Coal scheduling method and system adaptive to unit load change

Abstract

The embodiment of the application provides a kind of coal scheduling method and system of adapting to unit load change, related to the field of coal scheduling into furnace, by real-time sensing key operating parameter, the real, variable transmission time lag of coal powder is predicted from coal to combustion in furnace.And with this dynamic delay as the core, the future heat demand curve corresponding to the unit load instruction is compensated in advance in time domain, and the optimal coal scheduling instruction at the current time is reversely solved.The scheme changes the control logic from the passive lagging adjustment based on experience to the active prediction compensation based on physical model, and through accurate prediction and hedging transmission delay, it ensures that the actual heat release rate of coal powder can be accurately synchronized with the real-time heat demand of the boiler when the unit load changes sharply, thereby fundamentally solving the main steam pressure fluctuation and over-temperature risk caused by delay misplacement.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of furnace coal scheduling technology, specifically to a method and system for scheduling coal feed into the furnace to adapt to changes in unit load. Background Technology

[0002] As modern power grids increasingly demand greater flexibility and responsiveness from power generation systems, coal-fired power generating units, the mainstay of power supply, frequently need to participate in deep peak shaving using Automatic Generation Control (AGC) to cope with rapid load fluctuations. In this process, maintaining the safe, stable, and economical operation of the units hinges on achieving precise and dynamic matching between boiler heat input and unit load demand. Therefore, constructing a coal dispatching scheme capable of adapting to rapid load changes is indispensable for ensuring grid frequency stability and improving unit response quality.

[0003] However, existing coal feeding scheduling strategies generally exhibit inherent limitations when dealing with rapid load changes in the unit. Most current DCS systems simplify the complex pulverizing and conveying system into a linear link with a fixed, large inertial time delay, assuming that the delay from the coal feeder issuing the command to the pulverized coal burning and releasing heat in the furnace is a static constant. But in reality, this dynamic delay is a non-linear variable that changes in real time with operating parameters such as primary air volume and the amount of pulverized coal stored in the mill. This disconnect between the control model and physical reality leads to a severe misalignment between scheduling commands and actual heat release during rapid load increases or decreases. When the actual delay exceeds the system's set fixed value, the pulverized coal supply will lag significantly behind heat demand, potentially causing a sudden drop in main steam pressure and affecting unit safety; conversely, if the actual delay is shortened, it will cause heat supply overshoot, leading to main steam temperature exceeding limits and posing a risk of overheating and tube rupture.

[0004] Therefore, existing technologies cannot perceive and predict the variable dynamic delay of the pulverizing system in real time, and can only make ex-post adjustments based on static empirical parameters. They have fundamental defects in terms of rapid response capability and control accuracy, and are difficult to meet the higher requirements of modern power grids for unit flexibility. Summary of the Invention

[0005] The present invention aims to at least solve one of the technical problems existing in the prior art, and provides a method and system for scheduling coal feed into the furnace to adapt to changes in unit load.

[0006] In a first aspect, embodiments of the present invention provide a method for scheduling coal feed into the furnace to adapt to changes in unit load, comprising: Based on the corrected weight matrix of the current period, the operating condition characteristics of the original sensor data stream of the current period are calibrated and extracted to obtain the calibration state matrix. The original sensor data stream of the current period includes the coal feeder speed, primary air volume, primary air pressure, coal mill inlet and outlet pressure difference, and real-time AGC load command of the unit. Dynamic delay scalars of the coal mill system are obtained by performing dynamic delay prediction on the state subsets in the calibration state matrix that represent the coal feeder speed, primary air volume and coal mill inlet and outlet pressure difference. The target heat feedforward demand is calculated by performing target heat demand curves on the real-time AGC load commands and their time derivatives in the calibration state matrix. Using a dynamic delay scalar as a time-domain translation operator, the heat demand curve is optimized by performing variable-delay reverse scheduling instructions to obtain the coal feeding scheduling vector.

[0007] Secondly, embodiments of the present invention also provide a coal feeding scheduling system adapted to changes in unit load, comprising: The operating condition calibration module is used to calibrate and extract the operating condition features of the raw sensor data stream of the current period based on the correction weight matrix of the current period to obtain the calibration state matrix. The raw sensor data stream of the current period includes the coal feeder speed, primary air volume, primary air pressure, coal mill inlet and outlet pressure difference, and real-time AGC load command of the unit. The dynamic delay prediction module is used to perform dynamic delay prediction of the coal milling system on the state subsets in the calibration state matrix that represent the coal feeder speed, primary air volume and coal mill inlet and outlet pressure difference to obtain the dynamic delay scalar. The heat demand curve generation module is used to perform target heat feedforward demand calculation on the real-time AGC load command and its time derivative of the unit in the calibration state matrix to obtain the heat demand curve. The coal feeding scheduling module is used to optimize the heat demand curve by applying variable-delay reverse scheduling instructions using a dynamic delay scalar as a time-domain shift operator to obtain the coal feeding scheduling vector.

[0008] The coal feeding scheduling method and system adapted to unit load changes in this invention, by real-time sensing of key operating parameters such as primary air volume and coal mill differential pressure, abandons the traditional static, fixed delay assumptions and instead accurately predicts the real, variable transmission time delay of pulverized coal from coal feeding to combustion in the furnace. Based on this dynamic delay, it performs time-domain advance compensation on the future heat demand curve corresponding to the unit load command, and inversely solves for the optimal coal feeding scheduling command at the current moment. This scheme transforms the control logic from experience-based passive lagging adjustment to active prediction and compensation based on a physical model. By accurately predicting and offsetting transmission delays, it ensures that the actual heat release rate of pulverized coal is precisely synchronized with the boiler's real-time heat demand when the unit load changes drastically, thereby fundamentally solving the risk of main steam pressure fluctuations and overheating caused by delay misalignment. Attached Figure Description

[0009] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0010] Figure 1 A flowchart illustrating a method for adjusting coal feed to adapt to unit load changes according to an embodiment of the present invention; Figure 2 This is a data flow diagram illustrating a method for scheduling coal feed into the furnace to adapt to changes in unit load, according to an embodiment of the present invention. Figure 3 This is a flowchart of step 1 in the coal dispatching method for adapting to unit load changes according to an embodiment of the present invention; Figure 4 This is a flowchart of step 2 in the coal dispatching method for adapting to unit load changes according to an embodiment of the present invention; Figure 5 This is a flowchart of step 3 in the coal dispatching method for adapting to unit load changes according to an embodiment of the present invention; Figure 6 This is a flowchart of step 4 in the method for scheduling coal feed into the furnace to adapt to changes in unit load according to an embodiment of the present invention. Detailed Implementation

[0011] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0012] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0013] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0014] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0015] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0016] Based on the shortcomings of the aforementioned technical fields, this invention proposes a method for scheduling coal feed into the furnace that adapts to changes in unit load. For example... Figure 1 and Figure 2 As shown, Figure 1 This is a flowchart of a method for scheduling coal feed into the furnace to adapt to changes in unit load, according to an embodiment of the present invention. Figure 2 This is a data flow diagram illustrating the coal feeding scheduling method for adapting to unit load changes according to an embodiment of the present invention. Specifically, step 1: Based on the corrected weight matrix of the current period, the operating condition characteristics of the original sensor data stream of the current period are calibrated and extracted to obtain a calibration state matrix. The original sensor data stream of the current period includes the feeder speed, primary air volume, primary air pressure, mill inlet and outlet pressure difference, and real-time AGC load command of the unit; step 2: The dynamic delay prediction of the coal mill system is performed on the state subsets representing the feeder speed, primary air volume, and mill inlet and outlet pressure difference in the calibration state matrix to obtain a dynamic delay scalar; step 3: The target heat feedforward demand is calculated on the real-time AGC load command of the unit and its time derivative in the calibration state matrix to obtain the heat demand curve; step 4: Using the dynamic delay scalar as a time-domain translation operator, the heat demand curve is optimized by variable delay reverse scheduling command to obtain the coal feeding scheduling vector. The process also includes: Step 5: Sending the current control step action command in the coal feeding scheduling vector to the underlying control system for execution, and simultaneously extracting the state parameters representing the actual boiler heat response results from the calibration state matrix to obtain the execution feedback vector; Step 6: Subtracting the expected theoretical heat supply represented by the coal feeding scheduling vector from the actual heat generated represented by the execution feedback vector to obtain the residual vector; Step 7: Updating the weights using the gradient descent algorithm with the goal of minimizing the residual vector to obtain the corrected weight matrix for the next control cycle.

[0017] Specifically, step 1: Based on the corrected weight matrix of the current period, the operating condition characteristics of the raw sensor data stream of the current period are calibrated and extracted to obtain the calibration state matrix. The raw sensor data stream of the current period includes coal feeder speed, primary air volume, primary air pressure, coal mill inlet and outlet pressure difference, and real-time AGC load commands of the unit. It should be understood that in the actual operating environment of a coal-fired unit, the raw data stream directly collected from field sensors, such as coal feeder speed and air pressure, is essentially a non-ideal signal superimposed with high-frequency measurement noise and low-frequency sensor static drift. Furthermore, these data originate from different physical measurement dimensions, with significant differences in their dimensions, units, and numerical ranges. If such raw, heterogeneous data is directly used as input to the subsequent dynamic delay prediction and heat demand calculation models without preprocessing, it will not only introduce interference and reduce model accuracy, but may also lead to numerical instability or even divergence in the algorithm's solution process due to the significant numerical differences. Therefore, performing a series of standardized preprocessing steps on the raw sensor data before performing core algorithm calculations is fundamental to ensuring the robustness and accuracy of the entire scheduling method. Step 1 involves constructing a mapping from the original measurement to the normalized state space. Through adaptive compensation, dimensional unification, and normalization, the mixed original data is transformed into a noise-free, order-of-magnitude aligned, and range-controlled calibration state matrix, laying a high-quality data foundation for subsequent accurate dynamic delay prediction and reverse optimization scheduling.

[0018] Figure 3 This is a flowchart of step 1 in the method for scheduling coal feed into the furnace to adapt to changes in unit load according to an embodiment of the present invention. Figure 3 As shown, in one embodiment of step 1, the steps include: Sub-step 1.1: Based on the correction weight matrix of the current period, compensation is performed on the measurement noise and sensor static drift in the original sensor data stream of the current period to obtain a filtered data vector; Sub-step 1.2: Based on the pre-calibrated reference scaling diagonal matrix, linear projection and scaling are performed on the filtered data vector to obtain an alignment feature vector; Sub-step 1.3: Pre-defined upper and lower limit vectors of physical extrema are extracted, and minimax normalization mapping is performed on each element in the alignment feature vector to obtain a calibration state matrix.

[0019] The specific processing is as follows: The process begins with receiving two key inputs: the raw sensor data stream for the current period and the corrected weight matrix for the current period. The raw sensor data stream for the current period is a real-time acquired column vector, the dimension of which is determined by the monitored physical quantity. In this embodiment, the vector consists of five key operating parameters: feeder speed, primary air volume, primary air pressure, mill inlet and outlet pressure difference, and real-time AGC load command of the unit. It is acquired in real time from the distributed control system (DCS) in each control cycle, for example, every two seconds, reflecting the unit's current raw operating state. The corrected weight matrix for the current period is a square matrix with the same dimension as the raw sensor data stream vector. This matrix carries the correction information fed back from the previous control cycle and is used to dynamically compensate the raw data for the current period. This matrix is ​​iteratively updated by the adaptive learning algorithm in step 7 with the goal of minimizing the residual, thereby achieving continuous tracking and calibration of system changes. In the initial cycle when the system is first started, since there is no historical feedback information, the corrected weight matrix is ​​initialized as a unit diagonal matrix, that is, no compensation is applied to the raw data.

[0020] Sub-step 1.1 aims to compensate the raw data stream using the corrected weight matrix of the current cycle, filtering out measurement noise and sensor static drift. In industrial settings, sensors may experience zero-point drift due to long-term operation, or electromagnetic interference may be introduced into the signal transmission link, causing the acquired data to deviate from the true value. The corrected weight matrix is ​​designed to address this issue. The main diagonal elements of this matrix are primarily used to correct the sensitivity drift of each sensor, while the off-diagonal elements are used to decouple cross-interference between different measurement variables. The processing is achieved by performing matrix-vector multiplication operations.

[0021] Next, sub-step 1.2 is executed, the purpose of which is to address the problem of vastly different physical dimensions and orders of magnitude among the elements in the filtered data vector. Directly processing data with huge differences, such as rotational speed (hundreds of thousands) and wind pressure (thousands of thousands), would unreasonably assign higher weights to variables with larger values ​​in subsequent calculations, resulting in ill-conditioned numerical values. To address this, a pre-calibrated benchmark scaling diagonal matrix is ​​introduced. This matrix is ​​a diagonal matrix pre-calculated and fixed in the system based on the statistical characteristics of a large amount of historical operating data. Each element on its diagonal is a scaling factor for the corresponding physical quantity, set as the reciprocal of the standard deviation of the historical operating data of that physical quantity, aiming to project all variables onto a scale space with a variance approximately equal to 1. This linear projection and scaling operation essentially performs a standard deviation normalization on the data, eliminating the influence of different physical dimensions, making all features on a comparable numerical scale, generating an aligned feature vector. Each element in this vector is dimensionless, and its value fairly reflects its deviation from its normal fluctuation range.

[0022] Finally, sub-step 1.3 is executed, the core task of which is to normalize the alignment feature vector and construct the final calibration state matrix. To meet the strict requirements of subsequent models such as state observers or neural networks that the input data is within the [0,1] interval, a minimax normalization mapping needs to be performed on each element of the alignment feature vector. This process requires a preset physical extreme value upper and lower limit vector, which defines the absolute boundary of each physical quantity within the safe operating range. This vector is determined based on the unit equipment design manual, historical extreme operating condition data, and operating procedures, and it contains two parts: a minimum value vector and a maximum value vector. After normalizing all elements, a normalized column vector is obtained, where each element is between 0 and 1. To facilitate subsequent steps, especially in scenarios requiring multi-dimensional decoupling operations, this one-dimensional column vector needs to be converted into a two-dimensional matrix. Specifically, the elements of the normalized column vector are used as the main diagonal elements to construct a diagonal matrix, with all other elements being zero. This final generated matrix is ​​the calibration state matrix.

[0023] Specifically, step 2 involves predicting the dynamic delay scalar of the pulverizing system by performing dynamic delay prediction on the subset of states in the calibration state matrix that represent the feeder speed, primary air volume, and pulverizer inlet and outlet pressure difference. Correspondingly, in the process of adjusting the coal feed scheduling to adapt to changes in unit load, although step 1 has already transformed the raw, mixed sensor data into a standardized calibration state matrix, this only solves the data preprocessing problem. The matrix itself is merely a static snapshot of all current operating conditions of the unit and does not reveal the core dynamic characteristics of the pulverizing system. Background technology has clarified that the main deficiency of existing technologies lies in the inability to accurately and in real-time grasp the variable transmission delay of pulverized coal from feed to furnace. Without quantifying this crucial dynamic parameter, any feedforward control will be blind and unable to achieve accurate pre-compensation for load changes. Therefore, step 2 is performed to solve this core technical problem by performing in-depth analysis and physical modeling on a specific subset of states in the calibration state matrix, thereby predicting the dynamic delay scalar of the pulverizing system under the current operating conditions in real time.

[0024] Figure 4 This is a flowchart of step 2 in the method for scheduling coal feed into the furnace to adapt to changes in unit load according to an embodiment of the present invention. Figure 4As shown, in one embodiment of step 2, the steps include: sub-step 2.1: extracting a subset of states representing the feeder speed, primary air volume, and mill inlet and outlet pressure difference from the calibration state matrix to obtain the mill state vector; sub-step 2.2: inputting the mill state vector into a discrete-time differential observer based on the law of mass conservation to obtain a two-phase flow delay solution parameter vector; sub-step 2.3: performing dynamic time delay calculation on the gas-solid two-phase flow transport dynamic delay of the two-phase flow delay solution parameter vector to obtain a dynamic delay scalar.

[0025] The specific processing is as follows: The starting point of this process is receiving the calibration state matrix generated in step 1. As mentioned earlier, this matrix is ​​a diagonal matrix, and its diagonal elements are multiple operating parameters normalized to the [0,1] interval. First, sub-step 2.1 is executed, the task of which is to accurately extract the features most closely related to the dynamic characteristics of the coal mill system from this global state matrix, which contains numerous operating conditions across the entire plant. To avoid introducing irrelevant variables in the subsequent solution of differential equations, which would lead to the curse of dimensionality and a surge in computational complexity, a preset dimensionality reduction selection matrix for the coal mill system is used here. This selection matrix is ​​a sparse matrix, and its structure and content are determined based on prior knowledge of the physical processes of the pulverizing system, aiming to select the variables of interest from the global state. In this embodiment, the physical quantities of interest are the feeder speed, primary air volume, and the inlet and outlet pressure difference of the coal mill. If the calibration state matrix is ​​a 5×5 diagonal matrix, and its diagonal elements correspond to these five quantities in sequence, then the dimensionality reduction selection matrix for the coal mill system can be constructed as a 3×5 matrix. By performing matrix multiplication and multiplying the result by a single-column vector, the features of the diagonal matrix can be compressed into a high-density column vector. This process not only achieves feature dimensionality reduction but also focuses the computational attention on the core physical state of the coal mill system. The final output is a three-dimensional column vector, which is the coal mill state vector, whose three components represent the normalized feeder speed, primary air volume, and coal mill inlet and outlet pressure difference at the current moment.

[0026] Next, sub-step 2.2 is executed. The core of this step lies in solving a key state variable inside the coal mill that cannot be directly measured—the real-time coal inventory. The amount of coal inventory directly affects the conveying resistance of pulverized coal in the primary air duct and is the core factor determining the transmission delay. To this end, a discrete-time differential observer based on the law of conservation of mass is established. The essence of this observer is a state estimator, which follows the fundamental physical law that current coal inventory = previous coal inventory + input coal quantity - output coal quantity. The observer uses the feeder speed component in the coal mill state vector as a measure of mass input, because the speed is proportional to the coal feeding rate per unit time. At the same time, it uses the product of the primary air volume and the pressure difference between the inlet and outlet of the coal mill as a measure of mass output, because the primary air is the carrier for conveying pulverized coal, while the pressure difference reflects the concentration and fluidization state of pulverized coal in the duct, and the product of the two can characterize the rate at which pulverized coal is carried out of the coal mill. By integrating and accumulating the net input mass in each discrete control cycle Δt, the current coal inventory inside the coal mill can be deduced in real time. Subsequently, in order to ensure that downstream steps can obtain the complete information required for delay calculation, the latest calculated coal inventory is combined with the current primary air volume component to generate a two-dimensional column vector, which is the two-phase flow delay solution parameter vector.

[0027] Finally, sub-step 2.3 is executed, using the two-phase flow delay solution parameter vector obtained in the previous step to calculate the final dynamic delay scalar. Based on the fluid dynamics theory of gas-solid two-phase flow, the transmission delay of pulverized coal in the primary air duct is not a static constant. It is positively correlated with the conveying resistance formed by the amount of pulverized coal already in the duct (i.e., the amount of coal stored), meaning that the more coal stored, the slower the subsequent pulverized coal passes through, and the longer the delay. At the same time, it is negatively correlated with the kinetic energy of the primary air (i.e., the air volume), meaning that the higher the air velocity, the faster the conveying and the shorter the delay. This complex nonlinear relationship can be accurately characterized by an exponential function with a smoothing factor.

[0028] Specifically, step 3: Calculate the target heat feedforward demand for the unit's real-time AGC load command and its time derivative in the calibration state matrix to obtain the heat demand curve. It is understandable that after accurately predicting the dynamic delay of the pulverizing system, the coal dispatching method has solved the problem of when to respond. However, simply knowing the timing of the response is insufficient; the more crucial issue is determining the magnitude of the response, i.e., how much heat the boiler needs at each moment to match the load command issued by the power grid. The unit's real-time AGC load command is a target representing the output electrical power, while the boiler combustion system responds with the input thermal power. The relationship between these two is not a simple linear proportional relationship because the boiler itself is a complex thermodynamic system with enormous thermal inertia. When the load changes rapidly, in addition to providing the heat required to maintain steady-state operation under the new load, additional heat is needed to heat the boiler's metal components and internal steam-water working fluid; otherwise, the main steam pressure cannot be maintained, which could seriously endanger the unit's safety. Therefore, step 3 is performed to build this translation bridge. By performing in-depth analysis of the unit load command and its dynamic trends based on physical energy balance, a smooth and continuous future target heat demand curve can be obtained.

[0029] Figure 5 This is a flowchart of step 3 in the method for scheduling coal feed into the furnace to adapt to changes in unit load according to an embodiment of the present invention. Figure 5 As shown, in one embodiment of step 3, the steps include: Sub-step 3.1: Based on the load selection mask vector, performing scheduling command feature stripping and dynamic trend extraction on the calibration state matrix to obtain a load command vector, which includes the absolute value and rate of change of the load command; Sub-step 3.2: Performing reverse demand deduction based on the steam-water energy balance on the absolute value and rate of change of the load command in the load command vector to obtain a heat release rate sequence; Sub-step 3.3: Using the time series within the prediction window as the independent variable, performing cubic spline continuous smoothing on the heat release rate sequence to obtain a heat demand curve.

[0030] The specific processing is as follows: First, sub-step 3.1 is executed, the task of which is to accurately extract the real-time AGC load command components of the units directly related to the scheduling target from the multi-dimensional calibration state matrix and capture their dynamic changing trends. To achieve accurate feature extraction, a dedicated load selection mask vector is used here. This vector is a preset row vector with the same number of columns as the calibration state matrix, and most of its internal elements are zero, except for the positions corresponding to the load command components, which are 1. For example, if the fifth element in the five diagonal elements of the calibration state matrix represents the normalized load command, then the load selection mask vector can be set to [0,0,0,0,1]. By performing matrix multiplication with the calibration state matrix using this mask vector, the current normalized value of the load command can be uniquely extracted. Subsequently, in order to capture the dynamic ramp-up or slip-down trend of the power grid scheduling, it is necessary to calculate the transient time derivative of the load command, i.e., the rate of change. The calculation is implemented using a first-order backward difference algorithm, which involves subtracting the historical load command value stored in the previous control cycle from the current load command value, and then dividing by the control cycle duration Δt. Finally, the extracted absolute value of the load command and the calculated rate of change are vertically concatenated to form a two-dimensional column vector, which is the load command vector. If the normalized load command value extracted from the calibration state matrix at the current time t is 0.625, and the stored value at the previous time t-Δt, such as 2 seconds ago, is 0.600, then the rate of change is calculated as (0.625-0.600) / 2=0.0125 (unit normalized value / second). The resulting load command vector is [0.625, 0.0125]T, which fully represents the energy demand situation where the unit not only needs to maintain a load level of approximately 62.5%, but is also climbing upward at a certain rate.

[0031] Next, sub-step 3.2 is executed to perform inverse demand extrapolation based on steam-water energy balance on the load command vector generated in the previous step, in order to obtain the heat release rate sequence over a future period. During rapid load changes in the unit, the heat provided by the furnace needs to meet two demands: first, the static heat used to heat, evaporate, and superheat the feedwater to drive the turbine, the magnitude of which is proportional to the absolute value of the load command; second, the additional dynamic heat storage required to overcome the huge thermal inertia of the boiler body metal and internal fluid medium when the load increases, the magnitude of which is proportional to the rate of load change. Based on this physical principle, through a pre-constructed nonlinear dynamic energy balance equation for the boiler, the instantaneous furnace heat release power necessary to maintain stable main steam parameters is inversely extrapolated within a preset forward prediction window, such as covering the next 200 seconds.

[0032] Finally, sub-step 3.3 is executed, which transforms the discrete, stepped heat release rate sequence generated in the previous step into a physically achievable, smooth, and continuous curve. Directly using the discrete heat release rate sequence as the control target will cause actuators such as the coal feeder to make step adjustments in each control cycle, resulting in severe mechanical shocks and control quantity chattering, which seriously affects equipment life and control quality.

[0033] Specifically, step 4: Using the dynamic delay scalar as a time-domain shift operator, the heat demand curve is optimized using a variable-delay reverse scheduling command to obtain the coal feeding scheduling vector. It should be understood that in the previous steps, the dynamic delay of the pulverizing system was successfully predicted in real time, and the ideal heat demand curve of the boiler for a future period was accurately calculated. However, these two key pieces of information are misaligned in the time dimension: heat demand is a future-oriented goal, while control commands must be issued at the current moment. If the current heat demand is simply converted into a coal feeding command, due to the physical transmission delay, by the time this portion of pulverized coal is actually put into combustion, the unit's load demand may have already changed, leading to severe control lag and overshoot. Therefore, step 4 is executed to pre-compensate the heat demand curve by using the dynamic delay as a time-domain shift operator, and to construct a quadratic programming optimization space with multiple physical constraints, thereby inversely solving for a physically feasible coal feeding scheduling vector that can be issued at the current moment and accurately meet the heat demand in the future.

[0034] Figure 6 This is a flowchart of step 4 in the method for scheduling coal feed into the furnace to adapt to changes in unit load according to an embodiment of the present invention. Figure 6 As shown, in one embodiment of step 4, the steps include: Sub-step 4.1: Using the dynamic delay scalar as a time-domain translation operator, subtracting the dynamic delay scalar from the independent variable of the heat demand curve on the time axis to obtain the translation demand curve; Sub-step 4.2: Constructing an objective function based on the translation demand curve, with the goal of minimizing the sum of squares of the difference between the expected coal supply heat and the translation demand curve, and penalizing drastic changes in control actions, extracting the absolute upper and lower limits of the physical speed of the coal feeder and the maximum ramp-down rate of the control actuator, and converting them into a multidimensional linear inequality constraint matrix to obtain the optimization space; Sub-step 4.3: Solving the optimization space using the interior point method to obtain the coal feeding scheduling vector.

[0035] The specific processing is as follows: First, sub-step 4.1 is executed. Its core task is to use a dynamic delay scalar to shift the heat demand curve in reverse along the time axis to actively compensate for physical transmission delays. Mathematically, this operation is equivalent to bringing the future heat demand target back to the present, thereby guiding the current control actions.

[0036] After obtaining the time-aligned translation demand curve, sub-step 4.2 is executed, which involves constructing a standard optimization problem solvable by a mathematical optimizer. This requires defining an objective function and a series of constraints, which together constitute the optimization space. The objective function needs to consider two core objectives: first, accurately tracking the translation demand curve; and second, ensuring the smoothness of control actions to protect the physical equipment. Therefore, the objective function is constructed as a weighted multi-objective functional. Its first part is the sum of squares of the differences between the expected coal supply heat and the target value of the translation demand curve over a future period, i.e., within the control window, used to minimize the tracking error. The second part is the change in control actions, i.e., the sum of squares of the adjustments in the feeder speed during adjacent control cycles, used to penalize drastic control actions. Meanwhile, any practical control system is limited by its physical capabilities. Therefore, the physical constraints of the feeder need to be transformed into mathematical inequalities. These constraints include: the absolute upper and lower limits of the physical speed of the coal feeder, for example, according to the equipment nameplate and operating procedures, the speed must be between 0 rpm and 600 rpm; and the maximum ramp and descent rate of the control actuator, for example, due to the inertia of the motor and frequency converter, the change in speed per second must not exceed 10 rpm. These constraints are uniformly expressed as a multidimensional linear inequality constraint matrix.

[0037] Finally, sub-step 4.3 is executed to efficiently solve the quadratic programming optimization space constructed in the previous step. Since this problem is a typical convex optimization problem (the objective function is quadratic, and the constraints are linear), it has a unique optimal solution. Here, an efficient and robust numerical optimization algorithm—the interior-point method with a logarithmic barrier function—is used for solving it. The core idea of ​​this method is to transform the original optimization problem with a series of inequality constraints into an approximate, unconstrained optimization problem. This is achieved by adding a barrier term to the original objective function. The barrier term consists of the logarithms of all constraint functions; as the solution approaches the constraint boundary, the value of this term increases sharply towards infinity, thus preventing the solution from crossing the boundary.

[0038] Specifically, step 5 involves sending the current control step action command from the coal feeding scheduling vector to the underlying control system for execution, and simultaneously extracting the state parameters representing the actual boiler thermal response from the calibration state matrix to obtain the execution feedback vector. Correspondingly, although step 4, through complex optimization calculations, yields a theoretically optimal coal feeding scheduling vector under the current information, this is merely a decision-making link in the control loop. The effectiveness of any model-based predictive control must ultimately be verified and closed-loop through actual execution and feedback in the physical world. Due to the inevitable mismatch between the model and the actual process (such as unmodeled fluctuations in coal quality, temporary interference from sensors, etc.), and the existence of external disturbances, simply issuing open-loop commands is insufficient. Therefore, step 5 constructs a comprehensive execution feedback vector containing control input, theoretical expected output, and actual feedback output by issuing and executing the optimal scheduling command and capturing the boiler's actual thermal response to the command in real time. This provides an indispensable data foundation for subsequent residual calculation and model adaptive updates.

[0039] In one embodiment of step 5, the specific processing is as follows: First, following the basic principle of rolling optimization control, the first coal feeding action command corresponding to the current control step, k=1, is precisely extracted from the input coal feeding scheduling vector (i.e., a sequence containing multiple future control steps). This command is a scalar representing the optimal setpoint value of the coal feeder speed at the current moment. For example, if the optimal coal feeding scheduling vector V calculated in step 4... (fs,t)=[452.1,455.3,458.0,…]T (unit: revolutions per minute), then only the first element, 452.1, is extracted as the action command for the current moment, and it is sent to the underlying DCS or PLC actuator through the communication interface to drive the coal feeder frequency converter to adjust the speed to the set value. Simultaneously with the command, to facilitate subsequent effect evaluation, this physical speed command needs to be converted into its theoretically generated heat. This conversion is based on a priori mechanism model, namely, thermal power equals the coal feeding rate multiplied by the calorific value of the coal. The coal feeding rate is related to the coal feeder speed and the coal bed thickness. In the simplified model, it can be regarded as a linear function of the speed. At the same time, the actual boiler thermal response results are obtained from the feedback of the physical world. Since the instantaneous total heat release in the furnace cannot be directly measured, it is characterized by some indirect core state parameters. In this embodiment, the main steam pressure and its rate of change are selected as the core characterization of the actual generated heat. These state parameters are already included in the calibration state matrix. Through matrix trace operations or specific vectorization operations, the normalized value of the corresponding main steam pressure and its time derivative can be extracted from the matrix. Subsequently, these pressure-related state parameters can be used to estimate the actual heat generated at the current moment. Finally, to ensure that subsequent steps can acquire a complete data stream and form an information loop, the key information involved in this step is vertically concatenated to generate a comprehensive execution feedback vector. This vector contains at least the following parts: the currently issued coal feeding action command of 452.1 rpm, the theoretical expected heat supply calculated from this command of 1582.4 MJ / s, the actual heat generated estimated from the real state parameters, and some important global state features added to give the subsequent adaptive algorithm a global perspective (such as load commands and air volume extracted from the calibration state matrix).

[0040] Specifically, step 6: Subtract the expected theoretical heat supply represented by the coal feeding scheduling vector from the actual heat generated represented by the execution feedback vector to obtain the residual vector. It is understandable that in the previous steps, the theoretically optimal scheduling command has been successfully issued to the physical world for execution, and the actual response state of the system has been acquired simultaneously. However, this only completes one open-loop execution and observation of the control loop. The long-term effectiveness of any model-based advanced control strategy rests on its ability to continuously self-correct. Due to unforeseen factors such as coal quality fluctuations, equipment wear, and environmental changes, there will inevitably be deviations between the prediction model and the actual physical process. If this deviation cannot be accurately quantified, the control system will be unable to perceive its own performance degradation, let alone make adaptive adjustments, and will ultimately degenerate into a static, unadaptable open-loop controller. Therefore, step 6 obtains a residual vector that clearly represents the current model prediction accuracy by accurately subtracting the expected theoretical heat supply from the actual heat generated.

[0041] In one embodiment of step 6, the specific processing is as follows: First, the coal feeding scheduling vector and execution feedback vector are indexed and parsed to accurately extract two key scalar information: one is the expected theoretical heat supply calculated based on the currently issued coal feeding command, and the other is the actual heat generated estimated based on the boiler's actual state parameters. Before making a direct comparison, the measurement delay and dynamic response characteristics commonly found in industrial field sensors need to be considered. If a theoretical value calculated instantaneously is directly subtracted from a measured value with dynamic lag characteristics, false, high-frequency error signals may be generated due to slight time asynchrony, interfering with the subsequent adaptive learning process. To solve this problem, a first-order dynamic forgetting factor is introduced to provide temporal smoothing compensation for the measured value of the actual heat generated.

[0042] Specifically, step 7: The weights are updated using the gradient descent algorithm with the goal of minimizing the residual vector, resulting in the corrected weight matrix for the next control cycle. In other words, after accurately quantifying the model prediction error, the entire control and feedback chain has formed a complete information loop. However, if the residuals are merely calculated without being utilized, the system lacks the ability to self-evolve and adapt to change. A coal-fired power unit is a complex industrial system operating for a long time, and its internal characteristics change slowly and steadily over time. For example, the calorific value of coal fluctuates between batches, and the grinding efficiency of the mill rollers decreases due to long-term wear. These slowly changing factors cause the parameters preset in the model to gradually deviate from the actual physical process, thus causing the residuals calculated in the previous steps to persist or even expand. Therefore, step 7 is executed to construct this core adaptive learning loop. By minimizing the thermal residuals, the gradient descent algorithm is used to iteratively update the corrected weight matrix used for data preprocessing in step 1, thereby achieving long-term robustness and high-precision performance of the entire scheduling method.

[0043] In one embodiment of step 7, the specific processing is as follows: the input to this process is the residual vector generated in step 6. Its core task is to eliminate long-term model mismatch caused by factors such as gradual changes in coal quality and equipment wear. This embodiment uses the stochastic gradient descent (SGD) algorithm for online closed-loop adaptive learning. First, a clear optimization objective needs to be defined, namely the cost function. Here, the cost function is set to minimize the square of the thermal residual in the residual vector. The square form is chosen because it not only ensures that the cost function is non-negative, but also its derivative is simple to calculate, and it can smoothly converge to the optimal solution. Next, the gradient of the cost function with respect to the parameters that need to be updated, namely the modified weight matrix in step 1, needs to be calculated. According to the chain rule, this gradient can be decomposed into the product of the derivative of the cost function with respect to the residual, the derivative of the residual with respect to the actual heat, the derivative of the actual heat with respect to the state variables, and the derivative of the final state variables with respect to the modified weight matrix. Under the simplified framework of stochastic gradient descent, the partial derivative gradient of the cost function with respect to the system state parameters can be directly calculated. First, the error term and the system state term at the time the error occurred (i.e., the global state feature part in the execution feedback vector) are accurately extracted from the input residual vector. Then, the gradient of the cost function with respect to the currently resident modified weight matrix in memory is calculated. In actual calculations, this gradient can be approximated as the product of the residual term and the input state vector. After obtaining the gradient, the modified weight matrix can be updated. The update rule follows the basic idea of ​​gradient descent, adjusting in small steps along the opposite direction of the gradient. The calculated gradient increment is multiplied by a preset learning rate η, and then subtracted from the current cycle's weight baseline matrix resident in system memory, thereby deducing the optimized modified weight matrix suitable for the next control cycle.

[0044] In summary, the coal feeding scheduling method adapted to unit load changes based on the embodiments of the present invention has been clarified. It abandons the traditional static, fixed-delay assumptions by real-time sensing of key operating parameters such as primary air volume and mill pressure differential, and instead accurately predicts the actual, variable transmission time delay of pulverized coal from coal feeding to combustion in the furnace. Using this dynamic delay as the core, it performs advance compensation in the time domain for the future heat demand curve corresponding to the unit load command, and inversely solves for the optimal coal feeding scheduling command at the current moment. This scheme transforms the control logic from experience-based passive lagging adjustment to active prediction and compensation based on a physical model. By accurately predicting and offsetting transmission delays, it ensures that the actual heat release rate of pulverized coal can be precisely synchronized with the real-time heat demand of the boiler when the unit load changes drastically, thereby fundamentally solving the risk of main steam pressure fluctuations and overheating caused by delay misalignment.

[0045] According to an embodiment of the present invention, a coal feeding scheduling system adapted to unit load changes includes: an operating condition calibration module, used to calibrate and extract operating condition features from the current period's original sensor data stream based on a correction weight matrix of the current period to obtain a calibration state matrix, wherein the current period's original sensor data stream includes feeder speed, primary air volume, primary air pressure, mill inlet and outlet pressure difference, and real-time AGC load command of the unit; a dynamic delay prediction module, used to perform dynamic delay prediction of the coal mill system on the state subsets representing feeder speed, primary air volume, and mill inlet and outlet pressure difference in the calibration state matrix to obtain a dynamic delay scalar; a heat demand curve generation module, used to perform target heat feedforward demand calculation on the real-time AGC load command of the unit and its time derivative in the calibration state matrix to obtain a heat demand curve; and a coal feeding scheduling module, used to use the dynamic delay scalar as a time-domain translation operator to optimize the heat demand curve using variable delay reverse scheduling commands to obtain a coal feeding scheduling vector. It also includes: an execution feedback module, used to send the current control step action command in the coal feeding scheduling vector to the underlying control system for execution, and simultaneously extract the state parameters representing the actual boiler heat response results in the calibration state matrix to obtain the execution feedback vector; a residual calculation module, used to subtract the expected theoretical heat supply represented by the coal feeding scheduling vector from the actual heat generated represented by the execution feedback vector to obtain the residual vector; and a corrected weight matrix generation module, used to update the weights using the gradient descent algorithm with the goal of minimizing the residual vector, to obtain the corrected weight matrix for the next control cycle.

[0046] Furthermore, the operating condition calibration module is specifically used to: compensate for measurement noise and sensor static drift in the original sensor data stream of the current period based on the correction weight matrix of the current period to obtain a filtered data vector; perform linear projection and scaling on the filtered data vector based on a pre-calibrated reference scaling diagonal matrix to obtain an alignment feature vector; extract the preset physical extreme value upper and lower limit vectors, and perform maximum and minimum normalization mapping on each element in the alignment feature vector to obtain a calibration state matrix.

[0047] Furthermore, the dynamic delay prediction module is specifically used to: extract a subset of states representing the feeder speed, primary air volume, and mill inlet and outlet pressure difference from the calibration state matrix to obtain a mill state vector; input the mill state vector into a discrete-time differential observer based on the law of mass conservation to obtain a two-phase flow delay solution parameter vector; and perform dynamic time delay calculation on the gas-solid two-phase flow transport dynamic delay of the two-phase flow delay solution parameter vector to obtain a dynamic delay scalar.

[0048] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for scheduling coal feed into the furnace to adapt to changes in unit load, characterized in that, Includes the following steps: Based on the corrected weight matrix of the current period, the operating condition characteristics of the original sensor data stream of the current period are calibrated and extracted to obtain the calibration state matrix. The original sensor data stream of the current period includes the coal feeder speed, primary air volume, primary air pressure, coal mill inlet and outlet pressure difference, and real-time AGC load command of the unit. Dynamic delay scalars of the coal mill system are obtained by performing dynamic delay prediction on the state subsets in the calibration state matrix that represent the coal feeder speed, primary air volume and coal mill inlet and outlet pressure difference. The target heat feedforward demand is calculated by performing target heat demand curves on the real-time AGC load commands and their time derivatives in the calibration state matrix. Using a dynamic delay scalar as a time-domain translation operator, the heat demand curve is optimized by performing variable-delay reverse scheduling instructions to obtain the coal feeding scheduling vector.

2. The method for adjusting coal feed to adapt to unit load changes according to claim 1, characterized in that, Also includes: The current control step action command in the coal feeding scheduling vector is sent to the underlying control system for execution, and the state parameters characterizing the actual boiler heat response result in the calibration state matrix are extracted simultaneously to obtain the execution feedback vector. The residual vector is obtained by subtracting the expected theoretical heat supply represented by the coal feeding scheduling vector from the actual heat generated represented by the execution feedback vector. The weights are updated using the gradient descent algorithm with the goal of minimizing the residual vector, resulting in a corrected weight matrix for the next control cycle.

3. The method for adjusting coal feed to adapt to unit load changes according to claim 1, characterized in that, Based on the corrected weight matrix for the current period, the operating condition characteristics of the original sensor data stream for the current period are calibrated and extracted to obtain the calibration state matrix, including: Based on the corrected weight matrix of the current period, the measurement noise and sensor static drift in the original sensor data stream of the current period are compensated to obtain the filtered data vector. The filtered data vector is linearly projected and scaled based on a pre-calibrated reference scaling diagonal matrix to obtain an aligned feature vector; Extract the preset upper and lower limit vectors of physical extrema, and perform maximum and minimum normalization mapping on each element in the aligned feature vector to obtain the calibration state matrix.

4. The method for adjusting coal feed to adapt to unit load changes according to claim 1, characterized in that, Dynamic delay scalars of the coal mill system are obtained by performing dynamic delay prediction on the state subsets representing the feeder speed, primary air volume, and coal mill inlet and outlet pressure difference in the calibration state matrix, including: Extract the state subsets representing the feeder speed, primary air volume and the inlet and outlet pressure difference of the coal mill from the calibration state matrix to obtain the coal mill state vector; The state vector of the coal mill is input into a discrete-time differential observer based on the law of conservation of mass to obtain the parameter vector for solving the two-phase flow delay. The dynamic time delay scalar of gas-solid two-phase flow is obtained by solving the parameter vector of the two-phase flow delay solution.

5. The method for adjusting coal feed to adapt to unit load changes according to claim 1, characterized in that, The target heat feedforward demand is calculated from the real-time AGC load commands and their time derivatives in the calibration state matrix to obtain the heat demand curve, including: Based on the load selection mask vector, the load command vector is obtained by stripping scheduling command features and extracting dynamic trends from the calibration state matrix. The load command vector includes the absolute value and rate of change of the load command. The absolute value and rate of change of the load command in the load command vector are used to perform inverse demand deduction based on the steam-water energy balance to obtain the heat release rate sequence; Using the time series within the prediction window as the independent variable, the heat release rate series is smoothed by cubic spline continuum to obtain the heat demand curve.

6. The method for adjusting coal feed to adapt to unit load changes according to claim 1, characterized in that, Using a dynamic delay scalar as the time-domain translation operator, the heat demand curve is optimized using variable-delay reverse scheduling commands to obtain the coal feeding scheduling vector, including: The dynamic delay scalar is used as a time-domain translation operator. The independent variable of the heat demand curve is subtracted from the dynamic delay scalar on the time axis to obtain the translated demand curve. Based on the translation demand curve, an objective function is constructed with the goal of minimizing the sum of squares of the difference between the expected coal supply heat and the translation demand curve, and penalizing drastic changes in control actions. The absolute upper and lower limits of the physical speed of the coal feeder and the maximum climbing and sliding rate of the control actuator are extracted and transformed into a multidimensional linear inequality constraint matrix to obtain the optimization space. The coal feeding scheduling vector is obtained by solving the optimization space using the interior point method.

7. A coal feeding dispatching system adaptable to changes in unit load, characterized in that, include: The operating condition calibration module is used to calibrate and extract the operating condition features of the raw sensor data stream of the current period based on the correction weight matrix of the current period to obtain the calibration state matrix. The raw sensor data stream of the current period includes the coal feeder speed, primary air volume, primary air pressure, coal mill inlet and outlet pressure difference, and real-time AGC load command of the unit. The dynamic delay prediction module is used to perform dynamic delay prediction of the coal milling system on the state subsets in the calibration state matrix that represent the coal feeder speed, primary air volume and coal mill inlet and outlet pressure difference to obtain the dynamic delay scalar. The heat demand curve generation module is used to perform target heat feedforward demand calculation on the real-time AGC load command and its time derivative of the unit in the calibration state matrix to obtain the heat demand curve. The coal feeding scheduling module is used to optimize the heat demand curve by applying variable-delay reverse scheduling instructions using a dynamic delay scalar as a time-domain shift operator to obtain the coal feeding scheduling vector.

8. The coal feeding dispatching system adapting to unit load changes according to claim 7, characterized in that, Also includes: The execution feedback module is used to send the current control step action command in the coal feeding scheduling vector to the underlying control system for execution, and simultaneously extract the state parameters that characterize the actual boiler heat response results in the calibration state matrix to obtain the execution feedback vector. The residual calculation module is used to subtract the expected theoretical heat supply represented by the coal feeding scheduling vector from the actual heat generated represented by the execution feedback vector to obtain the residual vector. The modified weight matrix generation module is used to update the weights using the gradient descent algorithm with the goal of minimizing the residual vector, so as to obtain the modified weight matrix for the next control cycle.

9. The coal dispatching system for adapting to unit load changes according to claim 7, characterized in that, The operating condition calibration module is further used for: Based on the corrected weight matrix of the current period, the measurement noise and sensor static drift in the original sensor data stream of the current period are compensated to obtain the filtered data vector. The filtered data vector is linearly projected and scaled based on a pre-calibrated reference scaling diagonal matrix to obtain an aligned feature vector; Extract the preset upper and lower limit vectors of physical extrema, and perform maximum and minimum normalization mapping on each element in the aligned feature vector to obtain the calibration state matrix.

10. The coal feeding dispatching system adapting to unit load changes according to claim 7, characterized in that, The dynamic delay prediction module is further used for: Extract the state subsets representing the feeder speed, primary air volume and the inlet and outlet pressure difference of the coal mill from the calibration state matrix to obtain the coal mill state vector; The state vector of the coal mill is input into a discrete-time differential observer based on the law of conservation of mass to obtain the parameter vector for solving the two-phase flow delay. The dynamic time delay scalar of gas-solid two-phase flow is obtained by solving the parameter vector of the two-phase flow delay solution.

Images