Intelligent control method and system for coal-fired boiler based on data analysis

Abstract

The invention discloses a coal-fired boiler intelligent control method and system based on data analysis, and relates to the technical field of efficient combustion of fire coal. A coal-fired boiler intelligent control system based on data analysis comprises a data acquisition module, a twin updating module, a manifold system module, an event updating module, a candidate evaluation module, a comprehensive evaluation module and a manifold correction module. According to the method, a multi-layer physical constraint manifold system comprising a physically feasible bottom manifold, a middle manifold for safe operation and a top manifold for optimal operation is constructed in an operation state-controlled quantity combined space, and physical constraint satisfaction, a safe operation area and an efficient low-emission area are simultaneously represented in a unified space; layered description and integrated expression of physical feasibility, safety and economical efficiency targets of the coal-fired boiler are achieved.

Description

Technical Field

[0001] This invention relates to the field of high-efficiency coal combustion technology, and in particular to an intelligent control method and system for coal-fired boilers based on data analysis. Background Technology

[0002] Coal-fired boilers, as key equipment in large-scale thermal power generation and industrial heating systems, exhibit highly nonlinear, strongly coupled, and multi-constrained combustion processes, making them susceptible to various factors such as fuel quality fluctuations, load changes, equipment aging, and environmental disturbances. Traditional control systems primarily rely on DCS, conventional PID regulation, and some advanced process control (APC) techniques, using empirically tuned control parameters and limited mechanistic models to maintain stable key operating indicators such as main steam pressure, temperature, and oxygen content. In recent years, with the development of sensor technology and industrial big data, some solutions have begun to incorporate data-driven modeling and digital twin concepts to simulate, predict, and optimize boiler operation processes, aiming to improve combustion efficiency, reduce energy consumption, and decrease pollutant emissions.

[0003] However, existing coal-fired boiler control methods based on data analysis and digital twins often impose safety and operational constraints on the optimization problem in the form of fixed thresholds or static inequalities. Performance optimization is typically performed only within a single model or single-layer constraint domain, with feasibility and safety verification of the calculation results afterward. Such methods struggle to simultaneously distinguish different levels of regions (physical, safety, and effect) within a unified operating state-control space. They fail to adequately consider the impact of operating condition changes and equipment degradation at different time scales, and lack targeted modeling and dynamic correction mechanisms for local high-risk areas under target event scenarios such as deep peak shaving at low loads, sensor failures, and fluctuations in the denitrification system. This results in control strategies that are either overly conservative and fail to tap the boiler's operational potential, or face the risk of physical constraints being approximated or even potentially breached under extreme conditions, making it difficult to achieve deep synergy between performance optimization and physical feasibility and safety. Summary of the Invention

[0004] To address the challenge of simultaneously considering physical feasibility constraints, safe operation requirements, and high-efficiency, low-emission objectives under complex operating conditions in existing technologies, this invention constructs a multi-layered physically constrained manifold system with spatiotemporal decoupling characteristics within the joint space of the operating state and control variables of a coal-fired boiler. This system hierarchically characterizes the physically feasible region, the safe operating region, and the optimal operating region, and can incrementally update relevant local regions using high-value samples when target events occur. This manifold system is tightly coupled with digital twin models, counterfactual inference, and multi-objective optimization processes, thereby achieving a synergistic trade-off and integrated consideration of physical constraints, safety risks, and performance optimization objectives during the control decision-making process, and enhancing the intelligent control capability of coal-fired boilers under complex and dynamic operating conditions.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A data analysis-based intelligent control method for coal-fired boilers includes: Real-time acquisition and preprocessing of operating data from coal-fired boilers yields an operating state vector. A digital twin model of a coal-fired boiler is constructed based on historical operating data, and the parameters of the digital twin model are corrected using operating state vectors. In the joint space of the operating state and control variables of the coal-fired boiler corresponding to the digital twin model, a spatiotemporally decoupled multi-layered physically constrained manifold system is constructed. Real-time detection of target events during boiler operation; based on the detection results, trigger incremental updates of parameters, boundaries, and confidence weights of relevant local regions in the digital twin model and multi-layered physical constraint manifold. Multiple sets of candidate control commands are generated based on the type of the target event, and then input into the updated digital twin model for counterfactual inference to obtain performance prediction results. The state-control trajectory generated by the inference is mapped to the updated multi-layered physical constraint manifold system to perform hierarchical safety assessment and obtain hierarchical assessment indicators. The coupling performance prediction results and hierarchical evaluation indicators are used to conduct a multi-objective comprehensive evaluation of each candidate control command, and the control command with the highest comprehensive score is selected as the target control command under the preset constraints. The target control command and the current operating state vector are jointly mapped to the joint space of the multi-layered physical constraint manifold system. Layered projection is performed, and the resulting modified control command is output as the final control command of the coal-fired boiler.

[0006] As a preferred embodiment of the present invention, the correction includes: inputting the current operating state vector into the digital twin model to obtain the model prediction output; calculating the prediction error between the model prediction output and the corresponding measured operating data; constructing a correction objective function based on the prediction error; and incrementally correcting the parameters of the digital twin model using a recursive update method under the constraints of a preset parameter change range and update rate.

[0007] As a preferred embodiment of the present invention, the multi-layered physically constrained manifold system includes: establishing a physical constraint model based on the first-principles equations of a coal-fired boiler, and determining a bottom-layer manifold that satisfies the constraints of mass conservation, energy conservation, and momentum conservation; within the bottom-layer manifold, establishing a middle-layer manifold for safe operation based on historical safe operation data through manifold learning; within the middle-layer manifold, obtaining a top-layer manifold for optimal operation based on historical multi-objective optimization operation data, wherein the historical safe operation data is a subset of historical operation data characterizing the safe operation of a coal-fired boiler, and the historical multi-objective optimization operation data is a subset of historical operation data characterizing the efficient, economical, and low-emission operation of a coal-fired boiler.

[0008] As a preferred technical solution of the present invention, the multi-layer physical constraint manifold system is decoupled into three update scales in the time dimension: short-term, medium-term, and long-term. The short-term manifold is updated based on real-time data at the second / minute level and is used for immediate safety assessment during counterfactual inference. The medium-term manifold is updated based on hourly operating condition cycle data and is used for operating condition pattern recognition and strategy template matching. The long-term manifold is updated based on monthly / yearly equipment aging data and is used to reflect the long-term evolution impact of structural parameters and performance degradation on the bottom and middle manifolds.

[0009] As a preferred embodiment of the present invention, the incremental update includes: detecting target running events based on the running state vector and the prediction error of the current digital twin model; when a target running event is detected, constructing a high-value sample from the running state vector and control quantity data of the corresponding time period, and determining the local region corresponding to the high-value sample in the joint space; updating the incremental parameters triggered by the execution event of the digital twin model based on the high-value sample, locally adjusting the manifold parameters and boundaries related to the local region in the multi-layered physically constrained manifold system, and updating the confidence weight for the local region, thereby realizing local incremental update oriented towards the target event.

[0010] As a preferred technical solution of the present invention, the counterfactual deduction includes: according to the type of the target event, calling the corresponding control strategy template from a pre-established scenario-based strategy library to generate multiple sets of candidate control instructions; under the combination of the current running state vector and the candidate control instructions, calling the digital twin model to perform rolling prediction within a set prediction time window, outputting the corresponding prediction sequence as the performance prediction result, and generating a state-control trajectory corresponding to the prediction sequence; the state-control trajectory is a sequence composed of the running state vector and the candidate control instructions at each moment within the prediction time window.

[0011] As a preferred technical solution of the present invention, the hierarchical security assessment includes: calculating the degree of violation of the state-control trajectory relative to the bottom layer manifold to obtain the degree of violation of physical constraints; calculating the degree of spatial geometric deviation of the state-control trajectory in the bottom layer manifold relative to the middle layer manifold to obtain the degree of security deviation; and calculating the distance of the state-control trajectory in the middle layer manifold relative to the top layer manifold to obtain the performance optimization potential index.

[0012] As a preferred technical solution of the present invention, the multi-objective comprehensive evaluation includes: constructing a comprehensive evaluation function based on the performance prediction results and hierarchical evaluation indicators corresponding to the candidate control commands; the performance items of the comprehensive evaluation function include main steam pressure and temperature fluctuations, combustion stability indicators, NOx emission levels and ammonia slip levels; the constraint penalty items include physical constraint violation degree and safety deviation degree; and the performance optimization potential index is used as a guiding item for approaching the top manifold; under the premise of comprehensively considering the performance items, constraint penalty items and guiding items, and satisfying the preset physical constraints, safety constraints and operational constraints, a comprehensive score is calculated and sorted for each candidate control command, and the candidate control command with the best comprehensive score is selected as the target control command.

[0013] As a preferred embodiment of the present invention, the layered projection includes: adjusting the target control command to a physically feasible bottom-level manifold so that the adjusted control quantity satisfies the constraints of mass conservation, energy conservation, and momentum conservation; projecting the adjusted control command to a safe-operating middle-level manifold while satisfying the bottom-level manifold, so that the state-control combination corresponding to the target control command falls into the safe-operating region; and optimizing the control command while satisfying the middle-level manifold, so that its distance from the optimal top-level manifold in the joint space is reduced, thereby obtaining a corrected control command and outputting it for execution.

[0014] A data analysis-based intelligent control system for a coal-fired boiler includes: Data acquisition module: Collects and preprocesses the operating data of the coal-fired boiler in real time to obtain the operating status vector; The digital twin update module constructs a digital twin model of a coal-fired boiler based on historical operating data and uses operating state vectors to correct the parameters of the digital twin model. Manifold System Module: Construct a spatiotemporally decoupled multi-layered physically constrained manifold system in the joint space of the operating state and control variables of the coal-fired boiler corresponding to the digital twin model; Event Update Module: Real-time detection of target events during boiler operation, and triggering incremental updates of parameters, boundaries, and confidence weights of relevant local regions in the digital twin model and multi-layered physical constraint manifold based on the detection results; Candidate evaluation module: Generates multiple sets of candidate control commands based on the type of the target event, and inputs them into the updated digital twin model for counterfactual inference to obtain performance prediction results; uses the state-control trajectory generated by the inference to map to the updated multi-layered physical constraint manifold system to perform hierarchical safety evaluation and obtain hierarchical evaluation indicators; The comprehensive evaluation module couples the performance prediction results with hierarchical evaluation indicators to perform a multi-objective comprehensive evaluation of each candidate control command, and selects the one with the highest comprehensive score as the target control command under preset constraints. Manifold Correction Module: Maps the target control command and the current operating state vector together to the joint space of the multi-layered physical constraint manifold system, performs layered projection, and the resulting corrected control command is output as the final control command of the coal-fired boiler.

[0015] The present invention has the following advantages: This invention improves the prediction accuracy and robustness of the boiler process model under different operating conditions by constructing a digital twin model based on historical data and using real-time operating state vectors to recursively correct its parameters during operation, thus providing a reliable model foundation for subsequent counterfactual inference and optimization decision-making.

[0016] This invention constructs a multi-layered physically constrained manifold system in the joint space of operating state and control variables, which includes a physically feasible bottom manifold, a safe operating middle manifold, and an optimal operating top manifold. It simultaneously characterizes the physical constraint satisfaction, safe operating region, and high-efficiency and low-emission region in a unified space, thereby achieving a hierarchical characterization and integrated expression of the physical feasibility, safety, and economic objectives of coal-fired boilers.

[0017] This invention decouples a multi-layered physically constrained manifold system into three update scales—short-term, medium-term, and long-term—in the time dimension, which are used for real-time safety assessment, operating condition pattern recognition, and equipment aging evolution modeling, respectively. This enables the manifold structure to evolve step by step with changes in real-time operating conditions, operating cycles, and long-term degradation characteristics, significantly enhancing the adaptability of control strategies to dynamic operating conditions and changes in equipment performance.

[0018] This invention introduces a scenario-based strategy library to generate multiple sets of candidate control instructions based on the type of target event. It then performs counterfactual inference within a set time window on a digital twin model to obtain the state-control trajectory and performance prediction results. The trajectory is then mapped to a multi-layered manifold system for hierarchical safety assessment to obtain indicators such as physical constraint violation degree, safety deviation degree, and performance optimization potential. This allows the control optimization process to fully perceive potential physically infeasible and safety risk areas during the optimization phase, thus suppressing high-risk control strategies from the source.

[0019] This invention constructs a multi-objective comprehensive evaluation function that includes performance terms, constraint penalty terms, and top-level manifold guidance terms. It couples the performance prediction results of candidate control commands with hierarchical evaluation indicators. Under the premise of satisfying physical constraints, safety constraints, and operational constraints, the candidate control commands are comprehensively scored and ranked. This achieves a coordinated trade-off between combustion stability, main steam parameter fluctuations, NOx emissions and ammonia slip levels, and the ability to approach the optimal operating range, thereby obtaining control commands with better overall performance.

[0020] This invention performs layered projection of the target control command within the joint space of the multi-layered physically constrained manifold system. First, the control command is adjusted to the physically feasible bottom manifold, then projected to the safe operating middle manifold, and finally approximates the optimal operating top manifold while satisfying the physical feasibility and safety constraints. This ensures that the final output control command not only strictly satisfies the basic physical laws and safe operating requirements, but also approaches the desired efficient, economical and low-emission operating state as closely as possible. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only schematic diagrams of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort. Figure 1 This is a schematic diagram of the structure of an intelligent control system for a coal-fired boiler based on data analysis, as used in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0023] Example 1: A data analysis-based intelligent control method for coal-fired boilers, comprising the following steps: Step S1: Collect and preprocess the operating data of the coal-fired boiler in real time to obtain the operating status vector; In one embodiment of the present invention, the operating data of the coal-fired boiler is collected by various sensors, transmitters, metering devices and power acquisition devices deployed on the boiler body and its auxiliary equipment, and uploaded to the data acquisition module through fieldbus or data acquisition network.

[0024] The operational data includes various physical quantities reflecting boiler thermal status, flue gas composition, air and coal supply, pressure and flow rate, load commands, and environmental conditions.

[0025] Specifically, the thermal state data in the operational data includes main steam pressure, main steam temperature, reheat steam pressure, and reheat steam temperature, used to characterize the boiler steam-side output conditions; it also includes economizer inlet water temperature, economizer outlet water temperature, inlet and outlet flue gas temperatures of each stage of superheater and reheater, flue gas temperatures before and after the air preheater, and pressure and pressure difference at relevant measuring points, used to characterize the heat transfer status of the boiler's heating surfaces and the temperature distribution on the flue gas side. The above data are collected by primary instruments such as pressure transmitters, thermocouples, and resistance temperature detectors (RTDs).

[0026] The flue gas composition data in the operational data includes oxygen content (O2), carbon monoxide (CO) concentration, nitrogen oxide (NOx) concentration, and sulfur dioxide (SO2) concentration (collected when necessary), which reflects the degree of combustion, excess air level, and pollutant emission levels. Flue gas composition data is collected by flue gas analyzers and continuous emission monitoring devices installed at different locations in the flue and uploaded at set intervals.

[0027] The operational data, including coal and air supply data, includes the coal feed rate or speed of each coal feeder, total primary and secondary air flow rate, primary and secondary air pressure, primary and secondary air temperature, secondary air damper opening in each floor or area, and the output, current, or speed of multiple forced draft fans, primary air fans, and induced draft fans. This data is used to characterize the current fuel input and air supply status. The above data is collected by belt scales, flow meters, differential pressure gauges, temperature sensors, and damper opening feedback devices.

[0028] The pressure and furnace status data in the operational data include furnace negative pressure, pressure differential in each section of the flue, pressure measurement data at key parts of the boiler body, and related safety monitoring quantities. These data describe the furnace ventilation status, flue gas flow status, and potential blockages, ash accumulation, or abnormal flow conditions. Furnace negative pressure and flue pressure differential are collected by relevant pressure transmitters.

[0029] The load and operating condition command data in the operational data include the unit's current active power output, unit load setpoint, unit load change rate, grid connection parameters, and mode commands related to the operating mode (such as base load operation, peak load operation, etc.), which are used to reflect the boiler's dispatching condition. When necessary, external disturbance data such as outdoor temperature and ambient temperature are also collected to help determine changes in heating demand or cooling conditions.

[0030] Based on the directly collected data, the operational data undergoes preprocessing, including time alignment, data cleaning, and feature construction. Time alignment involves unifying the timestamps and interpolating data from different sampling periods and acquisition channels, ensuring that the operational status at the same moment is expressed in a unified vector form. Data cleaning includes removing obvious erroneous and distorted values, filling in short-term missing data, and smoothing isolated spikes using median filtering or moving averages.

[0031] After preprocessing, a number of derived features are calculated based on direct measurement data to form a more expressive operating state vector. These derived features include boiler thermal efficiency estimated based on main steam parameters, fuel input, and flue gas parameters; excess air coefficient calculated based on flue gas oxygen content and theoretical air volume; and rates of change of key variables (e.g., main steam pressure change rate, flue gas O2 content change rate, NOx concentration change rate) and control deviations (differences between setpoints and measured values) for each control loop, calculated based on continuous time series data.

[0032] The measurement data, which has been processed by time alignment, cleaning and feature construction, are concatenated with the derived features in a unified order to form an operating state vector for subsequent modeling and control decisions. The operating state vector at each moment synchronously expresses the current thermal state, flue gas composition state, air and coal supply state, pressure and flow state, load and operating condition state of the boiler, as well as the corresponding derived features.

[0033] Step S2: Construct a digital twin model of the coal-fired boiler based on historical operating data, and use the operating state vector to correct the parameters of the digital twin model; The correction process includes: inputting the current operating state vector into the digital twin model to obtain the model's predicted output; calculating the prediction error between the model's predicted output and the corresponding measured operating data; constructing a correction objective function based on the prediction error; and incrementally correcting the parameters of the digital twin model using a recursive update method under the constraints of a preset parameter change range and update rate.

[0034] In one embodiment of the present invention, the "digital twin model" refers to a virtual model constructed for a coal-fired boiler that mathematically corresponds highly to the actual boiler's operating behavior. This model receives operating state vectors and control quantity data at its input end and provides predictions of key operating variables for future moments at its output end, used to characterize the dynamic response characteristics of the boiler under different operating conditions and control commands. The construction of the digital twin model is based on the historical operating data collected and preprocessed in step S1. The historical operating data covers multiple load ranges, different coal quality conditions, and various operating modes, and includes state and control quantities related to boiler thermal state, flue gas composition, air and coal supply, pressure and flow rate, and load commands.

[0035] In the specific construction process, historical operating data is organized into an input-output sample sequence in chronological order. The input part includes the operating state vector at each moment and the corresponding control commands, while the output part includes target predictions such as the main steam pressure, main steam temperature, furnace negative pressure, flue gas O2 / CO / NOx content, and key heating surface outlet flue gas temperature at the next moment or several steps later. Based on the above input-output samples, a model structure suitable for describing nonlinear and time-varying characteristics is selected, such as a hybrid model that superimposes data-driven correction terms on a simplified mechanistic model. The initial parameters of the digital twin model are obtained through training with historical samples, enabling the model to accurately characterize the dynamic behavior of the boiler under historical operating conditions.

[0036] To ensure the digital twin model remains adaptable to current operating conditions during online operation, a parameter correction process based on the operating state vector is introduced after modeling is completed. Specifically, at each sampling moment, the current operating state vector, along with the current control command, is input into the digital twin model to obtain the model's predicted output at that moment. Simultaneously, measured operating data corresponding to the model's predicted output is extracted from the real-time collected data to construct the prediction error. The prediction error encompasses the deviations between main steam pressure, main steam temperature, furnace negative pressure, flue gas composition, and other variables selected as prediction objects, and is used to characterize the degree of deviation of the model under the current operating condition.

[0037] Based on the prediction error, an objective function for parameter correction is constructed. This objective function is a weighted sum of the errors of each prediction variable, with weights set according to the variable's importance in the control decision. For example, main steam pressure, main steam temperature, and NOx emissions are given higher weights, while auxiliary variables are given relatively lower weights. To avoid drastic changes in model parameters within a short period, constraints on the parameter variation range and update rate are set during the correction process. By limiting the magnitude of each update and the cumulative adjustment interval, the parameter evolution process remains smooth and avoids entering the numerically unstable region.

[0038] Under the aforementioned constraints, a recursive update method is adopted to incrementally correct the parameters of the digital twin model. The recursive update method refers to adjusting the parameters in small steps based on the prediction error at the current moment, while retaining the parameter estimation results from the previous moment. This allows the parameters to gradually approach their optimal values ​​under the current operating conditions over time. This recursive process is continuously executed chronologically, constantly absorbing new operating samples. This ensures that the parameters of the digital twin model inherit the overall patterns formed under historical operating conditions during long-term operation, while also undergoing detailed adjustments based on current coal quality, load levels, and equipment status, thereby maintaining high prediction accuracy at different operating stages.

[0039] Step S3: Construct a spatiotemporally decoupled multi-layered physically constrained manifold system in the joint space of the operating state and control variables of the coal-fired boiler corresponding to the digital twin model; In one embodiment of the present invention, the multi-layer physically constrained manifold system includes: establishing a physical constraint model based on the first-principles equations of a coal-fired boiler, and determining a physically feasible bottom-layer manifold that satisfies the constraints of mass conservation, energy conservation, and momentum conservation by the physical constraint model; the first-principles equations include the mass conservation equation, the energy conservation equation, and the momentum conservation equation, and the physical constraint model is established by combining the state relationships and boundary conditions related to combustion, heat transfer, and flow; Within the bottom manifold, a middle manifold for safe operation is established through manifold learning based on historical safe operation data; within the middle manifold, a top manifold for optimal operation is obtained through learning based on historical multi-objective optimization operation data. The historical safe operation data is a subset of historical operation data representing the safe operation of coal-fired boilers, and the historical multi-objective optimization operation data is a subset of historical operation data representing the efficient, economical, and low-emission operation of coal-fired boilers.

[0040] The multi-layered physical constraint manifold system is decoupled into three update scales in the time dimension: short-term, medium-term, and long-term. The short-term manifold is updated based on real-time data at the second / minute level and is used for immediate safety assessment during counterfactual inference. The medium-term manifold is updated based on hourly operating condition cycle data and is used for operating condition pattern recognition and strategy template matching. The long-term manifold is updated based on monthly / yearly equipment aging data and is used to reflect the long-term evolution impact of structural parameters and performance degradation on the bottom and middle manifolds.

[0041] The "joint space of operating state and control variables of a coal-fired boiler" refers to a high-dimensional space composed of the operating state vector formed in step S1 and its corresponding control variable data. Each point in this space simultaneously provides the boiler's process state at a given moment (including main steam pressure and temperature, flue gas composition, temperatures of each heating surface, furnace negative pressure, air volume, coal volume, etc.) and the control commands at that moment (including coal feed rate of each coal feeder, primary and secondary air flow and distribution, output of forced and induced draft fans, opening degree of desuperheating water valve, ammonia injection rate of SCR, etc.). Within this joint space, the boiler's operating trajectory is represented as a state-control combination sequence that evolves over time. The multi-layered physically constrained manifold system geometrically characterizes the feasible region, safe region, and optimal region at different levels within this joint space.

[0042] The "multi-layered physically constrained manifold system" refers to a set of hierarchical high-dimensional geometric objects constructed within the aforementioned joint space. This system includes at least a physically feasible bottom-layer manifold, a safe-operating middle-layer manifold, and an optimal-operating top-layer manifold. The bottom-layer manifold characterizes the feasible region that satisfies the basic physical laws of coal-fired boilers. The middle-layer manifold is used to further eliminate operating regions with safety risks based on physical feasibility. The top-layer manifold is used to refine the preferred operating region that is efficient, economical, and low-emission based on safe operation. Through this multi-layered manifold system, physical feasibility, safety, and performance are described hierarchically using geometric boundaries at different levels within a unified state-control space, providing a structural foundation for subsequent hierarchical safety assessments and hierarchical projections.

[0043] When constructing the underlying manifold, firstly, representative control volumes and the material, energy, and momentum exchange relationships between them are selected based on the structural layout and process flow of the coal-fired boiler. Mass conservation equations, energy conservation equations, and momentum conservation equations are then written, and combined with the state relationships of combustion, heat transfer, and fluid flow, as well as boundary conditions, to form a physical constraint model. For example, the mass conservation equation constrains the mass balance between fuel, air, and generated flue gas entering the furnace; the energy conservation equation constrains the energy balance between fuel chemical energy, steam-side heat absorption, flue gas-side heat release, and losses; and the momentum conservation equation constrains the relationship between pressure drop and flow rate of flue gas in the furnace and flue. These equations, along with the state relationships and boundary conditions, collectively define the set of operating states and control quantities that satisfy physical laws under given operating conditions. This set forms a continuous region with a certain geometric structure in the joint space, i.e., a physically feasible underlying manifold.

[0044] Based on the underlying manifold, a subset of historical operating data characterizing boiler safety is obtained through screening and statistical analysis of historical safe operation data. This subset corresponds to a time period in which boiler protection and interlocking mechanisms were not triggered during actual operation, key operating indicators (such as main steam pressure, main steam temperature, furnace negative pressure, and boiler metal wall temperature) remained within predetermined safe ranges, and no events such as severe coking, tube rupture, or furnace vibration were recorded. Mapping this historical safe operation data to the state-control joint space yields a series of discrete sample points. These sample points are then modeled using manifold learning methods to obtain a mid-level manifold that encloses the safe sample points within the physically feasible underlying manifold.

[0045] "Manifold learning" refers to a nonlinear modeling method that extracts the essential low-dimensional structure from high-dimensional data. By considering the local proximity relationships and global geometric relationships between samples in a high-dimensional joint space, it fits a smooth, connected low-dimensional subspace to approximate the backbone structure of the data distribution. In specific implementations, neighborhood-preserving dimensionality reduction algorithms, autoencoder structures, or other nonlinear dimensionality reduction models can be used. Through training, the model forms a continuous surface or family of curves in the joint space. This surface or family of curves is the safe-operating layered manifold, representing the state-control combination that can operate safely for a long time under the premise of satisfying physical constraints.

[0046] Within the mid-level manifold, historical multi-objective optimization operation data is introduced to further construct the optimal top-level manifold. This historical multi-objective optimization operation data is selected from safe operation data, with selection criteria including low coal consumption per unit load, low NOx emissions and ammonia slip levels, and minimal fluctuations in main steam pressure and temperature, reflecting a comprehensive advantage in combustion efficiency, environmental emissions, and operational stability. Mapping these optimal operating condition state-control samples onto a joint space yields a cluster of high-quality sample points located within the safe region. A similar manifold learning process is then used to fit a high-dimensional subspace covering these sample points, representing the optimal top-level manifold. Geometrically, this top-level manifold lies within the mid-level manifold, representing the operating region achieving high efficiency, economy, and low emissions under safe operating conditions.

[0047] To reflect the "spatiotemporal decoupling" characteristic, the multi-layered physically constrained manifold system is divided into three update scales in the time dimension: short-term, medium-term, and long-term. The short-term manifold is updated based on real-time data at the second / minute level. Updates primarily include fine-tuning the local shape of the safety boundary and optimal region under current coal quality, load, and environmental conditions, used for immediate safety assessment of the operating trajectory within the next few minutes during counterfactual simulations. The medium-term manifold is updated based on hourly operating cycle data. It re-identifies and matches safe and preferred regions under different operating conditions based on changes in load curves and operating modes within a day or several hours, ensuring the manifold structure aligns with the operating conditions. The long-term manifold is updated based on monthly / yearly equipment aging data. By analyzing the impact of factors such as fouling on heating surfaces, heat transfer performance degradation, and changes in duct resistance on the physically feasible region and safe operating region over long time scales, it structurally adjusts the bottom and middle manifolds, enabling the multi-layered manifold system to reflect the long-term evolution characteristics of the boiler as it ages.

[0048] Step S4: Real-time detection of target events during boiler operation, and triggering incremental updates of parameters, boundaries, and confidence weights of relevant local regions in the digital twin model and multi-layer physical constraint manifold based on the detection results; The incremental update includes: detecting target running events based on the running state vector and the prediction error of the current digital twin model; when a target running event is detected, constructing high-value samples from the running state vector and control data of the corresponding time period, and determining the local region corresponding to the high-value samples in the joint space; updating the incremental parameters triggered by the execution event of the digital twin model based on the high-value samples, locally adjusting the manifold parameters and boundaries related to the local region in the multi-layered physically constrained manifold system, and updating the confidence weights for the local region, thereby realizing local incremental updates oriented towards the target event.

[0049] In one embodiment of the present invention, a "target event" refers to a specific operating condition or abnormal situation that has a significant impact on the safe operation and control performance of a coal-fired boiler. This includes both extreme changes in operating conditions caused by dispatching instructions and abnormalities caused by equipment or measurement systems. Typical target events include low-load deep peak-shaving conditions, rapid load increases and decreases in the unit, unstable combustion in the furnace, drift or failure of key sensors, and abnormal operation of the denitrification system. For example, during deep peak-shaving, the unit load drops from the rated value to a lower load, reducing the combustion stability margin and increasing fluctuations in furnace temperature and flue gas composition. When the oxygen probe drifts or fails, the invariant relationship between flue gas O2 measurements and other process quantities shifts significantly. When the SCR denitrification system malfunctions, the response relationship between ammonia injection and SCR outlet NOx concentration is disrupted, and abrupt changes occur in ammonia escape levels. These scenarios all constitute typical target events.

[0050] The detection of target events is based on the prediction error between the operating state vector and the predicted output of the digital twin model, as well as the changing characteristics of the operating state itself. When the digital twin model, under normal operating conditions, undergoes recursive correction in step S2, the error between its predicted output and the measured operating data remains within a predetermined normal range. However, when sudden changes occur in coal quality, drastic load changes, or performance degradation of key equipment, the amplitude or pattern of the model prediction error changes significantly on certain variables. For example, during deep peak shaving, the main steam pressure prediction error increases rapidly and its sign changes repeatedly; when the oxygen probe drifts, the residual between the measured O2 value and the oxygen calculated based on CO, load, and airflow continuously deviates from zero and exceeds an empirical threshold. The amplitude, duration, and pattern of the prediction error are compared with pre-set event criteria. When the target event criteria are met, the operating data at the current moment and within a certain time window before and after it are marked as the data corresponding to the target event.

[0051] After detecting a target event, a time period related to the target event is selected, and the continuously collected operating state vectors and corresponding control quantity data within this time period are combined to form a "high-value sample". A high-value sample refers to a sequence of state-control data covering the entire process of operating condition changes within a certain time window before and after the target event. This sequence reflects the actual response behavior of the boiler system under such events. Taking low-load deep peak shaving as an example, the high-value sample includes operating state vectors (such as main steam parameters, furnace negative pressure, flue gas composition, air volume, and coal quantity) and corresponding control quantities (such as coal feed, primary and secondary air volume and distribution, and ammonia injection) at various moments during the process of the load gradually decreasing from a higher level to a deep peak shaving level. This is used to characterize the entire process of reduced combustion stability margin, adjustment of excess air coefficient, and changes in emissions.

[0052] After obtaining high-value samples, these samples are mapped into a joint space of operational states and control variables. By setting a neighborhood scale or similarity metric, a "local region" corresponding to the concentrated distribution of high-value samples is determined within the multi-layered physically constrained manifold system. A local region refers to an area within the joint space centered on a high-value sample point and within a certain distance or similarity range. This region aggregates operational state-control combinations related to the target event and represents the concentrated distribution area of ​​target event data within the multi-layered manifold. The location and shape of local regions differ for different target events. For example, local regions related to low-load peak shaving are mainly distributed near lower loads and higher excess air coefficients, while local regions related to oxygen probe failure are concentrated in areas where O2 measurements are inconsistent with other variables.

[0053] Based on a defined local region, the digital twin model undergoes event-triggered incremental parameter updates. Compared to the recursive correction for the global operating condition in step S2, the incremental update in step S4 primarily uses high-value samples, focusing on adjusting model parameters related to the target event to reduce the prediction error of the digital twin model under such operating conditions. For example, in a low-load deep peak-shaving event, combustion-related parameters and heat balance-related parameters are fine-tuned based on high-value samples from the deep peak-shaving process, making the model's predictions of main steam pressure, main steam temperature, furnace negative pressure, and flue gas composition more accurate in the low-load region; in an oxygen probe drift event, the mapping relationship related to flue gas oxygen in the model is adjusted using high-value samples, ensuring consistency between the model's internal estimation of oxygen and other variables.

[0054] Simultaneously, local adjustments are made to the manifold parameters and boundaries related to local regions in the multi-layered physically constrained manifold system. For the bottom manifold, based on the actual physical behavior reflected by high-value samples, the margins and limits of mass conservation, energy conservation, and momentum conservation constraints under local operating conditions are modified, making the bottom-layer physical feasible region more accurately fit the real feasible region in this local area. For the safe-operating middle-layer manifold, the position of the safety boundary of this local area is fine-tuned according to whether the high-value samples trigger safety alarms, protection actions, or important operating indicator boundaries during the event, making the division of safe and unsafe areas in the middle-layer manifold more refined under the event conditions. For the optimal-operating top-layer manifold, the shape of the optimal region within this local area is refined based on the performance of indicators such as thermal efficiency, NOx emissions, ammonia slip, and main steam parameter fluctuations in the high-value samples, enhancing the ability to describe high-performance operating states under target event conditions.

[0055] For each local region, its confidence weight within the multi-layered physically constrained manifold system is updated. "Confidence weight" refers to the quantitative evaluation value of each local region's reliability regarding its own geometry and boundaries, reflecting the richness of historical data and the degree of support from recent event data. After a target event occurs and new high-value samples are acquired, the confidence weight of the corresponding local region is adjusted based on the number and distribution of high-value samples and their consistency with the existing manifold boundaries. When high-value samples are highly consistent with the existing manifold structure, the confidence weight of the local region is increased; when high-value samples exhibit systematic deviations from the existing manifold boundaries, the confidence weight of the original structure is appropriately reduced while adjusting the local boundaries.

[0056] Step S5: Generate multiple sets of candidate control commands based on the type of the target event, and input them into the updated digital twin model for counterfactual inference to obtain performance prediction results; use the state-control trajectory generated by the inference to map to the updated multi-layered physical constraint manifold system to perform hierarchical safety assessment and obtain hierarchical assessment indicators. The counterfactual deduction includes: based on the type of the target event, calling the corresponding control strategy template from a pre-established scenario-based strategy library to generate multiple sets of candidate control instructions; under the combined conditions of the current running state vector and the candidate control instructions, calling the digital twin model to perform rolling prediction within a set prediction time window, outputting the corresponding prediction sequence as the performance prediction result, and generating a state-control trajectory corresponding to the prediction sequence; the state-control trajectory is a sequence composed of the running state vector and the candidate control instructions at each moment within the prediction time window.

[0057] The hierarchical security assessment includes: calculating the degree of violation of the state-control trajectory relative to the bottom manifold to obtain the physical constraint violation degree; calculating the spatial geometric deviation of the state-control trajectory in the bottom manifold relative to the middle manifold to obtain the security deviation degree; and calculating the distance of the state-control trajectory in the middle manifold relative to the top manifold to obtain the performance optimization potential index.

[0058] In one embodiment of the present invention, "counterfactual deduction" refers to the virtual calculation and prediction of the boiler's operation over a future period of time under different candidate control commands, without actually changing the boiler's current control commands. The counterfactual deduction starts with the current operating state vector and different candidate control commands, and rolls forward step by step within a given prediction time window, outputting the predicted operating state and related performance indicators at each moment, thereby forming multiple prediction trajectories representing the effects of different control strategies in the virtual space.

[0059] The "scenario-based strategy library" refers to a set of control strategy templates pre-built according to different target event types. These templates are derived from operational experience, historical optimization results, and summaries of control principles for specific operating conditions. For example, for low-load deep peak-shaving scenarios, the strategy library includes a set of control modes for adjusting the ratio of primary and secondary air, the opening degree of stratified air distribution, and the degree of graded combustion, used to reduce NOx emissions and fly ash carbon content while maintaining stable combustion. For scenarios where oxygen probes drift or fail, the strategy library includes a set of air volume adjustment modes that enhance CO, flue gas temperature, and load as feedback quantities, used to maintain combustion completeness and appropriate excess air levels under unreliable oxygen measurement conditions. For scenarios where the denitrification system is abnormal, the strategy library includes a set of coordinated control modes that coordinate in-furnace air distribution and SCR ammonia injection, used to suppress ammonia escape while maintaining NOx outlet concentration compliance.

[0060] In practice, upon detecting a target event type, a set of control strategy templates matching that event type is selected from the scenario-based strategy library. Each control strategy template provides a set of control adjustment schemes based on the current control command. Examples include increasing or decreasing the coal feed rate of each coal feeder by different magnitudes, adjusting the total primary and secondary air volume and the opening of each damper by different proportions, and raising or lowering the SCR ammonia injection flow rate by different levels. By overlaying these adjustment schemes with the current control command, multiple sets of candidate control commands are obtained, each set corresponding to a control strategy to be evaluated.

[0061] In the counterfactual simulation, the current operating state vector is used as the initial state. For each set of candidate control commands, the digital twin model is invoked to perform rolling predictions within a set prediction time window. The length of the prediction time window is set according to the boiler capacity, system inertia, and the object being regulated, generally covering a time range of several minutes to more than ten minutes to ensure that the main dynamic processes after the control command takes effect are included. Within this time window, the digital twin model receives the predicted operating state and corresponding candidate control command from the previous time step at each discrete time step, and outputs the predicted operating state and related output quantities for the next time step, such as main steam pressure, main steam temperature, furnace negative pressure, flue gas O2 / CO / NOx concentration, and thermal efficiency estimate. The prediction outputs of multiple consecutive time steps are merged to form the prediction sequence corresponding to the candidate control command, i.e., the performance prediction result.

[0062] While obtaining the performance prediction results, the predicted operating state vectors at each moment within the prediction time window are combined with the corresponding candidate control commands in chronological order to form a "state-control trajectory". This trajectory fully describes the virtual evolution path of the boiler in the combined space from the current moment to the end of the prediction time window, under the premise of adopting a certain candidate control command. Different candidate control commands correspond to different state-control trajectories, providing basic data for subsequent hierarchical safety assessment in a multi-layered physically constrained manifold system.

[0063] In the hierarchical safety assessment process, the relationship between the state-control trajectory and the underlying manifold is first evaluated based on the "physical constraint violation degree." The physical constraint violation degree characterizes the extent to which the predicted trajectory deviates from the physically feasible region defined by first-principles equations such as mass conservation, energy conservation, and momentum conservation. Specifically, at each moment within the prediction time window, the balance between predicted state variables and control variables is assessed according to the physical constraint model. When the prediction results at certain moments lead to energy budget imbalances, mass budget imbalances, or discrepancies between flow pressure drop and flow rate, that moment is considered to have a tendency to violate physical constraints. The degree of violation at each moment is aggregated over time to obtain the physical constraint violation degree of the corresponding state-control trajectory. The larger this index, the closer the candidate control command is to or further away from the physically feasible boundary.

[0064] Secondly, the spatial geometric deviation of the state-control trajectory within the lower manifold relative to the middle manifold is assessed based on the "safety deviation." The safety deviation characterizes the proximity of the trajectory to the safe operating region under physically feasible conditions. At each moment within the prediction time window, the predicted state-control point is projected into the lower manifold in the joint space, and its relative position to the safe operating middle manifold is calculated. When the point is located within the safe operating region, the safety deviation is smaller; when the point approaches the boundary of the middle manifold or even crosses out of the safe region, the safety deviation increases. By statistically analyzing the degree of safety deviation at all moments within the entire prediction time window, the safety deviation of the state-control trajectory can be obtained. This index reflects whether the trajectory frequently approaches or crosses the safe operating boundary throughout the prediction process.

[0065] Next, the distance of the state-control trajectory within the mid-level manifold relative to the top-level manifold is evaluated based on the "performance optimization potential index." The performance optimization potential index quantitatively describes the gap between the trajectory and its historically efficient, economical, and low-emission operating range. While maintaining its position within the safe operating mid-level manifold, a certain geometric distance exists between the state-control point at each moment within the prediction time window and the optimal operating top-level manifold. This distance reflects the degree of deviation of the operating state from the optimal operating range at that moment. The distances at each moment within the prediction time window are aggregated according to a certain weight to obtain the performance optimization potential index of the trajectory. A larger index value indicates that the trajectory is generally farther from the historical optimal operating range, indicating greater potential for performance improvement; a smaller value indicates that the trajectory is generally closer to the optimal operating range.

[0066] Step S6: Couple the performance prediction results with the hierarchical evaluation index, perform a multi-objective comprehensive evaluation of each candidate control command, and select the one with the highest comprehensive score as the target control command under the preset constraints. The multi-objective comprehensive evaluation includes: constructing a comprehensive evaluation function based on the performance prediction results and hierarchical evaluation indicators corresponding to candidate control commands. The performance items of the comprehensive evaluation function include main steam pressure and temperature fluctuations, combustion stability indicators, NOx emission levels, and ammonia slip levels. The constraint penalty items include physical constraint violation degree and safety deviation degree. The performance optimization potential indicator is used as a guiding term to approach the top manifold. Under the premise of comprehensively considering the performance items, constraint penalty items, and guiding terms, and satisfying the preset physical constraints, safety constraints, and operational constraints, a comprehensive score is calculated and sorted for each candidate control command. The candidate control command with the best comprehensive score is selected as the target control command.

[0067] In one embodiment of the present invention, "multi-objective comprehensive evaluation" refers to simultaneously considering multiple performance indicators and multiple constraint-related indicators based on the prediction results corresponding to a single candidate control command, and quantifying their overall advantages and disadvantages through a unified evaluation function. The inputs to the comprehensive evaluation function include the performance prediction results obtained in step S5 and the hierarchical evaluation indicators, wherein the performance prediction results focus on describing the stability of main steam parameters, combustion stability, and emission levels, and the hierarchical evaluation indicators focus on describing the spatial relationship with the bottom manifold, the middle manifold, and the top manifold.

[0068] The "performance term" in the comprehensive evaluation function is derived from quantities such as main steam pressure and temperature fluctuations, combustion stability indicators, NOx emission levels, and ammonia slip levels within the prediction time window. Main steam pressure and temperature fluctuations are calculated by measuring the range, amplitude, or statistical dispersion of these parameters at various points within the prediction time window, reflecting the stability of the main steam parameters under this control strategy. The combustion stability indicator is constructed based on predicted furnace negative pressure changes, CO concentration, flame detector signals, and other data, describing whether combustion is in a stable region. For example, high CO concentrations and increased furnace negative pressure fluctuations indicate decreased combustion stability. NOx emission levels and ammonia slip levels are statistically analyzed based on the SCR outlet NOx concentration and ammonia slip monitoring data in the prediction sequence, providing the average emission levels and fluctuations within the prediction time window for evaluating environmental performance.

[0069] The "constraint penalty term" in the comprehensive evaluation function includes two parts: physical constraint violation degree and safety deviation degree. The physical constraint violation degree is provided by the index calculated for the lower-level manifold in step S5. This index describes the degree to which the state-control trajectory violates physical constraints such as mass conservation, energy conservation, and momentum conservation within the prediction time window. When the trajectory approaches or crosses the boundary of the physically feasible region, the physical constraint violation degree increases significantly, thus imposing a larger penalty on such candidate control commands in the comprehensive evaluation. The safety deviation degree is provided by the spatial geometric deviation degree calculated for the middle-level manifold in step S5. This index describes the distance of the trajectory from the safe operating region under the premise of physical feasibility. When the trajectory frequently approaches the safe boundary or crosses the safe operating region, the safety deviation degree increases, thus suppressing such candidate control commands in terms of safety in the comprehensive evaluation.

[0070] The "guiding term" in the comprehensive evaluation function uses a performance optimization potential index to guide candidate control commands toward the optimal operating top-level manifold. This performance optimization potential index is provided by the distance index calculated for the top-level manifold in step S5, reflecting the average distance between the predicted trajectory and historically efficient, economical, and low-emission operating areas. When the trajectory corresponding to a candidate control command is generally close to the optimal operating area, the guiding term provides a positive evaluation, encouraging the selection of this type of control strategy; when the trajectory is far from the optimal operating area, the guiding term's evaluation value is lower or turns negative, indicating that there is still significant room for performance improvement.

[0071] When constructing the comprehensive evaluation function, corresponding weights are set according to the importance of main steam parameter stability, combustion stability, environmental emissions, and physical and safety constraints in actual operation. The weight setting follows the principle of safety priority; that is, if the degree of violation of physical constraints or safety deviation exceeds a preset threshold, a mandatory penalty is imposed in the comprehensive evaluation, preventing the comprehensive score from meeting the selection criteria and thus excluding candidate control commands that are physically infeasible or pose significant safety risks. Under the premise of satisfying physical and safety constraints, the weights among main steam pressure and temperature fluctuations, combustion stability indicators, NOx emission levels, and ammonia slip levels are configured according to the unit's current operating objectives (e.g., prioritizing energy conservation or emission reduction) to ensure that the comprehensive evaluation results meet dispatch requirements and operating strategies.

[0072] In actual calculations, a comprehensive evaluation function is invoked for each group of candidate control commands, using the corresponding performance prediction results and hierarchical evaluation indicators as input to obtain a comprehensive score for that candidate control command. A higher comprehensive score indicates that, under the premise of satisfying physical feasibility and safety, the candidate control command performs better in terms of main steam parameter stability, combustion stability, and emission levels, and its overall operating state is closer to the optimal operating range. After completing the comprehensive evaluation of all candidate control commands, they are sorted from highest to lowest score. Candidate control commands that do not meet the preset physical constraints, safety constraints, and operational constraints are eliminated. The group with the highest comprehensive score among the remaining qualified candidate control commands is selected as the target control command.

[0073] Step S7: Map the target control command and the current operating state vector together to the joint space of the multi-layer physical constraint manifold system, perform layered projection, and the resulting corrected control command is used as the final control command output of the coal-fired boiler; The layered projection includes: adjusting the target control command to a physically feasible bottom-level manifold so that the adjusted control quantity satisfies the constraints of mass conservation, energy conservation, and momentum conservation; projecting the adjusted control command to a safe-operating middle-level manifold while satisfying the bottom-level manifold requirements, so that the state-control combination corresponding to the target control command falls into the safe-operating region; and optimizing the control command while satisfying the middle-level manifold requirements, so that its distance from the optimal top-level manifold in the joint space is reduced, thereby obtaining a corrected control command and outputting it for execution.

[0074] In one embodiment of the present invention, "layered projection" refers to the hierarchical geometric adjustment of the state-control point determined by the target control command and the current operating state vector within the joint space of operating state and control variables, according to the hierarchical order of a multi-layered physically constrained manifold system. This adjustment ensures that the manifold first meets physical feasibility requirements, then meets safe operation requirements, and, based on this, approximates the optimal operating region as closely as possible. The layered projection process transforms the bottom-layer, middle-layer, and top-layer manifolds constructed in step S3 from static constraint structures into dynamic correction tools for control commands, ensuring that the final output control commands maintain consistency across multiple dimensions, including physical, safety, and performance.

[0075] First, at the underlying manifold level, the physical feasibility of the target control commands is adjusted. The target control commands, together with the current operating state vector, determine the expected state-control combination in the near future, which may be located outside the underlying manifold or close to its physical boundary. Based on the physical constraint model, the mass balance, energy balance, and momentum balance caused by the target control commands are assessed. When deviations in these balance relationships are detected, the state-control combination is pushed back into the underlying manifold by adjusting key control quantities in the target control commands. For example, when adjusting the combination of coal feed, primary air volume, and secondary air volume, their proportional relationships are corrected to ensure that the mass balance between fuel input, air supply, and flue gas generation once again satisfies the physical constraints. When changing load and flow distribution, the configuration of flow rate and pressure difference in each segment is refined to align with the flow law described by the momentum equation. After projection onto the underlying manifold, a version of the control commands that satisfies the constraints of mass conservation, energy conservation, and momentum conservation is obtained.

[0076] Secondly, at the mid-manifold level, control commands that have met physical feasibility are adjusted for safe operation. At this point, the target control commands are physically feasible, but may still lead to the operating state approaching or crossing the safe operating boundary. Based on the geometry and boundary shape of the mid-manifold for safe operation, the relative position of the state-control combination within or near the boundary of the manifold is evaluated. When the state-control combination is in the middle of the safe zone, only the control commands projected from the bottom layer are retained. When the state-control combination is close to the safe boundary or outside the safe zone, the control commands are directionally adjusted, moving the corresponding points inward into the safe zone. For example, at high loads and when the furnace negative pressure is close to the limit boundary, the furnace negative pressure is brought back to a region with a larger safety margin by appropriately reducing the coal feed rate or redistributing the secondary air volume. When the denitrification intensity is high and ammonia escape is close to the limit level, the NOx and ammonia escape fall within the safe operating boundary simultaneously by reducing the ammonia injection rate or enhancing the degree of low-NOx combustion in the furnace. After mid-manifold projection, a version of the control commands that meets both physical feasibility and falls within the safe operating zone is obtained.

[0077] Finally, at the top-level manifold, the control commands that meet the physical feasibility and safe operation requirements are optimized and approximated. While ensuring no deviation from the intermediate safety manifold, the control commands are finely adjusted based on the geometric position of the optimal operating top-level manifold, reducing the distance between the state-control combination and the optimal operating region in the joint space. For example, among multiple safe and feasible air-coal ratio schemes, priority is given to adjusting to an air-coal ratio that ensures stable combustion, meets emission standards, and has high thermal efficiency; among multiple safe and feasible ammonia injection configurations, priority is given to adjusting to a configuration with lower NOx emissions and less ammonia slip. By making small adjustments to the coal feed rate, stratified air distribution opening, ammonia injection rate, and related control quantities, the state-control point gradually approaches the preferred operating region represented by the top-level manifold within the intermediate safety manifold, thereby improving combustion efficiency and environmental performance without sacrificing safety.

[0078] After completing the layered projection, the resulting corrected control command satisfies the physical feasibility conditions defined by the bottom-level physical constraint manifold, falls within the safe operating region defined by the middle-level manifold, and is as close as possible to the historical optimal operating region in terms of the top-level manifold. This corrected control command serves as the final control command output of the method in this embodiment and acts as the basis for execution in the next control cycle. It realizes the synergistic effect of digital twin prediction, spatiotemporal decoupling of multi-layered physical constraint manifolds, and multi-objective comprehensive optimization in the control decision-making process, thereby improving the safety and operational performance of coal-fired boilers under complex operating conditions.

[0079] Example 2: A data analysis-based intelligent control system for a coal-fired boiler, comprising the following modules: Data acquisition module: Collects and preprocesses the operating data of the coal-fired boiler in real time to obtain the operating status vector; The digital twin update module constructs a digital twin model of a coal-fired boiler based on historical operating data and uses operating state vectors to correct the parameters of the digital twin model. Manifold System Module: Construct a spatiotemporally decoupled multi-layered physically constrained manifold system in the joint space of the operating state and control variables of the coal-fired boiler corresponding to the digital twin model; Event Update Module: Real-time detection of target events during boiler operation, and triggering incremental updates of parameters, boundaries, and confidence weights of relevant local regions in the digital twin model and multi-layered physical constraint manifold based on the detection results; Candidate evaluation module: Generates multiple sets of candidate control commands based on the type of the target event, and inputs them into the updated digital twin model for counterfactual inference to obtain performance prediction results; uses the state-control trajectory generated by the inference to map to the updated multi-layered physical constraint manifold system to perform hierarchical safety evaluation and obtain hierarchical evaluation indicators; The comprehensive evaluation module couples the performance prediction results with hierarchical evaluation indicators to perform a multi-objective comprehensive evaluation of each candidate control command, and selects the one with the highest comprehensive score as the target control command under preset constraints. Manifold Correction Module: Maps the target control command and the current operating state vector together to the joint space of the multi-layered physical constraint manifold system, performs layered projection, and the resulting corrected control command is output as the final control command of the coal-fired boiler.

[0080] Example 3: The structure of an intelligent control system for a coal-fired boiler based on data analysis, see [link to example]. Figure 1 As shown, it includes the following: Data acquisition module obtains operational data / feedback signals from coal-fired boilers and on-site execution / measuring equipment.

[0081] Data acquisition module: Acquire sensor / process quantities; preprocess (alignment, noise reduction, outlier handling).

[0082] Output: Running state vector.

[0083] Twin update module: The digital twin model is initialized based on historical data; parameters are corrected online using the running state vector and prediction error.

[0084] Output: The updated digital twin model.

[0085] Manifold System Module: Construct a physically feasible bottom-level manifold, a safe-running middle-level manifold, and an optimally running top-level manifold in the joint space of "state-control variables", supporting spatiotemporal decoupling (short-term / medium-term / long-term).

[0086] Event update module: Based on the prediction error detection event of the running state vector + twin model, the corresponding data is marked as high-value samples; triggering incremental updates of local parameters of digital twin model, local boundaries of multi-layer physical constraint manifold and confidence.

[0087] Output upwards: High-value samples / local regions / confidence levels → manifold system module.

[0088] Output downwards: Target event → Candidate evaluation module.

[0089] Candidate evaluation module: Based on the target event type, multiple sets of candidate instructions are generated from the policy library; candidate instructions + current state → counterfactual inference of digital twin model: performance prediction results are obtained; the state-control trajectory is mapped to a multi-layer manifold for hierarchical security assessment → hierarchical assessment indicators (physical violation degree / security deviation degree / optimal distance).

[0090] Output: Performance prediction results + hierarchical evaluation indicators → comprehensive evaluation module.

[0091] Comprehensive evaluation module: Construct a comprehensive evaluation function: performance terms (pressure and temperature fluctuations, NOx, ammonia escape, etc.), constraint penalty terms (physical violation degree, safety deviation degree), and guidance terms (potential indicators for approaching the optimal manifold); calculate and rank the comprehensive scores of each candidate command, and select the target control command.

[0092] Output: Target control command + current running state vector → manifold correction module.

[0093] Manifold correction module: The target control command is mapped to a multi-layered physical constraint manifold system, and layered projection is performed sequentially: projection to the physically feasible bottom manifold (physical constraint), projection to the safe operating middle manifold (safety constraint), and projection to the top manifold that approximates the optimal operation as much as possible (performance optimization); the corrected control command is obtained and output.

[0094] Output: Corrected control commands → coal-fired boilers and field execution / measuring equipment.

[0095] The specific embodiments of the coal-fired boiler described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A data analysis-based intelligent control method for coal-fired boilers, characterized in that, include: Real-time acquisition and preprocessing of operating data from coal-fired boilers yields an operating state vector. A digital twin model of a coal-fired boiler is constructed based on historical operating data, and the parameters of the digital twin model are corrected using operating state vectors. In the joint space of the operating state and control variables of the coal-fired boiler corresponding to the digital twin model, a spatiotemporally decoupled multi-layered physically constrained manifold system is constructed. Real-time detection of target events during boiler operation; based on the detection results, trigger incremental updates of parameters, boundaries, and confidence weights of relevant local regions in the digital twin model and multi-layered physical constraint manifold. Multiple sets of candidate control commands are generated based on the type of the target event, and then input into the updated digital twin model for counterfactual inference to obtain performance prediction results. The state-control trajectory generated by the inference is mapped to the updated multi-layered physical constraint manifold system to perform hierarchical safety assessment and obtain hierarchical assessment indicators. The coupling performance prediction results and hierarchical evaluation indicators are used to conduct a multi-objective comprehensive evaluation of each candidate control command, and the control command with the highest comprehensive score is selected as the target control command under the preset constraints. The target control command and the current operating state vector are jointly mapped to the joint space of the multi-layered physical constraint manifold system. Layered projection is performed, and the resulting modified control command is output as the final control command of the coal-fired boiler.

2. The intelligent control method for a coal-fired boiler based on data analysis according to claim 1, characterized in that, The correction process includes: inputting the current operating state vector into the digital twin model to obtain the model's predicted output; calculating the prediction error between the model's predicted output and the corresponding measured operating data; constructing a correction objective function based on the prediction error; and incrementally correcting the parameters of the digital twin model using a recursive update method under the constraints of a preset parameter change range and update rate.

3. The intelligent control method for coal-fired boilers based on data analysis according to claim 1, characterized in that, The multi-layered physically constrained manifold system includes: establishing a physical constraint model based on the first-principles equations of a coal-fired boiler, and determining a bottom-layer manifold that satisfies the constraints of mass conservation, energy conservation, and momentum conservation; within the bottom-layer manifold, establishing a safe operating middle-layer manifold through manifold learning based on historical safe operating data; and within the middle-layer manifold, obtaining a top-layer manifold with optimal operation based on historical multi-objective optimization operating data. The historical safe operating data is a subset of historical operating data characterizing the safe operation of the coal-fired boiler, and the historical multi-objective optimization operating data is a subset of historical operating data characterizing the efficient, economical, and low-emission operation of the coal-fired boiler.

4. The intelligent control method for a coal-fired boiler based on data analysis according to claim 3, characterized in that, The multi-layered physical constraint manifold system is decoupled into three update scales in the time dimension: short-term, medium-term and long-term. The short-term manifold is updated based on real-time data at the second / minute level and is used for immediate security assessment during counterfactual inference. The intermediate manifold is updated based on hourly operating cycle data and is used for operating pattern recognition and strategy template matching. The long-term manifold is updated based on monthly / grade equipment aging data to reflect the impact of structural parameters and performance degradation on the long-term evolution of the bottom and middle manifolds.

5. The intelligent control method for a coal-fired boiler based on data analysis according to claim 1, characterized in that, The incremental update includes: detecting target running events based on the running state vector and the prediction error of the current digital twin model; when a target running event is detected, constructing high-value samples from the running state vector and control data of the corresponding time period, and determining the local region corresponding to the high-value samples in the joint space; updating the incremental parameters triggered by the execution event of the digital twin model based on the high-value samples, locally adjusting the manifold parameters and boundaries related to the local region in the multi-layered physically constrained manifold system, and updating the confidence weights for the local region, thereby realizing local incremental updates oriented towards the target event.

6. The intelligent control method for a coal-fired boiler based on data analysis according to claim 1, characterized in that, The counterfactual deduction includes: based on the type of the target event, calling the corresponding control strategy template from a pre-established scenario-based strategy library to generate multiple sets of candidate control instructions; under the combined conditions of the current running state vector and the candidate control instructions, calling the digital twin model to perform rolling prediction within a set prediction time window, outputting the corresponding prediction sequence as the performance prediction result, and generating a state-control trajectory corresponding to the prediction sequence; the state-control trajectory is a sequence composed of the running state vector and the candidate control instructions at each moment within the prediction time window.

7. The intelligent control method for a coal-fired boiler based on data analysis according to claim 1, characterized in that, The hierarchical security assessment includes: calculating the degree of violation of the state-control trajectory relative to the bottom manifold to obtain the physical constraint violation degree; calculating the spatial geometric deviation of the state-control trajectory in the bottom manifold relative to the middle manifold to obtain the security deviation degree; and calculating the distance of the state-control trajectory in the middle manifold relative to the top manifold to obtain the performance optimization potential index.

8. The intelligent control method for a coal-fired boiler based on data analysis according to claim 1, characterized in that, The multi-objective comprehensive evaluation includes: constructing a comprehensive evaluation function based on the performance prediction results and hierarchical evaluation indicators corresponding to candidate control commands. The performance items of the comprehensive evaluation function include main steam pressure and temperature fluctuations, combustion stability indicators, NOx emission levels, and ammonia slip levels. The constraint penalty items include physical constraint violation degree and safety deviation degree. The performance optimization potential indicator is used as a guiding term to approach the top manifold. Under the premise of comprehensively considering the performance items, constraint penalty items, and guiding terms, and satisfying the preset physical constraints, safety constraints, and operational constraints, a comprehensive score is calculated and sorted for each candidate control command. The candidate control command with the best comprehensive score is selected as the target control command.

9. The intelligent control method for a coal-fired boiler based on data analysis according to claim 1, characterized in that, The layered projection includes: adjusting the target control command to a physically feasible bottom-level manifold so that the adjusted control quantity satisfies the constraints of mass conservation, energy conservation, and momentum conservation; projecting the adjusted control command to a safe-operating middle-level manifold while satisfying the bottom-level manifold requirements, so that the state-control combination corresponding to the target control command falls into the safe-operating region; and optimizing the control command while satisfying the middle-level manifold requirements, so that its distance from the optimal top-level manifold in the joint space is reduced, thereby obtaining a corrected control command and outputting it for execution.

10. A data analysis-based intelligent control system for a coal-fired boiler, characterized in that, The system employs a data analysis-based intelligent control method for coal-fired boilers as described in any one of claims 1 to 9, comprising: Data acquisition module: Collects and preprocesses the operating data of the coal-fired boiler in real time to obtain the operating status vector; The digital twin update module constructs a digital twin model of a coal-fired boiler based on historical operating data and uses operating state vectors to correct the parameters of the digital twin model. Manifold System Module: Construct a spatiotemporally decoupled multi-layered physically constrained manifold system in the joint space of the operating state and control variables of the coal-fired boiler corresponding to the digital twin model; Event Update Module: Real-time detection of target events during boiler operation, and triggering incremental updates of parameters, boundaries, and confidence weights of relevant local regions in the digital twin model and multi-layered physical constraint manifold based on the detection results; Candidate evaluation module: Generates multiple sets of candidate control commands based on the type of the target event, and inputs them into the updated digital twin model for counterfactual inference to obtain performance prediction results; uses the state-control trajectory generated by the inference to map to the updated multi-layered physical constraint manifold system to perform hierarchical safety evaluation and obtain hierarchical evaluation indicators; The comprehensive evaluation module couples the performance prediction results with hierarchical evaluation indicators to perform a multi-objective comprehensive evaluation of each candidate control command, and selects the one with the highest comprehensive score as the target control command under preset constraints. Manifold Correction Module: Maps the target control command and the current operating state vector together to the joint space of the multi-layered physical constraint manifold system, performs layered projection, and the resulting corrected control command is output as the final control command of the coal-fired boiler.

Images