module, the state awareness module, and the twin construction module. It receives the candidate control strategy set output by the parallel simulation module, the real-time deviation signal output by the state awareness module, and the system state estimate output by the twin construction module. It calculates the fuel delivery command sequence through rolling optimization. The feedforward compensation module connects the rolling optimization module and the twin construction module. It receives the fuel air supply command sequence output by the rolling optimization module and the dynamic coupling matrix output by the twin construction module, performs decoupling compensation on the fuel air supply command sequence, and outputs the compensated command sequence. The execution intervention module connects the feedforward compensation module and the state perception module. It receives the compensated instruction sequence output by the feedforward compensation module and the safety monitoring signal output by the state perception module, distributes the compensated instruction sequence to the actuator, and performs safety intervention based on the safety monitoring signal.

[0021] The state awareness module is responsible for synchronously collecting multi-source heterogeneous data. This module obtains real-time electrical load command curves from the power grid dispatch system, reflecting the power system's power demand on generator units. Simultaneously, the module collects process variables such as main steam pressure, main steam flow, and wall temperatures of superheaters and reheaters at each stage through a sensor network deployed on the boiler body. The module also connects to the fuel supply system to obtain fuel parameters such as coal feed rate, air volume, and air distribution ratio, and retrieves historical data similar to the current operating conditions from the power plant's historical database, especially records of past combustion oscillations. All data, after time-stamp alignment, outlier removal, and standardization preprocessing, is fused to generate a multi-dimensional state matrix with strictly synchronized timestamps. This matrix serves as a digital snapshot of the boiler's operating state, providing a unified input for all subsequent analysis and decision-making.

[0022] The digital twin construction module receives the aforementioned multidimensional state matrix and uses it as the basis to drive the construction and online updating of the boiler's digital twin model. The construction process integrates mechanistic modeling and data-driven methods. The module's built-in mechanistic model, based on the laws of conservation of mass, energy, and momentum, describes the physicochemical processes such as combustion, heat transfer, and flow within the boiler. The module utilizes real-time data and a parameter identification algorithm to dynamically correct key parameters in the mechanistic model, such as the thermal inertia time constant reflecting the boiler's thermal dynamic response characteristics. Simultaneously, the module may also integrate machine learning-based models to capture complex nonlinear relationships that are difficult to accurately describe using mechanistic equations. The final generated digital twin is a high-fidelity virtual mapping of the boiler, capable of simulating the dynamic response of a real boiler under various commands and disturbances. The digital twin construction module extracts crucial control information from this model, including estimates of the current intangible state and a dynamic coupling matrix reflecting the interactions between multiple variables within the boiler.

[0023] The parallel simulation module creates a secure virtual testing environment using a digital twin model. This module receives future electrical load command sequences and inputs them into the digital twin for forward-looking simulation. The core objective of the simulation is to pre-evaluate the effectiveness and risks of different control strategies before the actual control commands are issued to the physical boiler. The module runs multiple candidate control strategies in parallel on the digital twin, such as different fuel and air ratio schemes. Each simulation records the predicted trajectories of key parameters such as steam production, furnace pressure, and temperature. By analyzing and comparing the simulation results, the module can evaluate the performance of each strategy in tracking load commands, maintaining combustion stability, and robustness, and assign risk assessment labels to them. Finally, the module outputs a filtered and ranked set of candidate control strategies, providing optimization options for subsequent real-time decision-making.

[0024] The rolling optimization module is the core of real-time decision-making. This module receives a candidate strategy set from the parallel simulation module, a real-time deviation signal between the boiler's current operating state and the target value from the state awareness module, and system state estimates from the digital twin construction module. Based on this information, the module solves a rolling time-domain optimization problem in each control cycle. The goal of the optimization problem is to find the optimal sequence of fuel and air supply commands that, over a future period, can quickly and accurately respond to changes in electrical load, minimize combustion oscillations, ensure boiler safety, and simultaneously accommodate the actuator's movement. During optimization, a digital twin model is used as a constraint to predict the controlled object's dynamics, and the candidate strategy set is referenced to accelerate the solution or provide high-quality initial values. After the solution is completed, only the first command value corresponding to the current moment in the command sequence is output. The next cycle then re-optimizes based on the latest measurements, forming a rolling forward decision-making mechanism.

[0025] The feedforward compensation module specifically addresses the dynamic coupling problem among multiple variables within the boiler. A boiler is a complex multi-input multi-output system; for example, adjusting the fuel quantity simultaneously affects steam output and furnace pressure. This coupling can cause a control action on one variable to adversely affect another. The feedforward compensation module obtains a dynamic coupling matrix describing the coupling strength between variables under the current operating conditions from the twin construction module. When the module receives the fuel and ventilation command sequence calculated by the rolling optimization module, it calculates the feedforward compensation amount required for "decoupling" based on this coupling matrix. For example, while increasing the fuel quantity command, a corresponding ventilation or induced draft compensation command is pre-calculated and superimposed to offset the expected impact of the fuel increase on furnace pressure, thereby allowing for relatively independent control of steam output and furnace pressure. After decoupling compensation, the dynamic quality of the command sequence is improved.

[0026] The intervention module assumes the dual responsibility of final instruction issuance and operational safety monitoring. It converts the compensated instruction sequence from the feedforward compensation module into control signals or communication messages recognizable by the field actuators, and issues them to equipment such as the coal feeder, forced draft fan, and induced draft fan. Simultaneously, an independent safety monitoring thread runs in parallel. This thread continuously acquires key signals directly reflecting the boiler's safety status from the status perception module, such as the temperature at specific points on the water-cooled wall and the frequency and amplitude of furnace pressure oscillations. After rapid risk assessment, if the system determines that the boiler's operating status deviates from safety boundaries, the safety monitoring thread immediately triggers intervention. Intervention methods include, but are not limited to, forcibly switching to a preset safety control mode or limiting the overshoot of the issued instructions. This design allows the safety protection loop to run in parallel with and prioritize the main control loop, providing a direct and rapid hardware safety barrier for boiler equipment, especially the membrane water-cooled wall, while the main control loop pursues performance optimization.

[0027] Specifically, the boiler steam production optimization device based on dynamic matching of electrical load according to the present invention includes a status sensing module comprising a data acquisition unit, an image acquisition unit, a component analysis unit, a historical query unit, and a preprocessing unit. The data acquisition unit synchronously acquires the electrical load command curve of the power grid dispatching system, the main steam pressure and flow signal of the boiler, and the thermocouple signals of the wall temperature of each stage of superheater at millisecond intervals, and outputs process variable data. The image acquisition unit continuously acquires a sequence of images of the flame distribution throughout the furnace through the observation hole of the burner using a high-temperature industrial endoscope, and outputs image data. The component analysis unit receives real-time coal composition spectral data provided by the online coal quality analyzer and outputs coal quality data. The historical query unit retrieves waveform segments of historical combustion oscillation events that are similar to the current load, coal quality, and air distribution mode from the relational database and outputs historical waveform data. The preprocessing unit receives process variable data output by the data acquisition unit, image data output by the image acquisition unit, coal quality data output by the component analysis unit, and historical waveform data output by the historical query unit. It performs time-stamp alignment, outlier removal, and standardization on the various types of data collected, and generates a multi-dimensional state matrix with strictly synchronized timestamps, which is then output to the twin construction module.

[0028] The status awareness module, through the coordinated operation of multiple dedicated units within it, achieves synchronous acquisition and fusion of multi-source, heterogeneous operational data. The data acquisition unit actively queries or receives electrical load command curves issued by the power grid dispatching system at millisecond intervals via communication protocols; these commands represent the power targets the boiler needs to track. Simultaneously, this unit acquires signals from the boiler's sensor network in parallel, including readings from pressure transmitters and flow meters installed on the main steam pipeline, and thermocouple temperature signals distributed across the walls of each stage of the superheater. All continuously changing measurements are categorized as process variable data, directly reflecting the boiler's real-time thermal state.

[0029] To address the complex process of combustion, which is difficult to characterize with a single parameter, the image acquisition unit deploys a high-temperature resistant industrial endoscope at the burner's observation port. The endoscope continuously captures images of the flame inside the furnace at a fixed frame rate, forming a sequence of images showing the flame distribution throughout the furnace. This image sequence visually records the flame's morphology, brightness distribution, and pulsation, providing crucial visual evidence for assessing combustion stability and uniformity.

[0030] The composition analysis unit focuses on fuel characteristics, a key perturbation factor. The unit receives real-time analysis results from an online coal quality analyzer (such as one employing laser-induced breakdown spectroscopy) via a data interface, acquiring spectral data of the volatile matter, fixed carbon, ash, and other components of the coal fed into the furnace, and quantifying these into coal quality data. Real-time changes in fuel composition directly affect the combustion reaction rate and calorific value, making them parameters that must be considered in feedforward control and robust design.

[0031] The historical query unit serves to provide experiential references for current decision-making. Based on key operating conditions such as current load levels, measured coal quality data, and air distribution patterns, the unit performs similarity matching searches in the historical operating archives of a relational database. Its goal is to identify combustion oscillation events that have occurred under similar historical conditions and extract the corresponding process variable waveform segments as historical waveform data. Historical patterns help the system identify whether the current state is approaching an unstable risk boundary.

[0032] The preprocessing unit is the convergence and normalization point of the aforementioned multi-source information streams. This unit simultaneously receives output data from the four aforementioned units, including process variable data, image data, coal quality data, and historical waveform data. Since the data originates from different independent systems, their sampling times and frequencies may not be naturally synchronized. Therefore, the primary task of the preprocessing unit is to time-stamp align all input data streams, typically using interpolation algorithms to unify the signals onto a common time series of the system's master clock. Next, the unit applies statistical methods (such as the Laida criterion) to identify and eliminate outliers in the process variable data, removing invalid data points caused by momentary sensor malfunctions or interference. Finally, the unit standardizes data with varying physical dimensions to eliminate the impact of orders of magnitude differences on subsequent model calculations. After this series of standardized operations, the preprocessing unit ultimately generates a multidimensional state matrix with strictly synchronized timestamps, consistent data quality, and multi-dimensional information. This matrix is ​​a complete and consistent snapshot of the boiler's operating state in digital space, which is output in real-time to the twin building module as direct input for constructing and updating the high-fidelity digital model.

[0033] Specifically, the boiler steam production optimization device based on dynamic matching of electrical load according to the present invention includes a twin construction module comprising a mechanism model identification unit, a visual feature extraction unit, a model fusion unit, and a derived information output unit. The mechanism model identification unit receives the process variables in the multidimensional state matrix output by the state perception module, inputs the process variables into the recursive least squares algorithm with forgetting factor, dynamically adjusts the thermal inertia time constant distribution parameters in the nonlinear distributed parameter mechanism model of the boiler, and outputs the updated mechanism model. The visual feature extraction unit receives the flame image sequence from the multidimensional state matrix output by the state perception module, inputs the flame image sequence into a pre-trained deep convolutional neural network, extracts combustion stability feature vectors representing the center position of flame brightness, pulsation frequency, and fullness, and outputs the feature vectors. The model fusion unit receives the updated mechanism model output by the mechanism model identification unit and the feature vector output by the visual feature extraction unit, injects the feature vector into the updated mechanism model in a weighted summation manner, corrects the combustion reaction rate term, and outputs the fused digital twin model. The derived information output unit receives the fused digital twin model output by the model fusion unit, extracts real-time thermal inertia parameters, system state estimates, and dynamic coupling matrices from the fused digital twin model, and sends them to the parallel simulation module, rolling optimization module, and feedforward compensation module, respectively.

[0034] The digital twin construction module receives a multi-dimensional state matrix from the state perception module and drives two technical paths in parallel based on this matrix to construct a high-fidelity digital twin model. The mechanism model identification unit specifically processes the process variable data stream in the multi-dimensional state matrix, such as main steam pressure, fuel supply rate, primary air volume, and secondary air volume. The unit employs a recursive least squares algorithm with a forgetting factor, the core of which is to assign an exponentially decaying weight to historical data. As new process variable measurements are continuously input, the algorithm dynamically and selectively reduces the influence of old data, thereby tracking the slow time-varying characteristics of the boiler system caused by coking, wear, or fuel changes. A key output of the algorithm is the real-time update of the thermal inertia time constant distribution parameter in the boiler's nonlinear distributed parameter mechanism model. This parameter physically characterizes the dynamic delay and inertial distribution from the release of fuel chemical energy to the output of main steam energy. The updated mechanism model can more accurately predict the boiler's thermodynamic response under steady-state and transient conditions.

[0035] The visual feature extraction unit simultaneously processes the flame image sequence in the multi-dimensional state matrix. The unit invokes a pre-trained deep convolutional neural network, typically trained on an image database containing various combustion state markers. The network analyzes the flame images frame by frame, automatically extracting high-level, abstract combustion visual features through its multi-layer convolution and pooling structure. These features are quantized into a feature vector, including dimensions such as the coordinates of the flame brightness centroid on the furnace cross-section, the pulsating frequency and amplitude of the brightness signal, and the degree to which the flame fills the furnace. Essentially, the feature vector transforms unstructured visual information into a structured numerical description, directly reflecting the stability, concentration, and dynamic characteristics of combustion.

[0036] The model fusion unit is responsible for organically integrating the results of the two paths mentioned above. The unit receives the updated mechanistic model output from the mechanistic model identification unit and the combustion stability feature vector output from the visual feature extraction unit. The fusion strategy employs a weighted summation method, injecting the feature vector into the combustion reaction rate kinetic equation of the mechanistic model with adjustable weights. For example, when the feature vector shows a significant increase in flame pulsation frequency, the fusion algorithm can correspondingly correct the parameters or terms characterizing combustion instability in the mechanistic model, enabling the digital twin model to not only deduce based on physical laws but also incorporate real-time visual observation information of the combustion state. The model output after fusion correction is a high-fidelity digital twin model that simultaneously reflects the physical laws of the boiler and the current actual combustion state.

[0037] The Derivative Information Output Unit serves as the information distribution hub for the digital twin construction module. This unit receives the fused digital twin model generated by the model fusion unit and extracts three types of derivative information crucial for subsequent control. Real-time thermal inertia parameters are directly obtained from the latest parameter set of the model, reflecting the system's current dynamic response speed. System state estimates utilize the digital twin model and the latest multidimensional state matrix, employing a state observer algorithm to estimate key states that are difficult to measure directly or have measurement lags, such as the detailed temperature field distribution within the furnace or the instantaneous value of combustion efficiency. The dynamic coupling matrix is ​​obtained by analyzing the linearized model of the digital twin model at the current operating point or through numerical perturbation analysis. This matrix quantifies the intensity of the dynamic interaction between control inputs (such as fuel and air supply commands) and controlled outputs (such as steam pressure and furnace negative pressure). The Derivative Information Output Unit sends the real-time thermal inertia parameters and system state estimates to the rolling optimization module, the dynamic coupling matrix to the feedforward compensation module, and the digital twin model to the parallel simulation module, thus providing a unified model and data foundation for all downstream optimization, compensation, and simulation functions.

[0038] Specifically, the boiler steam production optimization device based on dynamic matching of electrical load described in this invention includes a parallel simulation module comprising a baseline extrapolation unit, an optimization exploration unit, a robust testing unit, and a confidence assessment unit. The baseline simulation unit receives the digital twin model output by the twin construction module and the electrical load command sequence output by the state perception module. It inputs the electrical load command sequence into the digital twin model to perform open-loop simulation, predicts the baseline response curves of steam production and furnace pressure, and outputs baseline response data. The optimization exploration unit receives the digital twin model output by the twin construction module, runs an oscillation suppression policy network trained with deep deterministic policy gradient in the digital twin model, explores the optimal path to suppress pressure fluctuations by adjusting the virtual fuel command, and outputs the optimized control command trajectory. The robustness testing unit receives the digital twin model output by the twin construction module, injects a step change disturbance of coal quality parameters into the digital twin model, tests the stability boundary of the candidate control strategy under different disturbance amplitudes, and outputs the robustness test results. The confidence assessment unit receives the baseline response data output by the baseline extrapolation unit, the optimized control command trajectory output by the optimization exploration unit, and the robustness test results output by the robustness test unit. It compares the triple extrapolation results, calculates the steam production tracking error, pressure oscillation peak value, and robustness score of each candidate strategy, and generates a set of candidate control strategies with risk assessment labels, which is then sent to the rolling optimization module.

[0039] The parallel simulation module conducts forward-looking multi-threaded simulations in a secure virtual environment based on the received digital twin model and electrical load command sequence. The baseline simulation unit inputs the complete electrical load command sequence for a future period into the digital twin model, performing an open-loop simulation without any active control intervention. The simulation process mimics the boiler's natural dynamic response when it simply follows the current command without any optimization adjustments. The baseline response data output from the simulation specifically includes the predicted change curve of steam production under expected load changes, the predicted trajectory of main steam pressure, and the estimated fluctuation pattern of pressure at key points in the furnace. This data depicts the possible future state of the system without any intervention, providing an objective, non-interventional reference benchmark for subsequent optimization evaluation.

[0040] The optimization exploration unit deploys and runs an oscillation suppression policy network pre-trained using a deep deterministic policy gradient algorithm within the same digital twin model. This policy network acts as a virtual intelligent controller, aiming to explore control paths that effectively suppress or eliminate furnace pressure fluctuations in a simulation environment by adjusting virtual fuel commands, air supply commands, and other control variables in real time. Based on real-time pressure status and pressure change rate feedback from the simulation environment, and a reward function centered on minimizing pressure oscillation amplitude, the network dynamically generates and outputs an optimized virtual control command trajectory through trial and error learning. This optimized control command trajectory represents a potential high-performance control strategy.

[0041] The robustness testing unit also runs on the digital twin model, but its core task is to test the resilience of candidate control strategies in the face of uncertain disturbances. The unit injects simulated disturbances into the model, such as abrupt changes in the quality of coal entering the furnace, like a sudden decrease in calorific value or a sudden change in volatile matter. Under scenarios with different disturbance magnitudes and types, the robustness testing unit evaluates whether each candidate strategy can maintain stable system operation and records the critical disturbance conditions at which the strategy begins to fail, i.e., the stability boundary. The robustness test results output during the testing process quantify the ability of each strategy to cope with fluctuations in key parameters such as fuel characteristics.

[0042] The confidence assessment unit is the decision-making center of the parallel simulation module. This unit synchronously receives baseline response data from the baseline extrapolation unit, optimized control command trajectories from the optimization exploration unit, and robustness test results from the robustness testing unit. The unit comprehensively compares and analyzes these three extrapolation results. Specifically, it calculates the steam production tracking error of each candidate optimization strategy based on the baseline response data, evaluating its core performance in following load commands; it analyzes the suppression effect of furnace pressure oscillation peaks when applying the optimization strategy; and it scores the robustness of the strategy based on the robustness test results. Through a weighted comprehensive calculation of tracking error, oscillation peak suppression, and robustness score, the confidence assessment unit generates a quantified risk assessment label for each explored candidate strategy, ultimately forming a structured, priority-ranked set of candidate control strategies, which is then sent to the rolling optimization module to provide it with high-quality strategy options that have been fully verified through simulation for online real-time decision-making.

[0043] Specifically, the boiler steam production optimization device based on dynamic matching of electrical load according to the present invention includes a rolling optimization module comprising a data integration unit, a problem construction unit, and a rolling solution unit; The data integration unit receives candidate strategy sets from the parallel simulation module, real-time process variable deviation signals from the state awareness module, and system state estimates from the twin construction module, and outputs integrated data. The problem construction unit receives the integrated data output by the data integration unit, constructs an objective function using the sum of squared errors in steam production tracking, the predicted index of combustion oscillation intensity, and the penalty term for the rate of change of actuator action, defines the constraints using the dynamic safe operation envelope provided by the digital twin model, and outputs the optimization problem. The rolling solution unit receives the optimization problem output by the problem construction unit, takes the strategy with the highest evaluation in the candidate strategy set as the initial solution, solves the optimal fuel air delivery command sequence in the future finite time domain in each sampling period, and outputs only the command value at the current moment to the feedforward compensation module.

[0044] The rolling optimization module performs rolling time-domain optimization calculations based on real-time data and simulation results, ultimately generating a fuel supply air command sequence. The data integration unit, the input to this module, is responsible for receiving and integrating heterogeneous data from the three upstream modules. The unit receives a set of candidate control strategies output by the parallel simulation module, where each strategy includes a series of virtual control command sequences and their corresponding risk assessment labels. The unit also receives real-time process variable deviation signals from the state awareness module, such as the instantaneous deviation between the main steam pressure setpoint and the actual measured value, and the difference between the load command and the current power generation. Furthermore, the unit receives system state estimates from the twin construction module, which are the best current estimates for states inside the boiler that are difficult to measure directly (such as average furnace temperature and combustion efficiency). The data integration unit's task is to align, convert, and cache the input data, which differ in format, timing, and physical meaning, and package them into a unified, time-synchronized integrated data package, preparing for subsequent optimization problem construction.

[0045] The problem-building unit is the core of defining the optimization mathematical problem. The unit receives integrated data from the data integration unit and constructs an optimization model with an objective function and constraints based on this data. The objective function is designed to balance multiple conflicting control objectives. The sum of squared steam production tracking errors is used as the primary term, forcing the optimization result to closely follow changes in electrical load commands; this is fundamental to achieving dynamic matching. The combustion oscillation intensity prediction index is the second term, its calculation relying on the system dynamic characteristics provided by the digital twin building module. Its purpose is to proactively embed the tendency to suppress pressure and temperature oscillations into the command sequence. The actuator action change rate penalty term is the third term, used to constrain the variation amplitude of commands from equipment such as the fuel feeder and blower, avoiding excessively drastic actions that could impact mechanical equipment and worsen the combustion process. Constraints are defined by the dynamic safe operating envelope provided by the digital twin model, such as the upper and lower limits of main steam pressure, the maximum allowable value of superheater wall temperature, and the minimum air-fuel ratio required to maintain stable combustion. These constraints collectively constitute a time-varying, multi-dimensional safe and feasible domain, ensuring that the optimized command sequence always operates within the boiler's safety boundaries. The final output of the problem building unit is a fully defined mathematical optimization problem description with time-varying parameters and constraints.

[0046] The rolling solver unit is the stage that performs numerical calculations and outputs the final command. The unit receives the problem, constructs the optimization problem output by the unit, and solves it using a numerical optimization algorithm. To improve efficiency and quality, the rolling solver unit uses the strategy with the highest comprehensive risk assessment label score from the candidate control strategy set as the initial solution for optimization iterations. This is because the superior strategy evaluated by the parallel simulation module has a command trajectory close to the optimal region, and using this as a starting point can significantly reduce the convergence time of the optimization algorithm. At the beginning of each control sampling period, the unit re-solves the optimization problem in the finite future time domain, which may cover tens of seconds in the future. The core of the solution is to find an optimal sequence of fuel and air volume commands that minimizes the objective function value from the current moment to the end of the future time domain, while satisfying all safety constraints throughout. However, following the principles of model predictive control, the rolling solver unit does not output the entire future command sequence; instead, it only uses the first command value corresponding to the current moment as the final control command for the current sampling period and outputs it to the feedforward compensation module. At the next sampling cycle, the unit will reconstruct and solve a new optimization problem based on the latest measurement data, updated state estimates, and a new set of candidate strategies, thus "rolling" forward. This mechanism enables the system to continuously absorb the latest feedback information and correct control commands in real time, thereby addressing the challenges posed by dynamic changes in the boiler, fuel disturbances, and uncertainties in load commands.

[0047] Specifically, the boiler steam production optimization device based on dynamic matching of electrical load according to the present invention includes a feedforward compensation module comprising an instruction receiving unit, a relationship acquisition unit, a decoupling calculation unit, and a compensation output unit. The instruction receiving unit receives the fuel air delivery instruction sequence output by the rolling optimization module; The relationship acquisition unit receives the dynamic coupling matrix output by the twin construction module; The decoupling calculation unit receives the fuel air supply command sequence output by the command receiving unit and the dynamic coupling matrix output by the relationship acquisition unit, calculates the coupling gain of fuel command changes on furnace pressure based on the dynamic coupling matrix, and generates the corresponding induced draft command feedforward compensation amount. The compensation output unit receives the induced draft command feedforward compensation amount output by the decoupling calculation unit, superimposes the compensation amount onto the fuel supply air command sequence, and outputs the decoupled fuel, supply air, and induced draft coordinated command to the execution intervention module.

[0048] The feedforward compensation module receives the fuel and air supply command sequence from the rolling optimization module and the dynamic coupling matrix from the twin construction module, aiming to dynamically decouple and compensate the command sequence. The command receiving unit, acting as the module entry point, continuously receives the fuel and air supply command sequence output by the rolling optimization module in each control cycle. This sequence includes the setpoint changes in fuel quantity, primary air volume, and secondary air volume over a future period. Simultaneously, the relationship acquisition unit obtains the dynamic coupling matrix from the twin construction module, representing the dynamic interaction relationships between multiple variables at the current boiler operating point. This matrix is ​​a mathematical description that quantifies how changes in control inputs such as fuel and air supply commands affect and couple to key controlled outputs such as steam output and furnace pressure. The decoupling calculation unit synchronously receives the aforementioned command sequence and dynamic coupling matrix, performing calculations based on the specific coupling gain values ​​provided in the matrix. For example, when the received fuel supply command sequence indicates a need to increase fuel supply, the decoupling calculation unit calculates the precise induced draft fan guide vane opening compensation amount required to offset this coupling effect based on the parameter "coupling gain of fuel quantity change on furnace pressure" in the dynamic coupling matrix, thereby generating the corresponding induced draft command feedforward compensation amount. The compensation output unit finally receives the induced draft command feedforward compensation amount output by the decoupling calculation unit and algebraically superimposes it with the original fuel supply command sequence. After superposition, the output is the decoupled fuel, supply, and induced draft coordinated command. This compensation mechanism ensures that the induced draft command is pre-activated before the fuel command change takes effect, effectively offsetting the expected interference of fuel change on furnace pressure, achieving dynamic decoupling control among multiple variables, and finally sending the coordinated command to the execution intervention module to drive the field equipment.

[0049] Specifically, the boiler steam production optimization device based on dynamic matching of electrical load according to the present invention includes an execution intervention module comprising an instruction distribution unit, a safety monitoring unit, a fuzzy inference unit, a threshold comparison unit, and an intervention execution unit. The instruction distribution unit receives the compensated instruction sequence output by the feedforward compensation module, converts the instruction sequence into fieldbus protocol messages, and sends them to the fuel feeder frequency converter and the blower actuator. The safety monitoring unit receives the safety monitoring signal output by the status perception module and reads the water-cooled wall temperature, the tube wall temperature difference change rate and the furnace pressure oscillation mode signal in parallel. The fuzzy inference unit receives the monitoring signal output by the security monitoring unit, inputs the monitoring signal into the fuzzy logic system, performs inference through the rule base, and outputs a quantified risk value. The threshold comparison unit receives the quantized risk value output by the fuzzy inference unit and the system state estimate provided by the twin construction module, and compares the risk value with the safety threshold in the system state estimate in real time. The intervention execution unit receives the comparison result output by the threshold comparison unit. When the risk value exceeds the intervention threshold, it directly sends a control mode switching command and a gradient limiting signal to the execution mechanism.

[0050] The execution intervention module is the terminal link in the device responsible for the final execution of instructions and real-time monitoring of operational safety. It adopts an architecture design with parallel main control loop and safety protection loop. The instruction distribution unit receives the compensated instruction sequence from the feedforward compensation module, which has undergone dynamic decoupling optimization. The core function of this unit is to convert digital instructions representing the setpoints of fuel quantity, primary air volume, secondary air volume, and induced draft volume into physical signals that can be recognized and executed by the field industrial control system. Specifically, the unit encapsulates the instruction sequence into a message format of a standard fieldbus protocol (such as Profibus-DP or ModbusTCP) and sends it in real time through the industrial network to the variable frequency speed control device of the fuel feeder, the electric actuator of the blower, and the regulating damper controller of the induced draft fan. This process realizes the accurate and reliable transmission of optimization decision instructions from the digital space to the physical actuators.

[0051] Meanwhile, the safety monitoring unit operates independently and in parallel, forming a rapid safety barrier unaffected by delays in the main control loop. This unit directly receives raw, unprocessed safety monitoring signals from the state-aware module, reading multiple key safety parameters in parallel with the highest priority and fastest sampling rate. These parameters include the tube wall temperature at specific monitoring points on the boiler's membrane water-cooled walls, the real-time temperature difference and rate of change between adjacent tube banks, and the oscillation frequency and amplitude mode obtained after performing a fast Fourier transform analysis on the furnace pressure signal. The signals directly and rapidly reflect the thermal stress state of the boiler's heating surfaces and the pressure stability of the combustion chamber, serving as the core basis for assessing immediate equipment risks.

[0052] The fuzzy inference unit receives a set of pre-processed monitoring signals from the safety monitoring unit. This unit has a built-in fuzzy logic system and rule base based on expert experience. The system converts the input continuous signals (such as "high wall temperature," "rapid temperature rise rate," and "moderate pressure oscillation amplitude") into fuzzy linguistic variables using a membership function, and then applies predefined "IF-THEN" type inference rules for calculation. For example, a typical rule might be "If the temperature at a certain point on the water-cooled wall falls into the 'high' category, and the rate of temperature change at that point is 'positive and rapid,' then the output risk level is 'high.'" Through comprehensive inference and defuzzification calculation using multiple rules from the rule base, a quantitative value representing the current overall safety risk—the quantified risk value—is finally output.

[0053] The threshold comparison unit is responsible for real-time risk assessment. It receives the quantified risk value output by the fuzzy inference unit and extracts the safety threshold dynamically calculated by the digital twin model based on the current operating conditions from the system state estimate provided by the twin construction module. This safety threshold is not a fixed value but a dynamic boundary that adapts to factors such as boiler load, fuel characteristics, and equipment status. The threshold comparison unit continuously compares the real-time calculated quantified risk value with this dynamic safety threshold at the millisecond level.

[0054] The intervention execution unit is the final executor of the safety protection logic. The unit continuously monitors the comparison results output by the threshold comparison unit. Once the quantified risk value exceeds the pre-set dynamic intervention threshold, it means the system determines that the current risk exceeds the range that the main control loop can safely handle, and the intervention execution unit immediately activates the hardware-level protection logic. This unit bypasses the normal main control loop and directly sends the highest priority control mode switching command to field actuators such as the fuel feeder and blower, forcibly switching equipment control to a preset, absolutely conservative safe operating mode. Simultaneously, the unit applies strict gradient change rate limits (gradient amplitude limits) to all control commands being issued or about to be issued, physically prohibiting any large value or rapid action by the actuators. This provides a direct, fast, and reliable safety hardware barrier for boiler equipment, especially the vulnerable membrane water-cooled walls, preventing equipment damage risks that may be induced by excessively rapid changes in control commands at the execution level.

[0055] Specifically, in the boiler steam production optimization device based on dynamic matching of electrical load described in this invention, the confidence evaluation unit of the parallel simulation module feeds back the robustness test results obtained during the deduction process to the twin construction module, and the twin construction module updates the thermal inertia time constant distribution parameters of the mechanism model using the robustness test results; the safety intervention event data recorded by the execution intervention module is fed back to the oscillation suppression strategy network training dataset of the parallel simulation module, and the oscillation suppression strategy network adjusts the network weight parameters and exploration strategy according to the safety intervention event data.

[0056] The feedback mechanism in this invention forms the core of the system's continuous self-optimization and evolution. This mechanism improves the prediction accuracy of the digital twin model and the decision-making security of the oscillation suppression strategy network through two parallel data feedback paths.

[0057] The first feedback path focuses on the online correction of model parameters. After completing multiple rounds of virtual simulation, the confidence assessment unit of the parallel simulation module packages the stability boundary results obtained from robustness tests, such as the critical perturbation amplitude for maintaining combustion stability under a specific air-coal ratio, into structured data and feeds it back to the twin construction module. The mechanism model identification unit of the twin construction module receives the boundary data from the virtual environment test and treats it as a "stress test" result of the model's predictive ability. The unit compares the measured stability boundary conditions with the predicted boundary of the mechanism model under the same virtual perturbation. When it finds that the stability boundary predicted by the model is too optimistic or conservative, the unit uses the difference information to adjust the key parameters in the mechanism model that characterize the dynamic response speed of the boiler, namely the thermal inertia time constant distribution parameter. For example, if robustness tests show that the critical point of system instability under a certain disturbance occurs earlier than the model predicts, the unit will increase the thermal inertia time constant of the corresponding working area in the model accordingly, so that the updated digital twin model can more realistically predict similar risks in the future, thereby providing a more reliable safety constraint boundary for the rolling optimization module.

[0058] The second feedback path is dedicated to the continuous learning of the strategy network. During operation, the intervention execution module meticulously records comprehensive data on each event triggering a safety intervention. This data includes not only direct trigger signals such as the water-cooled wall temperature and furnace pressure oscillation modes at the moment of intervention, but also fuel commands, air supply command sequences, and critical boiler state responses over a period before and after the intervention. Real-world operational event data tagged with "high-risk" is fed back to the parallel simulation module in real time and integrated into the dedicated training dataset of the oscillation suppression strategy network. In the next training cycle of the strategy network, intervention cases from the real physical world become crucial training samples. The training algorithm forces the strategy network to deeply analyze the command trajectories and state evolutions that lead to intervention in the actual system, adjusting its internal weight parameters. This allows it to proactively avoid operational paths that might lead to similar high-risk states when exploring control strategies in the virtual environment. This mechanism enables the oscillation suppression strategy network to learn from real-world safety events, gradually shifting its exploration behavior from simply minimizing pressure oscillations to proactively avoiding safety boundaries that might trigger hardware protection while meeting performance objectives, thus achieving a progressive improvement in strategy security.

[0059] Specifically, in the boiler steam production optimization device based on dynamic matching of electrical load described in this invention, the rolling solution unit of the rolling optimization module outputs a sequence of instructions to the feedforward compensation module, which decouples the sequence of instructions. The instruction distribution unit of the execution intervention module sends the decoupled sequence of instructions to the field actuator. The state perception module collects the actual response data of the boiler and generates a new state matrix. The digital twin construction module receives the new state matrix and updates the model parameters. The updated digital twin model provides a new system state estimate and safe operation envelope constraint for the rolling optimization in the next sampling period.

[0060] After completing the calculation in each control cycle, the rolling solution unit of the rolling optimization module outputs a fuel and air supply command sequence to the feedforward compensation module. This sequence includes the optimal setpoint planning for fuel quantity and air supply within a finite time domain. The feedforward compensation module then performs a crucial decoupling process on this command sequence. The decoupling calculation unit calls the dynamic coupling matrix obtained from the twin construction module to calculate the corresponding induced draft command feedforward compensation amount for changes in the command sequence, especially in fuel quantity, and adds the compensation amount to the original command. The purpose of this step is to preemptively offset the dynamic coupling interference caused by fuel adjustment on furnace pressure before the command is executed, thereby outputting a set of decoupled fuel, air supply, and induced draft coordinated commands.

[0061] Upon receiving this set of coordination instructions, the instruction distribution unit of the intervention module immediately converts them into standard protocol messages recognizable by the fieldbus network, such as PROFIBUS-DP or ModbusTCP format, and distributes them in real time through the industrial network to the actuators of the fuel feeder's frequency converter, forced draft fan, and induced draft fan. The field equipment then operates according to the instructions, changing the fuel supply rate and airflow, thereby actually affecting the boiler's combustion process and steam output.

[0062] Meanwhile, the state perception module, which constitutes the system feedback loop, begins continuous operation. Its data acquisition unit synchronously collects actual response data of the boiler in terms of main steam pressure, flow rate, and superheater wall temperature at millisecond intervals; the image acquisition unit continues to acquire real-time furnace flame images. The preprocessing unit rapidly processes the newly acquired multi-source data reflecting the boiler's actual response to the control commands of the previous cycle, including time-stamp alignment, outlier removal, and standardization, ultimately generating a new multi-dimensional state matrix representing the latest operating state of the boiler.

[0063] This new multidimensional state matrix is ​​sent to the twin construction module in real time. The mechanistic model identification unit and visual feature extraction unit of the twin construction module use the latest field data to drive the model update process again. For example, the mechanistic model identification unit will fine-tune parameters such as the distribution of thermal inertia time constant, which characterizes the dynamic characteristics of the boiler, in the digital twin model based on the latest process variable response and using a recursive least squares algorithm with a forgetting factor. The model fusion unit may make minor corrections to the combustion model based on the latest flame image features. Through this process, the digital twin model is continuously updated, enabling it to track the characteristic drift of the boiler caused by coking, equipment wear, or slow changes in fuel characteristics, and maintain high-fidelity synchronization between the virtual model and the physical boiler.

[0064] The updated digital twin model immediately provides a new and more accurate basis for the rolling optimization calculations in the next sampling cycle. On one hand, the derived information output unit extracts the latest system state estimates from the updated model, such as estimates of the current combustion efficiency and heat storage status in the furnace. These estimates reflect the true internal state of the system better than relying solely on sensor readings. On the other hand, the model also provides updated dynamic safe operating envelopes, such as recalculating the safe upper and lower limits of the main steam pressure and the air-coal ratio boundary to ensure stable combustion based on the latest model parameters and states. When the rolling optimization module performs optimization in the next cycle, it constructs a new optimization problem based on the updated system state estimates and safe operating envelope constraints, enabling its decisions to adapt to the latest dynamic characteristics of the boiler. This iterative process forms a closed-loop adaptive optimization flow of "command output, execution, response acquisition, model update, and optimization re-decision," allowing the device to dynamically match changes in electrical load while continuously ensuring the stability and safety of boiler operation.

[0065] Specifically, the boiler steam production optimization device based on dynamic matching of electrical load described in this invention comprises: a state perception module that collects real-time data to drive a digital twin construction module to generate a digital twin model; a parallel simulation module that performs virtual simulations on the digital twin model to generate a set of candidate control strategies; a rolling optimization module that performs rolling optimization calculations based on the candidate control strategy set and real-time state data; a feedforward compensation module that decouples and compensates for optimization instructions; and an execution intervention module that enables safe execution of instructions and risk intervention.

[0066] The state awareness module, serving as the system's data starting point, continuously collects electrical load command curves from the power grid dispatching system, sensor network data from the boiler itself, real-time parameters from the fuel supply system, and relevant operating condition records from the historical database. After time-stamp alignment, outlier removal, and standardization by the preprocessing unit, the multi-source heterogeneous data is fused to generate a multi-dimensional state matrix with strictly synchronized timestamps. This matrix comprehensively represents the current instantaneous state of the boiler and external load commands, providing a unified, high-quality input data source for subsequent processing chains.

[0067] The digital twin construction module receives the multi-dimensional state matrix output in real time from the state perception module and uses it as the core to drive the construction and updating of a high-fidelity digital twin model. The module dynamically corrects the parameters of the boiler nonlinear model based on physical laws, especially the critical thermal inertia time constant, through the mechanism model identification unit; simultaneously, it extracts combustion stability feature vectors from flame image sequences through the visual feature extraction unit. The model fusion unit weighted and fused these two types of information to output a high-fidelity digital twin model that reflects both the physical mechanism and the current actual combustion state. The derived information output unit then extracts key derived information such as system state estimates and dynamic coupling matrices from this model and sends them, along with the model itself, to downstream modules.

[0068] The parallel simulation module utilizes the received digital twin model to perform forward-looking multi-strategy extrapolation in a virtual environment. The module inputs future electrical load command sequences into the model and evaluates the performance of different control strategies in tracking load, suppressing oscillations, and responding to disturbances through parallel simulations of baseline extrapolation, optimization exploration, and robustness testing. The confidence assessment unit synthesizes all simulation results, calculates tracking error, peak oscillation value, and robustness score for each strategy, and finally outputs a set of candidate control strategies with clear risk assessment labels. This process completes the screening and pre-evaluation of a large number of strategies in the digital space before the actual execution of the commands.

[0069] The rolling optimization module is the core component that integrates real-time information and simulation results to make the final decision. Its data integration unit receives candidate strategy sets from the parallel simulation module, real-time process variable deviation signals from the state awareness module, and the latest system state estimates from the digital twin construction module. Based on this information, the problem construction unit constructs an optimization problem aimed at minimizing steam production tracking error, combustion oscillation intensity, and actuator actions, using the dynamic safe operation envelope provided by the digital twin model as constraints. In each control cycle, the rolling solution unit starts with the highest-rated candidate strategy and solves for the optimal command sequence within the finite time domain, but only outputs the fuel supply command value at the current moment, achieving a rolling forward optimization decision.

[0070] The feedforward compensation module specifically addresses the dynamic coupling problem in the command sequence. The module receives the fuel and air supply command sequence output by the rolling optimization module and simultaneously obtains a dynamic coupling matrix representing the interaction relationships between variables under the current operating condition from the twin construction module. The decoupling calculation unit calculates the required induced draft feedforward compensation amount to offset the expected interference caused by changes in fuel commands on the furnace pressure based on the coupling gain parameters in the matrix. The compensation output unit superimposes this compensation amount onto the original commands, outputting a set of decoupled fuel, air supply, and induced draft coordination commands, thereby proactively eliminating mutual interference between multiple variables before command execution.

[0071] The intervention module bears the dual responsibility of final instruction issuance and real-time monitoring of operational safety. The instruction distribution unit converts the coordination instructions from the feedforward compensation module into fieldbus protocol messages and distributes them to actuators such as the feeder and blower. Simultaneously, an independent and parallel safety protection loop continues to operate. The safety monitoring unit directly reads key safety signals such as water-cooled wall temperature, temperature difference rate of change, and furnace pressure oscillation mode. The fuzzy inference unit applies an expert rule base to assess the risks of these signals and outputs quantified risk values. The threshold comparison unit compares the risk values ​​with dynamic safety thresholds provided by the digital twin model in real time. Once a risk exceeds the intervention threshold, the intervention execution unit immediately activates the highest-priority hardware protection logic, directly sending control mode switching instructions to the actuators and applying gradient limiting, forcing the system into a safe and conservative operating state, providing a fast and reliable safety hardware barrier for the main control loop.

[0072] The device forms a tight closed-loop connection through data flow between the aforementioned modules. Real-time data from the state-aware module drives model updates and strategy simulations. The rolling optimization module integrates simulation results with real-time status to make decisions. After decoupling and compensation, the decision instructions are safely executed by the intervention module. The actual response data generated by the boiler after execution is then collected by the state-aware module, thus initiating a new optimization cycle. Furthermore, the stability boundaries discovered by the simulation module and the safety events recorded by the intervention module serve as feedback data for continuous optimization of model parameters and the strategy network, enabling the system to possess adaptive and self-learning capabilities. All modules work collaboratively through feedforward, feedback, and parallel processing mechanisms to jointly constitute a complete intelligent control system capable of dynamically matching changes in electrical load while ensuring the safe and stable operation of the boiler.

[0073] In actual power plant operation, especially in scenarios involving grid frequency regulation, electrical load commands often fluctuate frequently and significantly within a short period. However, boiler systems, due to their large heat capacity, exhibit significant thermal inertia. This mismatch in dynamic response speed is the core issue inducing combustion oscillations and threatening equipment safety. The device described in this invention addresses this problem through the following operational methods.

[0074] The state awareness module continuously operates as the system's data entry point. Its data acquisition unit synchronously collects electrical load command curves from the power grid dispatch system and sensor network signals from the boiler itself, including main steam pressure and flow rate, and wall temperatures of various superheaters, at millisecond intervals, forming a process variable data stream. Simultaneously, the image acquisition unit continuously captures a sequence of images showing the flame distribution throughout the furnace using a high-temperature industrial endoscope deployed in the burner observation port. The composition analysis unit receives real-time spectral data from an online coal quality analyzer, analyzing key parameters such as volatile matter and fixed carbon in the coal fed into the furnace. The historical query unit retrieves waveform fragments of past combustion oscillation events similar to the current load and coal quality from the database. The preprocessing unit aggregates the aforementioned multi-source heterogeneous data, performs rigorous time-stamp alignment, removes outliers, and standardizes the data, ultimately generating a timestamped, consistent-quality, multi-dimensional state matrix, providing a unified digital twin input for all subsequent modules.

[0075] After receiving the multidimensional state matrix, the twin construction module drives two technical paths in parallel to build a high-fidelity virtual model. The mechanism model identification unit extracts process variables from the matrix and uses a recursive least squares algorithm with a forgetting factor to dynamically correct the key thermal inertia time constant distribution parameters in the boiler nonlinear mechanism model, enabling the mechanism model to track the characteristic drift of the boiler caused by coking and wear. The visual feature extraction unit processes the flame image sequence and extracts feature vectors representing combustion stability, such as brightness center and pulsation frequency, through a pre-trained deep convolutional neural network. The model fusion unit injects the feature vectors into the updated mechanism model in a weighted manner, corrects its combustion reaction rate term, and outputs a high-fidelity digital twin model that integrates physical laws and real-time combustion visual state. The derived information output unit then extracts key information such as system state estimates and dynamic coupling matrices from this model and distributes them to downstream modules.

[0076] The parallel simulation module uses this digital twin model to create a safe virtual testbed. Upon receiving a sequence of future electrical load commands, the module initiates multi-threaded simulations. The baseline simulation unit performs open-loop simulations on the model, predicting the baseline response of steam production and furnace pressure under no-intervention conditions, outlining potential risks. The optimization exploration unit runs an oscillation suppression strategy network trained with deep deterministic strategy gradients within the model. This network, like a virtual intelligent controller, actively explores optimal control paths that can smooth pressure fluctuations by adjusting virtual fuel and air supply commands. The robustness testing unit simultaneously injects simulated disturbances such as step changes in coal quality parameters into the model to test the stability boundaries of different control strategies under abnormal operating conditions. Finally, the confidence assessment unit integrates the results of these three simulations, calculates scores for each candidate strategy in terms of load tracking accuracy, pressure oscillation suppression, and disturbance rejection capability, and generates a set of candidate control strategies with clear risk assessment labels.

[0077] The rolling optimization module is the core of the system, integrating simulation pre-playback with real-time status to make the final decision. Its data integration unit receives the candidate strategy set, real-time process variable deviation signals from the state awareness module, and the latest system state estimates from the digital twin construction module. The problem construction unit defines a rolling time-domain optimization problem based on this information: the objective function aims to minimize steam production tracking error, combustion oscillation prediction intensity, and actuator action amplitude; constraints are defined by the dynamic safe operation envelope provided by the digital twin model, such as the upper limit of main steam pressure, the wall temperature safety threshold, and the minimum air-fuel ratio. The rolling solution unit uses the candidate strategy with the highest score as the initial solution and solves for the optimal fuel-air delivery command sequence within the future finite time domain in each control cycle, but only outputs the command value at the current moment, achieving a rolling forward closed-loop decision-making process.

[0078] The feedforward compensation module is specifically designed to handle the complex dynamic coupling problem among multiple variables in a boiler. The module receives the instruction sequence output from the rolling optimization module and simultaneously obtains a dynamic coupling matrix describing the interaction relationships between variables under the current operating conditions from the twin construction module. Based on the coupling gain in the matrix, the decoupling calculation unit accurately calculates the induced draft feedforward compensation amount required to offset the expected disturbance to furnace pressure caused by changes in fuel commands. The compensation output unit superimposes this compensation amount onto the original commands, thereby outputting a set of dynamically decoupled fuel, forced draft, and induced draft coordination commands, suppressing mutual interference between multiple variables at the source.

[0079] The intervention module bears the crucial responsibility of ensuring the safe execution of instructions. The instruction distribution unit converts coordination instructions into fieldbus protocol messages and distributes them to actuators such as the fuel feeder and blower. Simultaneously, an independent, parallel safety protection loop operates at high speed. The safety monitoring unit directly reads native safety signals such as water-cooled wall temperature, temperature difference rate of change, and furnace pressure oscillation mode. The fuzzy inference unit applies an expert rule base to comprehensively evaluate the signals and outputs a quantified risk value. The threshold comparison unit compares this risk value with the dynamic safety threshold provided by the digital twin model in real time. Once the risk value exceeds the intervention threshold, it means the main control loop can no longer safely handle the current risk. The intervention execution unit immediately activates the highest-priority hardware protection logic, directly sending a control mode switching instruction to the actuators and applying a gradient rate of change limit, forcing the system into an absolutely conservative operating state, providing a fast and reliable final safety barrier for the boiler membrane water-cooled wall.

[0080] The system forms a tight closed loop and continuously learns through data flow. The state perception module collects actual response data of the boiler to control commands, driving the twin construction module to update model parameters, keeping the digital twin synchronized with the physical boiler, and thus providing a more accurate predictive basis for the next cycle of rolling optimization. In addition, the stability boundary information discovered by the parallel simulation module in virtual simulation is fed back to correct the thermal inertia parameters of the twin model; the real safety intervention event data recorded by the execution intervention module is fed back to retrain the oscillation suppression strategy network, enabling it to learn to proactively avoid high-risk operations in virtual exploration. This complete intelligent closed loop of "perception-modeling-simulation-optimization-compensation-execution-learning" enables the device to dynamically match rapidly changing electrical loads while proactively suppressing combustion oscillations, ensuring the long-term safe and stable operation of the boiler.

[0081] In the mechanism model identification unit of the twin construction module, a recursive least squares algorithm with a forgetting factor is used to dynamically update the boiler mechanism model parameters. The core of this algorithm is iterative solution, and its calculation process involves the following formulas and data processing paths: First, the parameter vector of the boiler mechanism model to be identified is defined as θ(k). Here, k represents the current discrete-time sampling moment. The parameter vector θ(k) includes key model parameters such as the thermal inertia time constant distribution parameter that needs to be adjusted online.

[0082] At each time step, the algorithm receives preprocessed process variable data from the state-aware module. The data is constructed as a regression vector φ(k), which includes a combination of measured values ​​of process variables (such as pressure, flow rate, and fuel quantity) from the current and historical time steps. Its specific form depends on the structure of the boiler mechanism model. Simultaneously, the algorithm receives the corresponding actual measured values ​​y(k) from the model output.

[0083] The parameter update formula for the recursive least squares algorithm with a forgetting factor is as follows: ; In the formula: This represents the algorithm gain vector at time k.

[0084] Let k represent the parameter estimation error covariance matrix at time k, with initial values... It is usually set as a large diagonal matrix.

[0085] This represents the forgetting factor, which ranges from (0,1). Forgetting factor Its function is to configure exponentially decaying weights for historical data. When system characteristics change slowly, taking a value less than 1 can make the algorithm continuously track parameter changes.

[0086] Let k represent the regression vector at time k.

[0087] Represents the regression vector The transpose of .

[0088] This represents the actual measured value output by the model at time k.

[0089] This represents the parameter vector estimate after time k.

[0090] This represents the updated parameter vector estimate at time k-1 (i.e., the previous time).

[0091] The algorithm's data processing path is as follows: In each sampling period, the unit receives new process variable data, constructs a regression vector φ(k), and obtains the measured value y(k). Subsequently, the algorithm sequentially calculates the gain vector K(k), utilizing the information (i.e., the difference between the measured value and the model's predicted value). The parameter estimates are corrected to obtain the updated parameter vector θ(k), and the covariance matrix P(k) is updated simultaneously. The final output θ(k) is the core parameter of the updated mechanism model, which is used to reconstruct the high-fidelity digital twin model.

[0092] In the visual feature extraction unit of the twin construction module, a pre-trained deep convolutional neural network is used to process the flame image sequence. The computation process of the deep convolutional neural network can be summarized as multi-layer feature transformation, and its core operations include convolution, non-linear activation, and pooling. Although the specific structure of the network (number of layers, filter size, etc.) are fixed parameters after training, its processing of single-frame images... The data flow from which feature vectors are extracted can be described as follows: Let the input image be , where t represents the time index in the image sequence. The network consists of L layers, each layer having l pairs of input feature maps. Perform calculations and output new feature maps. For a convolutional layer, the calculation can be expressed as: ; In this formula: This represents a single-frame flame image acquired at time index t, which is the input to the network.

[0093] This represents the time index in the sequence of flame images, used to identify the order of the images.

[0094] This represents the input feature map of the l-th layer. For the first layer of the network (i.e., l=1), its input... .

[0095] This represents the output feature map of the l-th layer.

[0096] This represents the set of weight parameters for the l-th layer convolutional filter, which are fixed values ​​learned during the pre-training phase.

[0097] This represents the convolution operation.

[0098] This represents the bias parameter vector of the l-th layer, which is also obtained during pre-training.

[0099] This represents a nonlinear activation function, such as the ReLU (Rectified Unit) function.

[0100] The network typically ends with a fully connected layer, which maps the final feature map into a fixed-dimensional feature vector. eigenvectors Each dimension corresponds to the abstract visual features learned by the network, such as features related to the center of flame brightness, pulsation, and fullness. Each unit represents each frame in the image sequence. Performing the forward propagation calculations described above yields the corresponding feature vector sequence. This sequence is output for subsequent model fusion.

[0101] In the optimization exploration unit of the parallel simulation module, the oscillation suppression policy network trained with deep deterministic policy gradients runs in a virtual environment constructed by a digital twin model. The learning objective of the policy network π is to optimize the policy parameters. Its update depends on the policy gradient. The core update formula is based on the deterministic policy gradient theorem and can be expressed as the policy parameters... Gradient ascent: ; In this formula: Representation Strategy In parameters The expected cumulative reward (return) is the goal that needs to be maximized.

[0102] Describe the objective function Regarding strategy parameters Gradient estimation is used to guide The direction of updates.

[0103] This indicates the number of samples (state-action pairs) sampled from the experience replay cache for this gradient update.

[0104] This represents the state in the i-th sample, derived from the simulation environment of the digital twin model.

[0105] This indicates an action, specifically a control variable such as a virtual fuel command or an air supply command.

[0106] This represents a policy network whose function is to generate actions based on the state.

[0107] Indicates the policy network in parameters Given a state Deterministic actions output in real time.

[0108] The action value function (Critic network) has the following parameters: Used to evaluate in state Next action Its long-term value. Its output is a scalar value.

[0109] Action value function Network parameters.

[0110] Action value function Regarding the action The gradient in the action Calculation at point.

[0111] This indicates that the policy network outputs parameters about itself. The gradient in state Calculation at point.

[0112] The data processing path is as follows: In the virtual simulation, the policy network processes data according to the current state. Generate Actions It is applied to digital twin models and rewards are given. and new status This experience Store the data in the replay cache. During training, sample a batch of experience data from the cache, calculate the policy gradient using the gradient estimation formula described above, and update the policy network parameters using an optimizer (such as Adam). This causes the strategy to tend to output actions that yield higher cumulative rewards (i.e., better suppression of oscillations).

[0113] In the problem construction unit of the rolling optimization module, a rolling time-domain optimization problem needs to be constructed. The objective function of this problem is a combination of multiple performance metrics, and the constraints are defined by the digital twin model. The mathematical description of the optimization problem is as follows: Let the optimized time domain length be... The control time domain length is (generally At time k, the optimization problem aims to find the optimal sequence of control commands from time k to time k+N_c-1. .in, This represents the predicted value of the control commands (fuel, air supply) at time k+i.

[0114] objective function Typically designed as follows: ; In this objective function: This represents the value of the rolling time-domain optimization objective function that needs to be minimized at time k.

[0115] This indicates the length of the prediction time domain, i.e., how many future steps the optimization problem considers in the output.

[0116] This indicates the length of the control time domain, which is the length of the future control instruction sequence to be optimized.

[0117] This represents the control command sequence vector to be optimized starting from time k.

[0118] This represents the control command predicted at time k+i in the future.

[0119] This indicates a model based on a digital twin and a sequence of instructions. The system output (such as steam production and main steam pressure) predicted at time k in the future k+i time.

[0120] The reference trajectory at time k+i (such as the steam production setpoint) is derived from the electrical load command sequence.

[0121] This represents the weighted square norm of the output tracking error.

[0122] It is a positive definite or semi-positive definite weight matrix, used to assign different importance to the tracking accuracy of different output variables.

[0123] It indicates the amount of change in the control command.

[0124] This represents the weighted square norm of the rate of change of the control increment.

[0125] It is a positive definite weight matrix, used to penalize frequent or violent actions of the actuator, thereby increasing control stability.

[0126] This represents the weighting coefficient of the combustion oscillation intensity prediction term.

[0127] This represents the index function for predicting the intensity of combustion oscillations.

[0128] This represents the predicted system state sequence from time k+1 to k+N_p, generated by the digital twin model based on the control sequence. It is calculated based on the current state. (Function) Based on this predicted state sequence, the oscillation intensity index is calculated.

[0129] The constraints of the optimization problem include: System dynamic constraints: .in It is the discrete-time form of the digital twin model, describing the state transition relationships.

[0130] Input constraints: .in and It is the physical limit of the actuator.

[0131] State constraints (safe operating envelope): .in The dynamic safety operating boundaries provided by the digital twin model are defined, such as the upper limit of the main steam pressure. Upper limit of superheater wall temperature and the minimum air-coal ratio required to maintain stable combustion. The boundary values ​​may change dynamically with the operating conditions.

[0132] The rolling solution unit of the rolling optimization module solves the above problem with the objective function in each sampling period k. The optimal instruction sequence is obtained by solving the optimization problem with constraints. However, only the first element in the sequence... As the actual output, it is sent to the feedforward compensation module. In the next sampling period k+1, based on the new state measurement or estimate, a new optimization problem is solved again to achieve "rolling optimization and feedback correction".

[0133] Embodiment 1 of the present invention: In the scenario where a power plant participates in frequency regulation of automatic generation control (AGC) of the power grid; In a typical scenario where the power grid frequency fluctuates and the dispatch system issues rapid load increase / decrease commands, such as requiring the unit to increase output by 50 MW within 2 minutes, the electrical load command curve exhibits a steep incline. At this time, the data acquisition unit of the state perception module synchronously captures this command curve at a 100-millisecond cycle, while simultaneously reading the real-time values ​​of boiler main steam pressure, flow rate, and high-temperature superheater wall temperature. The image acquisition unit observes regional flickering of the furnace flame brightness through the endoscope in the observation port. The composition analysis unit provides feedback on the received basis lower heating value of the current coal fed into the furnace. The historical query unit finds several event records from the database showing pressure oscillations under similar load increase rates and coal qualities. The preprocessing unit aligns and cleans the data, generating a multi-dimensional state matrix including time, load command, process variables, flame image features, coal quality data, and historical risk pattern labels.

[0134] The twin construction module receives this matrix. The mechanism model identification unit, using the latest process variable data and an algorithm with a forgetting factor, calculates that the critical thermal inertia time constant from combustion to steam production in the current operating condition is approximately 8% longer than the previous cycle. The visual feature extraction unit extracts the vector indicating increased pulsation frequency from the flame image sequence. The model fusion unit incorporates this feature vector into the mechanism model, correcting the reaction rate prediction for the unstable combustion region, and generating a digital twin reflecting the latest characteristics of the boiler: "slower load response and combustion instability." The derived information output unit extracts the updated system state estimate and dynamic coupling matrix from it.

[0135] The parallel simulation module immediately took action. The baseline extrapolation unit input the load increase command for the next two minutes into the digital twin. Open-loop simulation results showed that if the current control strategy was followed, the main steam pressure would face a significant risk of dropping in the medium term. The optimization exploration unit activated the oscillation suppression strategy network and simulated the exploration within the digital twin, discovering a virtual command trajectory of "small initial fuel increase with simultaneous increased air supply, and accelerated coal feeding in the medium term," which could better stabilize the pressure. The robustness testing unit injected disturbances related to coal calorific value fluctuations into the model, verifying that the strategy could maintain stability even with slight deterioration in coal quality. After comprehensive evaluation, the confidence assessment unit marked the strategy as the "preferred strategy for high-risk scenarios" and added it to the candidate set.

[0136] The data integration unit of the rolling optimization module receives the candidate strategy, the deviation between the current steam pressure and the target value, and the "pressure drop risk" state estimate provided by the twin module. The problem construction unit then constructs an optimization problem based on this, aiming to rapidly track load commands while strictly controlling pressure fluctuations and wall temperature changes within the dynamic safety envelope calculated by the twin model. The rolling solution unit, starting with the candidate optimization strategy, quickly solves for the optimal fuel and air supply command sequence for the next tens of seconds and issues the first command.

[0137] The feedforward compensation module calculates the induced draft fan guide vane opening compensation command that needs to be pre-acted based on the fueling command output by the rolling optimization module and the coupling matrix provided by the twin module that "the current fueling has a significant impact on the furnace negative pressure", thereby achieving decoupling.

[0138] The instruction distribution unit of the intervention module issues coordinated fuel, air, and induced draft commands. Simultaneously, the safety monitoring unit detects a rapid increase in the temperature of a water-cooled wall section on the sidewall due to heat flux changes, and the fuzzy inference unit determines that the local overheating risk level has increased. Since the risk value has not yet exceeded the dynamic safety threshold provided by the latest digital twin model, the main control loop continues to operate. Through the aforementioned forward-looking simulation optimization and decoupling compensation, the device ensures a smooth transition of the main steam pressure during this rapid load increase, effectively suppressing drastic fluctuations in combustion pressure and avoiding the impact on the membrane water-cooled wall caused by pressure oscillation energy transfer.

[0139] Embodiment 2 of the present invention: Under the condition of co-firing biomass in a coal-fired power plant or encountering a significant change in coal quality; When boiler fuel is switched from pure coal to a high proportion of biomass, or when the quality of the coal entering the furnace undergoes an unpredictable and abrupt change, the fuel ignition characteristics and burnout time change significantly, easily inducing combustion instability. The composition analysis unit of the state sensing module is the first to detect a significant increase in fuel volatiles through spectral analysis. The image acquisition unit almost simultaneously observes a brightening of the flame color and a more dispersed shape. The historical query unit immediately retrieves historical alarm data under similar operating conditions with sudden changes in fuel characteristics.

[0140] The visual feature extraction unit of the digital twin construction module quickly extracts new feature vectors from the abruptly changed flame images, indicating an upward shift in the brightness center and a decrease in the degree of combustion. The mechanism model identification unit, based on the latest process variables, identifies changes in the distribution of the thermal inertia time constant required for the combustion reaction. The model fusion unit injects the new feature vectors, rapidly correcting the predictions regarding burnout rate and furnace temperature field in the digital twin model, enabling the virtual model to more quickly reflect the dynamic characteristics of the boiler after fuel mutations.

[0141] The parallel simulation module, utilizing the updated digital twin, immediately performs robustness testing on the current operating strategy. The simulation injects detected fuel composition step disturbances into the model, revealing that under the original control strategy, furnace pressure oscillations intensify. The optimization exploration unit then re-explores strategies based on the new model, quickly deriving a new virtual strategy that "appropriately increases the secondary air ratio and adjusts the burner angle" to stabilize the flame. A new set of candidate strategies is rapidly generated and transmitted.

[0142] When constructing the optimization problem, the rolling optimization module receives new candidate strategies and a safety operating envelope provided by the twin module that has tightened due to fuel mutations (e.g., a smaller allowable range of steam pressure fluctuations). The solved fuel-air supply command sequence automatically includes air-fuel ratio adjustments to adapt to the new fuel characteristics.

[0143] The feedforward compensation module recalculates the decoupling compensation amount based on the updated dynamic coupling matrix after fuel mutation, ensuring that interference between control commands can still be effectively canceled under the new combustion characteristics.

[0144] The safety monitoring unit of the intervention module continuously monitored the furnace pressure for low-frequency oscillations and the accelerated rate of change in local wall temperature. The fuzzy inference unit comprehensively judged the continuous accumulation of risk. Suddenly, the risk quantification value at a certain moment exceeded the dynamic safety threshold calculated by the digital twin model for the current abrupt change in operating conditions. The threshold comparison unit immediately triggered an alarm. The intervention execution unit responded within milliseconds, directly issuing the highest priority command to the actuator, forcibly switching the control system into the preset "conservative mode for burning special fuel," and significantly limiting the gradient amplitude of all adjustment commands, instantly suppressing drastic changes in operation, and gaining safe time for operators to intervene and adjust, and for the system to readjust, thereby protecting the boiler's heating surfaces.

Claims

1. A boiler steam production optimization device based on dynamic matching of electrical load, characterized in that, include: The status perception module collects electrical load commands, boiler sensor data, fuel parameters, and historical operating data, and outputs a multi-dimensional status matrix. The digital twin construction module connects to the state perception module, receives the multi-dimensional state matrix output by the state perception module, constructs a digital twin model of the boiler based on the multi-dimensional state matrix, and outputs the digital twin model, system state estimate, and dynamic coupling matrix. The parallel simulation module connects the twin construction module and the state perception module. It receives the digital twin model output by the twin construction module and the electrical load command sequence output by the state perception module, performs virtual simulation on the digital twin model, and outputs a set of candidate control strategies. The rolling optimization module connects the parallel simulation module, the state awareness module, and the twin construction module. It receives the candidate control strategy set output by the parallel simulation module, the real-time deviation signal output by the state awareness module, and the system state estimate output by the twin construction module. It calculates the fuel delivery command sequence through rolling optimization. The feedforward compensation module connects the rolling optimization module and the twin construction module. It receives the fuel air supply command sequence output by the rolling optimization module and the dynamic coupling matrix output by the twin construction module, performs decoupling compensation on the fuel air supply command sequence, and outputs the compensated command sequence. The execution intervention module connects the feedforward compensation module and the state perception module. It receives the compensated instruction sequence output by the feedforward compensation module and the safety monitoring signal output by the state perception module, distributes the compensated instruction sequence to the actuator, and performs safety intervention based on the safety monitoring signal.

2. The boiler steam production optimization device based on dynamic matching of electrical load according to claim 1, characterized in that, The state perception module includes a data acquisition unit, an image acquisition unit, a component analysis unit, a historical query unit, and a preprocessing unit; The data acquisition unit synchronously acquires the electrical load command curve of the power grid dispatching system, the main steam pressure and flow signal of the boiler, and the thermocouple signals of the wall temperature of each stage of superheater at millisecond intervals, and outputs process variable data. The image acquisition unit continuously acquires a sequence of images of the flame distribution throughout the furnace through the observation hole of the burner using a high-temperature industrial endoscope, and outputs image data. The component analysis unit receives real-time coal composition spectral data provided by the online coal quality analyzer and outputs coal quality data. The historical query unit retrieves waveform segments of historical combustion oscillation events that are similar to the current load, coal quality, and air distribution mode from the relational database and outputs historical waveform data. The preprocessing unit receives process variable data output by the data acquisition unit, image data output by the image acquisition unit, coal quality data output by the component analysis unit, and historical waveform data output by the historical query unit. It performs time-stamp alignment, outlier removal, and standardization on the various types of data collected, and generates a multi-dimensional state matrix with strictly synchronized timestamps, which is then output to the twin construction module.

3. The boiler steam production optimization device based on dynamic matching of electrical load according to claim 2, characterized in that, The twin construction module includes a mechanism model identification unit, a visual feature extraction unit, a model fusion unit, and a derived information output unit; The mechanism model identification unit receives the process variables in the multidimensional state matrix output by the state perception module, inputs the process variables into the recursive least squares algorithm with forgetting factor, dynamically adjusts the thermal inertia time constant distribution parameters in the nonlinear distributed parameter mechanism model of the boiler, and outputs the updated mechanism model. The visual feature extraction unit receives the flame image sequence from the multidimensional state matrix output by the state perception module, inputs the flame image sequence into a pre-trained deep convolutional neural network, extracts combustion stability feature vectors representing the center position of flame brightness, pulsation frequency, and fullness, and outputs the feature vectors. The model fusion unit receives the updated mechanism model output by the mechanism model identification unit and the feature vector output by the visual feature extraction unit, injects the feature vector into the updated mechanism model in a weighted summation manner, corrects the combustion reaction rate term, and outputs the fused digital twin model. The derived information output unit receives the fused digital twin model output by the model fusion unit, extracts real-time thermal inertia parameters, system state estimates, and dynamic coupling matrices from the fused digital twin model, and sends them to the parallel simulation module, rolling optimization module, and feedforward compensation module, respectively.

4. The boiler steam production optimization device based on dynamic matching of electrical load according to claim 3, characterized in that, The parallel simulation module includes a baseline extrapolation unit, an optimization exploration unit, a robust testing unit, and a confidence assessment unit; The baseline simulation unit receives the digital twin model output by the twin construction module and the electrical load command sequence output by the state perception module. It inputs the electrical load command sequence into the digital twin model to perform open-loop simulation, predicts the baseline response curves of steam production and furnace pressure, and outputs baseline response data. The optimization exploration unit receives the digital twin model output by the twin construction module, runs an oscillation suppression policy network trained with deep deterministic policy gradient in the digital twin model, explores the optimal path to suppress pressure fluctuations by adjusting the virtual fuel command, and outputs the optimized control command trajectory. The robustness testing unit receives the digital twin model output by the twin construction module, injects a step change disturbance of coal quality parameters into the digital twin model, tests the stability boundary of the candidate control strategy under different disturbance amplitudes, and outputs the robustness test results. The confidence assessment unit receives the baseline response data output by the baseline extrapolation unit, the optimized control command trajectory output by the optimization exploration unit, and the robustness test results output by the robustness test unit. It compares the triple extrapolation results, calculates the steam production tracking error, pressure oscillation peak value, and robustness score of each candidate strategy, and generates a set of candidate control strategies with risk assessment labels, which is then sent to the rolling optimization module.

5. The boiler steam production optimization device based on dynamic matching of electrical load according to claim 4, characterized in that, The rolling optimization module includes a data integration unit, a problem construction unit, and a rolling solution unit; The data integration unit receives candidate strategy sets from the parallel simulation module, real-time process variable deviation signals from the state awareness module, and system state estimates from the twin construction module, and outputs integrated data. The problem construction unit receives the integrated data output by the data integration unit, constructs an objective function using the sum of squared errors in steam production tracking, the predicted index of combustion oscillation intensity, and the penalty term for the rate of change of actuator action, defines the constraints using the dynamic safe operation envelope provided by the digital twin model, and outputs the optimization problem. The rolling solution unit receives the optimization problem output by the problem construction unit, takes the strategy with the highest evaluation in the candidate strategy set as the initial solution, solves the optimal fuel air delivery command sequence in the future finite time domain in each sampling period, and outputs only the command value at the current moment to the feedforward compensation module.

6. The boiler steam production optimization device based on dynamic matching of electrical load according to claim 5, characterized in that, The feedforward compensation module includes an instruction receiving unit, a relationship acquisition unit, a decoupling calculation unit, and a compensation output unit; The instruction receiving unit receives the fuel air delivery instruction sequence output by the rolling optimization module; The relationship acquisition unit receives the dynamic coupling matrix output by the twin construction module; The decoupling calculation unit receives the fuel air supply command sequence output by the command receiving unit and the dynamic coupling matrix output by the relationship acquisition unit, calculates the coupling gain of fuel command changes on furnace pressure based on the dynamic coupling matrix, and generates the corresponding induced draft command feedforward compensation amount. The compensation output unit receives the induced draft command feedforward compensation amount output by the decoupling calculation unit, superimposes the compensation amount onto the fuel supply air command sequence, and outputs the decoupled fuel, supply air, and induced draft coordinated command to the execution intervention module.

7. The boiler steam production optimization device based on dynamic matching of electrical load according to claim 6, characterized in that, The execution intervention module includes an instruction distribution unit, a security monitoring unit, a fuzzy inference unit, a threshold comparison unit, and an intervention execution unit; The instruction distribution unit receives the compensated instruction sequence output by the feedforward compensation module, converts the instruction sequence into fieldbus protocol messages, and sends them to the fuel feeder frequency converter and the blower actuator. The safety monitoring unit receives the safety monitoring signal output by the status perception module and reads the water-cooled wall temperature, the tube wall temperature difference change rate and the furnace pressure oscillation mode signal in parallel. The fuzzy inference unit receives the monitoring signal output by the security monitoring unit, inputs the monitoring signal into the fuzzy logic system, performs inference through the rule base, and outputs a quantified risk value. The threshold comparison unit receives the quantized risk value output by the fuzzy inference unit and the system state estimate provided by the twin construction module, and compares the risk value with the safety threshold in the system state estimate in real time. The intervention execution unit receives the comparison result output by the threshold comparison unit. When the risk value exceeds the intervention threshold, it directly sends a control mode switching command and a gradient limiting signal to the execution mechanism.

8. The boiler steam production optimization device based on dynamic matching of electrical load according to claim 7, characterized in that, The confidence evaluation unit of the parallel simulation module feeds back the robustness test results obtained during the inference process to the twin construction module. The twin construction module uses the robustness test results to update the thermal inertia time constant distribution parameters of the mechanism model. The safety intervention event data recorded by the execution intervention module is fed back to the oscillation suppression strategy network training dataset of the parallel simulation module. The oscillation suppression strategy network adjusts the network weight parameters and exploration strategy according to the safety intervention event data.

9. The boiler steam production optimization device based on dynamic matching of electrical load according to claim 8, characterized in that, The rolling optimization module's rolling solution unit outputs a sequence of instructions to the feedforward compensation module, which decouples the instruction sequence. The execution intervention module's instruction distribution unit sends the decoupled instruction sequence to the field actuators. The state perception module collects the boiler's actual response data and generates a new state matrix. The digital twin construction module receives the new state matrix and updates the model parameters. The updated digital twin model provides new system state estimates and safe operation envelope constraints for the rolling optimization in the next sampling cycle.

10. The boiler steam production optimization device based on dynamic matching of electrical load according to claim 9, characterized in that, The state perception module collects real-time data to drive the twin construction module to generate a digital twin model; the parallel simulation module performs virtual simulation on the digital twin model to generate a set of candidate control strategies; the rolling optimization module performs rolling optimization calculations based on the candidate control strategy set and real-time state data; the feedforward compensation module performs decoupling compensation on the optimization instructions; and the execution intervention module realizes the safe execution of instructions and risk intervention.

Images