Control optimization method for optimal output condition of coal mill

Abstract

The present application relates to the field of coal mill control, and discloses a control optimization method for the optimal output working condition of a coal mill, which comprises collecting multi-source operation data of the coal mill, cleaning the collected data, and constructing historical training samples; dynamic working condition feature mining and adaptive safety constraint modeling of the coal mill; improved particle swarm optimization control parameter optimization by fusing working condition priori and dynamic safety constraints; and issuing the optimized set values of the coal feed quantity, primary air quantity and separator speed obtained by optimization to the on-site control station of the coal mill to drive the action of the actuator, thereby realizing closed-loop dynamic regulation of the output of the coal mill. The present application has the beneficial effect that, under the premise of ensuring that the outlet temperature of the coal mill and the current of the coal mill do not exceed the limit, the combination of the coal feed quantity, primary air quantity and separator speed is closer to the historical high-efficiency operation area.

Description

Technical Field

[0001] This invention relates to the field of coal mill control, specifically to a control optimization method for the optimal output conditions of a coal mill. Background Technology

[0002] During coal mill operation, its output is influenced by control parameters such as coal feed rate, primary air volume, and separator speed. These parameters need to be adjusted collaboratively to maintain the mill outlet temperature and current within safe ranges, thus affecting unit coal grinding energy consumption. During variable load periods, uncoordinated control parameter adjustments can easily lead to changes in the safety boundary under dynamic operating conditions, and the coal mill exhibits various energy consumption characteristic patterns at different coal feed rates. Therefore, extracting the trend coupling degree between control variables based on operating data, constructing a dynamic safety boundary that changes with operating conditions, and identifying operating modes with different efficiencies are key technological foundations for optimizing coal mill control parameters while balancing safety and energy conservation. However, existing technologies still have the following problems: 1. Conventional optimization algorithms typically initialize particles randomly and rely solely on the update speed of individual and global historical best positions. This can easily lead to local optima on the multi-mode energy consumption surface of a coal mill, and the search results may lack historical operating conditions, resulting in poor engineering feasibility.

[0003] 2. Existing technologies typically use fixed or steady-state safety limits, failing to consider safety risks caused by response lag during load changes. Under dynamic conditions such as load increases, the possibility of over-temperature or over-current is easily overlooked.

[0004] 3. Existing methods, when constructing surrogate models, typically only measure sample similarity based on the spatial distance of control parameters, without considering the impact of different dynamic trend coordination and operating condition mode affiliation on model prediction, resulting in insufficient prediction accuracy under complex operating conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a control optimization method for optimal coal mill output conditions. This method forms a complete closed loop through trend coupling degree extraction, multi-mode prior modeling, dynamic safety boundary contraction, local composite weighted proxy prediction, mode-guided particle swarm update, and dynamic safety constraint repair. Under the premise of ensuring that the coal mill outlet temperature and coal mill current do not exceed the limits, it outputs a combination of coal feed rate, primary air volume, and separator speed that is closer to the historical high-efficiency operating range.

[0006] A control optimization method for optimal output conditions of a coal mill includes the following steps: Multi-source operating data of the coal mill are collected and cleaned to construct a historical training sample library. The multi-source operating data includes coal feed rate sequence, primary air volume sequence, separator speed sequence, coal mill outlet temperature sequence, coal mill current sequence, and coal mill power sequence. Based on historical training sample library and current online operation data, the trend coupling degree between coal feed rate, primary air volume and separator speed is extracted, a dynamic safety boundary model is constructed that is adjusted with coal feed rate and coal feed rate change, and the operating mode is identified to obtain the operating condition prior information for subsequent particle swarm optimization. Based on the probabilistic initialization of operating conditions, the local composite weighted proxy model, the fitness function of safety constraints, the mode-oriented speed update and the dynamic safety constraint repair, the safe and energy-saving optimization of control parameters is achieved. The optimized setpoints for coal feed rate, primary air volume, and separator speed obtained through optimization are sent to the coal mill field control station to drive the actuators and achieve closed-loop dynamic adjustment of the coal mill output.

[0007] Furthermore, the trend coupling degree between coal feed rate, primary air volume, and separator rotation speed is extracted, specifically: By extracting a fixed-length sliding window from the coal feed rate sequence, primary air volume sequence, and separator speed sequence, we obtain the coal feed rate window sequence, primary air volume window sequence, and separator speed window sequence. Calculate the coal feed rate change sequence, primary air volume change sequence, and separator speed change sequence respectively within the sliding window; The coal feed rate change rate sequence, the primary air volume change rate sequence, and the separator speed change rate sequence are combined into a rate change sequence group, and the cross-correlation matrix is ​​calculated based on the rate change sequence group; The cross-correlation matrix is ​​decomposed into eigenvalues ​​to obtain three eigenvalues, which are then sorted in descending order and L1 normalized. The eigenvalue spectral entropy is calculated based on the normalized eigenvalues, and then converted into trend coupling degree.

[0008] Furthermore, a dynamic safety boundary model is constructed that adjusts with the coal feed rate and its variation, specifically as follows: Steady-state operation samples are selected from the historical training sample library. Steady-state operation samples refer to historical training samples when the coal mill is in a relatively stable output state. The upper limits of the steady-state coal mill outlet temperature and the upper limit of the steady-state coal mill current were obtained by kernel smoothing quantile boundary fitting. Finally, based on the change in coal feed rate and boundary sensitivity, transient margin contraction is performed to obtain the upper limit of dynamic coal mill outlet temperature and the upper limit of dynamic coal mill current.

[0009] Furthermore, the specific identification of operating conditions is as follows: For each historical training sample in the historical training sample library, construct a working condition identification feature vector. The working condition identification feature vector includes coal feed rate, unit coal grinding energy consumption and trend coupling degree. A Gaussian mixture model is used to cluster the normalized working condition identification feature vectors to obtain multiple working condition patterns. The mean value of control parameters corresponding to low-energy consumption samples is extracted from each operating mode and used as a priori guiding center. The mean of the control parameter vectors of all historical training samples within the operating mode is calculated, and then the joint fluctuation of each control parameter component relative to the mean is calculated to obtain the covariance matrix. The diagonal elements of the covariance matrix are used to characterize the fluctuation range of coal feed, primary air volume and separator speed, respectively; the off-diagonal elements are used to characterize the linkage relationship between different control parameters.

[0010] Furthermore, the specific initialization of the operating condition mode probability is as follows: Set the particle swarm size and upper and lower bounds of the control parameters. The upper and lower bounds of the control parameters can be determined by the physical limitations of the equipment, the allowable setting range of the control system, and the historical safe operating extreme values. The first element is drawn from the prior probability distribution of the working condition mode. Each particle operating condition mode label, the prior probability of the operating condition mode belonging to the historical training sample is... The overall probability of each working condition mode; With the first The prior guidance center of the working condition mode to which each particle belongs is used as the sampling mean, and the covariance matrix of the working condition mode is used as the sampling scattering scale to generate the initial position within the upper and lower bounds of the control parameters. Initialize particle velocity. The initial velocity is randomly generated within a preset proportion of the allowable range of each control parameter.

[0011] Furthermore, the method for constructing a locally composite weighted proxy model is as follows: Establish historical regression inputs and target values ​​for agent modeling; The control parameter spatial distance weight between the query point and the historical training samples is calculated based on the normalized control parameter spatial distance. Calculate the similarity weight of trend coupling degree between the query point and the historical training samples based on the difference in trend coupling degree. The posterior probability of the query point belonging to each working condition mode in the control parameter space is calculated based on the position of the query point in the control parameter space, and the posterior probability is used as the mode similarity weight. The composite weight is obtained by multiplying the control parameter spatial distance weight, the trend coupling degree similarity weight, and the pattern similarity weight. Three locally weighted linear regression models were constructed with unit coal grinding energy consumption, coal mill outlet temperature and coal mill current as target values, respectively.

[0012] Furthermore, the method for constructing the fitness function for safety constraints is as follows: Based on particle position Construct query points and call the surrogate model to calculate the surrogate model prediction value for each particle position; The change in coal feed rate is calculated based on the candidate coal feed rate and the current actual coal feed rate, and then input into the dynamic safety boundary model to obtain the upper limit of the dynamic coal mill outlet temperature and the upper limit of the dynamic coal mill current. Calculate the penalties for exceeding the limits of the coal mill outlet temperature and the coal mill current based on the dynamic safety boundary and the predicted values ​​of the surrogate model, respectively. The deviation penalty for each operating mode is calculated based on the particle position and the Mahalanobis distance between them. The predicted unit coal grinding energy consumption, safety constraint penalty, and mode deviation penalty are combined into a fitness function. The fitness function is used to evaluate the overall quality of candidate control parameter combinations. The smaller the value, the better the control parameter combination.

[0013] Furthermore, the mode-oriented speed update and dynamic safety constraint repair are specifically as follows: Calculate the most likely operating mode for each particle based on its current position; The particle velocity is updated jointly based on the inertia term, individual cognitive term, social learning term, and pattern guidance term; Calculate the new particle position based on the updated velocity and perform upper and lower bound repair of the control parameters; The particle locations where dynamic safety constraint repair needs to be performed are determined based on the surrogate model predictions and dynamic safety boundaries. For particle positions that violate dynamic safety constraints, dynamic safety constraints are repaired along the descent direction of the constraint function; Calculate the fitness of the particle position after repair, and update the individual historical best position and the global historical best position; The final global historical optimal position is used as the recommended control parameters for the coal mill. The final global historical optimal position includes the recommended coal feed rate, recommended primary air volume, and recommended separator speed.

[0014] Furthermore, the multiple operating modes include a high-efficiency mode, a transitional mode, and an inefficient mode. The operating mode with the lowest average unit coal grinding energy consumption is defined as the high-efficiency mode, the operating mode with the middle average unit coal grinding energy consumption is defined as the transitional mode, and the operating mode with the highest average unit coal grinding energy consumption is defined as the inefficient mode.

[0015] To achieve the above objectives, the present invention employs the following technical solution: Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention calculates the eigenvalue spectral entropy of the sequence of coal feed rate, primary air volume and separator rotation speed change rate to obtain the trend coupling degree, and introduces it as a dynamic coordination evaluation index into the optimization process to guide the optimization algorithm to prioritize the combination of control parameters with high coordination.

[0016] 2. This invention fits a safety boundary based on steady-state operation samples, and performs transient margin contraction on the safety boundary according to the change in coal feed rate and boundary sensitivity, to obtain the upper limit of dynamic coal mill outlet temperature and the upper limit of dynamic coal mill current that are adaptively tightened with the output adjustment range.

[0017] 3. Based on the operating condition patterns identified by the Gaussian mixture model, this invention introduces the prior knowledge of the operating condition patterns into the particle swarm optimization algorithm through pattern probability initialization, pattern-guided speed update, and pattern deviation penalty. This guides the search to approach the historical high-efficiency operating condition region. Furthermore, it constructs a local composite weighted surrogate model that integrates control parameter distance, trend coupling degree similarity, and pattern similarity to improve prediction accuracy. Attached Figure Description

[0018] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram illustrating the trend coupling degree of coal feed rate, primary air volume, and separator rotation speed in this invention. Figure 3 This is a schematic diagram of dynamic safety boundary modeling that incorporates transient margin correction; Figure 4 Schematic diagram of working condition pattern recognition and prior guidance center extraction based on energy consumption and output characteristics; Figure 5 This is a schematic diagram of the online low-energy safety control application process for coal mills. Detailed Implementation

[0019] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0020] like Figure 1 As shown, this invention proposes a control optimization method for optimal output conditions of coal mills, the main contents of which are as follows: S1. Multi-source operation data acquisition and historical sample database construction of coal mill The operating status of a coal mill is jointly affected by the coal feed rate, primary air volume, separator speed, mill outlet temperature, mill current, and mill power. The coal feed rate determines the instantaneous output of the coal mill, the primary air volume determines the coal powder conveying and drying capacity, and the separator speed determines the coal powder fineness and return powder intensity.

[0021] This step first collects multi-source operating data from the coal mill control system and historical operating database, and constructs historical training samples according to a unified timeline, so that each historical training sample includes control parameters, operating response, and energy consumption indicators. The specific steps are as follows: 1> Collect the coal feed rate sequence, primary air volume sequence, separator speed sequence, coal mill outlet temperature sequence, coal mill current sequence, and coal mill power sequence from the coal mill distributed control system.

[0022] Specifically, the coal feed rate sequence is in the first... The value at the sampling time is denoted as , Indicates the first The amount of coal entering the coal mill per unit time at the sampling time, in units of (tons / hour); primary air volume sequence in the first The value at the sampling time is denoted as , Indicates the first The primary air flow rate entering the coal mill at the sampling time, in units of (tons / hour); separator speed sequence in the first... The value at the sampling time is denoted as , Indicates the first The rotational speed of the separator at the sampling time, in units of (Revolutions per minute); The coal mill outlet temperature sequence is in the... The value at the sampling time is denoted as , Indicates the first The temperature of the pulverized coal gas flow at the pulverizer outlet at the sampling time is expressed in units of... (degrees Celsius); the coal mill current sequence is in the... The value at the sampling time is denoted as , Indicates the first The operating current of the coal mill motor at the sampling time is expressed in units of... (Ampere); The power sequence of the coal mill is in the first... The value at the sampling time is denoted as , Indicates the first The active power consumed by the coal mill motor at the sampling time, in units of (kilowatt).

[0023] In the specific implementation, the sampling period is denoted as... , This represents the time interval between two adjacent sampling times, in units of... (seconds), which can be 1 second, 5 seconds, or 10 seconds. If the original sampling times of different measuring points are not completely consistent, the sampling time of the coal feed rate sequence is used as a unified time axis, and the primary air volume sequence, separator speed sequence, pulverizer outlet temperature sequence, pulverizer current sequence, and pulverizer power sequence are mapped to the unified time axis through linear interpolation. If a measuring point has a short-term missing value, and the missing length does not exceed the preset missing length threshold (for example, 3 sampling points can be used), then adjacent valid sampling values ​​are used to linearly fill in the missing value. If the missing length exceeds the preset missing length threshold, then that time period is removed and does not participate in the construction of historical training samples. Based on this, it can be ensured that the control parameters and operating responses within the same historical training sample have a clear correspondence.

[0024] In one embodiment, for example, if the sampling period is 5 seconds, and the coal feed sampling time starts from second 0, then proceeds sequentially to 0 seconds, 5 seconds, 10 seconds..., and the primary air volume sampling time starts from second 2, then proceeds sequentially to 2 seconds, 7 seconds, 12 seconds...; if the coal feed sequence times are used as a unified time axis, in The primary air volume value at a given second is determined by the adjacent primary air volume value. and Linear interpolation is obtained, i.e. If the sampling period is 5 seconds, the coal feed sequence is in Seconds missing, adjacent valid values , If a measurement point is missing, linear interpolation is used to fill in the missing point. .

[0025] 2> Perform outlier removal on the time-aligned runtime data to obtain cleaned runtime data. Outliers include sensor disconnection points, points that clearly cross physical boundaries, and invalid points during control system switching.

[0026] In practical implementation, hard thresholds can be set according to the allowable range of the equipment. For example, the coal feed rate must not be less than 0, the primary air volume must not be less than 0, and the separator speed must not be less than 0. For the coal mill outlet temperature, coal mill current, and coal mill power, reasonable boundaries can be set in combination with the equipment protection settings and historical operating ranges. If any key variable at a sampling point exceeds the corresponding reasonable boundary, the sampling point is marked as an anomaly and removed.

[0027] Furthermore, to avoid distortion of subsequent trend calculations caused by a single spike, a sliding median filter can be performed on each variable. Taking the coal mill outlet temperature sequence as an example, several sampling points (e.g., two sampling points) are taken before and after the current sampling point to form a local window, and the local median and the local absolute deviation median are calculated. If the difference between the current sampling point and the local median exceeds a preset multiple of the local absolute deviation (e.g., five times), the current sampling point is identified as a spike point and replaced with the local median, thereby reducing the impact of sampling noise on trend coupling, safety boundaries, and surrogate models.

[0028] In one embodiment, for example, regarding the coal mill outlet temperature, if the five sampled values ​​within a local window are... The local median is Calculate the absolute deviation of each point from the median. The median absolute deviation is If the preset multiplier is 5, then the threshold is... Sample value The deviation from the median is Greater than Therefore, it was identified as a peak and replaced with a local median. .

[0029] 3> Construct historical training samples based on the cleaned running data.

[0030] Specifically, the first The control parameter vector of each historical training sample is denoted as . ,in, Indicates the index of historical training samples. ; This represents the total number of historical training samples; Indicates the first Coal feed rate corresponding to each historical training sample; Indicates the first One air volume corresponding to each historical training sample; Indicates the first Separator rotation speeds corresponding to each historical training sample; This represents the transpose operation of a vector. The runtime response of each historical training sample includes the coal mill outlet temperature. Coal mill current Coal mill power ,in, Indicates the first The coal mill outlet temperature corresponding to each historical training sample. Indicates the first The coal mill current corresponding to each historical training sample Indicates the first The power of the coal mill corresponding to each historical training sample.

[0031] In practical implementation, if the historical training samples are constructed in the form of single sampling points, then the coal feed rate, primary air volume, separator speed, mill outlet temperature, mill current, and mill power at the same sampling moment are directly combined into a single historical training sample. If the historical training samples are constructed in the form of time segments, then a short window is extracted with the current sampling moment as the right endpoint, and the mean value of each variable within this short window is used as the representative value of the historical training sample. The length of the short window can be 6 to 12 sampling points to reduce the influence of single-point noise while maintaining the time correspondence between control parameters and operating response.

[0032] In one embodiment, as an example, if historical training samples are constructed using a time-segment method, the short window length is 6 sampling points, the current sampling time is the 100th second, the sampling period is 5 seconds, and the coal feed rate within the window is 32.1, 32.3, 32.0, 32.2, 32.4, 32.0 (unit: If the coal feed rate is 0, then the average value is taken. .

[0033] 4> Calculate the unit coal grinding energy consumption for each historical training sample based on the cleaned operational data. The unit coal grinding energy consumption is denoted as... , It refers to the energy consumption of the coal mill corresponding to a unit coal feed rate, and is used to characterize the economic efficiency of coal mill operation.

[0034] In one implementation, for the first... A historical training sample, in terms of coal feed rate When the amount exceeds the minimum effective coal feed threshold, it can be calculated as follows: The calculation is performed in the following manner; where, Indicates the first The unit coal grinding energy consumption of a historical training sample. If the primary air fan's allocated power needs to be included in the coal grinding energy consumption, then it can be... Replace with the sum of the coal mill power and the primary air fan's allocated power. The minimum effective coal feed threshold is used to exclude periods of mill shutdown, start-up, and extremely low output. It can be set according to the equipment operation procedures and can be taken as 5% of the rated coal feed.

[0035] 5> The first The control parameter vectors, operational responses, and unit coal grinding energy consumption of each historical training sample are combined to form a historical training sample record, which constitutes a historical training sample library. , It is a dataset consisting of multiple historical training sample records, used for subsequent trend coupling degree calculation, dynamic safety boundary fitting, working condition pattern recognition, and local proxy model construction.

[0036] S2. Dynamic Operating Condition Feature Mining and Adaptive Safety Constraint Modeling of Coal Mills During the output adjustment of the coal mill, the coal feed rate, primary air volume, and separator speed need to be adjusted in a coordinated manner. If the coal feed rate increases but the primary air volume or separator speed is not adjusted synchronously, it can easily lead to insufficient conveying capacity inside the mill, insufficient drying of pulverized coal, or imbalance in the separation process, which in turn can cause mill blockage, excessive outlet temperature of the coal mill, or excessive current of the coal mill.

[0037] This step, based on historical training sample database and current online operating data, extracts the trend coupling degree between coal feed rate, primary air volume, and separator speed, constructs a dynamic safety boundary that adjusts with coal feed rate and its variation, and identifies efficient, transitional, and inefficient modes to obtain prior operating information for subsequent particle swarm optimization. The specific steps are as follows: S201. Extraction of the trend coupling degree between coal feed rate, primary air volume and separator speed. Coal feed rate, primary air volume, and separator speed reflect the fuel input, conveying and drying conditions, and pulverized coal separation intensity of the coal mill, respectively. The more coordinated the trends of these three control variables, the easier it is for the coal mill to maintain stable operation; conversely, the more dispersed the trends, the more likely the coal mill is to enter a state of mismatch between conveying, drying, and separation. This step calculates the correlation between the rate-of-change sequences of the three control variables within a sliding window and obtains the trend coupling degree based on the concentration of the eigenvalue spectrum of the correlation matrix. This trend coupling degree characterizes the strength of the coordinated regulation of the control variables under the current operating conditions, providing a dynamic coordination evaluation index for subsequent optimization and guiding the optimization algorithm to prioritize combinations of control parameters with high coordination, such as... Figure 2 As shown, the specific steps are as follows: 1> Using the current sampling time as the right endpoint, extract a fixed-length sliding window from the coal feed rate sequence, primary air volume sequence, and separator speed sequence to obtain the coal feed rate window sequence, primary air volume window sequence, and separator speed window sequence.

[0038] Specifically, the width of the sliding window is denoted as... , This indicates the number of sampling points used for each trend analysis, which can be set to 60. (Sampling period...) seconds, then N w =60 indicates that data from the most recent 60 seconds is used to depict the current dynamic trend; if the sampling period is... seconds, then N w =300 indicates that data from the most recent 300 seconds is used to characterize the current dynamic trend. The coal feed window sequence is denoted as... , It refers to the continuous movement within the current sliding window. A sequence consisting of coal feed rate sampling values; the primary air volume window sequence is denoted as... , It refers to the continuous movement within the current sliding window. A sequence consisting of primary air volume sampling values; the separator speed window sequence is denoted as... , It refers to the continuous movement within the current sliding window. A sequence consisting of sampled values ​​of the separator rotation speed.

[0039] 2> Calculate the coal feed rate change sequence, primary air volume change sequence, and separator speed change sequence respectively within the sliding window.

[0040] Specifically, the series of coal feed rate changes is denoted as... , This refers to the sequence formed by differentiating adjacent coal feed rate samples within the current sliding window, used to characterize the increasing or decreasing trend of coal feed rate within the window; the primary air volume change rate sequence is denoted as... , This refers to the sequence formed by differentiating adjacent primary air volume samples within the current sliding window, used to characterize the increasing or decreasing trend of primary air volume within the window; the separator speed change rate sequence is denoted as... , It refers to the sequence formed by the difference of adjacent separator speed sampling values ​​within the current sliding window, which is used to characterize the increasing or decreasing trend of separator speed within the window.

[0041] In one implementation, the rate of change within the sampling interval is obtained by subtracting the previous sampling point value from the current sampling point value; if it needs to be converted to a rate of change per unit time, the rate of change is divided by the sampling period. That's fine. For example, if the coal feed rate is 32.0 at two adjacent sampling points... and 32.4 If the sampling period is 5 seconds, then the change in coal feed rate within this sampling interval is 0.4. The corresponding rate of change per unit time is 0.08. By using a rate of change series, the impact of differences in the absolute values ​​of variables under different load levels can be reduced, and the trend coupling degree can focus more on whether the control variables rise synchronously, fall synchronously, or change in opposite directions.

[0042] 3> The coal feed rate change rate sequence, the primary air volume change rate sequence, and the separator speed change rate sequence are combined into a rate change sequence group, and the cross-correlation matrix is ​​calculated based on the rate change sequence group.

[0043] Specifically, the cross-correlation matrix is ​​denoted as , The size is The cross-correlation matrix is ​​used to characterize the degree of linear synchronization between each pair of the coal feed rate change sequence, the primary air flow rate change sequence, and the separator speed change sequence. The diagonal elements of the cross-correlation matrix are 1, indicating that each rate change sequence is perfectly correlated with itself; the off-diagonal elements are used to characterize the degree of correlation between the changing trends of different control variables.

[0044] In the specific implementation, firstly, respectively... , and Perform mean elimination (mean elimination means subtracting the arithmetic mean of each element in the rate of change sequence from the sequence itself, making the mean of the sequence zero) to obtain the mean-free rate of change sequence; then calculate the covariance of the two mean-free rate of change sequences and divide it by the product of the standard deviations of the two rate of change sequences to obtain the corresponding Pearson correlation coefficient. If a control variable remains almost unchanged within a certain sliding window, causing the standard deviation of the rate of change sequence of that control variable to be less than a preset fluctuation threshold (for example, a threshold value of 0.5%), then... If the correlation coefficient between the control variable and other control variables is set to 0, the sliding window will be marked as a low dynamic window.

[0045] In one embodiment, as an example, if the mean-removed coal feed rate change sequence is... The primary air volume change rate sequence is The calculated covariance of the two sequences is 0.0026, and the product of their standard deviations is 0.0072. Therefore, the corresponding Pearson correlation coefficient is... .

[0046] 4> Cross-correlation matrix Eigenvalue decomposition is performed to obtain three eigenvalues, which are then sorted in descending order and L1 normalized.

[0047] Specifically, the first The eigenvalues ​​are denoted as , Used to characterize the changing trends of the three control variables in the first... The relevant energy in each principal direction, among which... Indicates the eigenvalue index. If the trends of coal feed rate, primary air volume, and separator speed are highly consistent, most of the relevant energy will be concentrated in the first principal direction; if the trends of the three control variables are dispersed or conflict with each other, the relevant energy will be dispersed in multiple principal directions.

[0048] Furthermore, L1 normalization is performed on the three eigenvalues ​​to obtain normalized eigenvalues. , Indicates the first The proportion of each eigenvalue in the sum of all eigenvalues ​​is used to eliminate the influence of the absolute value of the eigenvalue on subsequent calculations. The value of is in the range of 0 to 1, and the sum of the three normalized eigenvalues ​​is 1, which can be expressed as: ;in, Indicates the index for normalized summation. ; Indicates the first Each eigenvalue is normalized. Subsequent trend coupling reflects only whether the relevant energy distribution is concentrated, and is unaffected by the overall fluctuation magnitude across different windows.

[0049] 5. Calculate the eigenvalue spectral entropy based on the normalized eigenvalues, and convert the eigenvalue spectral entropy into trend coupling degree. Trend coupling degree can introduce the current dynamic adjustment state into the optimization process, avoiding insufficient security caused by relying solely on static control parameters for optimization.

[0050] In one implementation, because the normalized eigenvalue spectrum is more concentrated and the eigenvalue spectrum entropy is lower when the control variables are highly coordinated, and more dispersed and with higher eigenvalue spectrum entropy when the control variables are uncoordinated, the trend coupling degree is obtained by subtracting the normalized eigenvalue spectrum entropy from 1. The calculation method is expressed as follows: ; in, This indicates the trend coupling degree within the current sliding window. It is used to measure the degree of coordination of the changing trends of the three control variables: coal feed rate, primary air volume, and separator speed. The value ranges from 0 to 1. The closer the value is to 1, the more concentrated the changing trends of the three control variables are in the same main direction, and the stronger the multi-variable coordination. The closer the value is to 0, the more dispersed the changing trends of the three control variables are, and the worse the working condition coordination. Represent the natural logarithm function; This represents the normalized coefficient of the spectral entropy of the largest eigenvalue corresponding to the three control variables. When a certain normalized eigenvalue... When it is 0, This item is treated as 0.

[0051] In one embodiment, for example, if the coal feed rate, primary air volume, and separator speed increase roughly synchronously within a certain sliding window, the normalized eigenvalues ​​of the three eigenvalues ​​of the cross-correlation matrix are approximately 0.90, 0.07, and 0.03, respectively. In this case, the eigenvalue spectrum is relatively concentrated, and the calculated trend coupling degree is 0.65, indicating that the current output adjustment is relatively coordinated. If, in another sliding window, the coal feed rate increases while the primary air volume lags behind and the separator speed fluctuates significantly, the normalized eigenvalues ​​are approximately 0.45, 0.35, and 0.20, respectively. In this case, the eigenvalue spectrum is dispersed, and the calculated trend coupling degree is 0.05, indicating that the current control variable coordination is poor.

[0052] S202, Dynamic safety boundary modeling with transient margin correction The upper limits of the coal mill outlet temperature and current are not only related to the equipment protection settings, but also to the coal feed rate, ventilation conditions, and load adjustment range. When the coal feed rate increases rapidly, the coal mill outlet temperature and current typically exhibit a lag response. If fixed safety limits are still used or safety boundaries are obtained solely based on steady-state operating samples, the risk of over-temperature or over-current may be underestimated during load increases. This step first selects steady-state operating samples from the historical training sample library, then fits the steady-state upper limits of the coal mill outlet temperature and current corresponding to the coal feed rate, and finally performs transient margin contraction based on the change in coal feed rate and boundary sensitivity to obtain dynamic upper limits of the coal mill outlet temperature and current. This allows the safety constraints to dynamically tighten with the output adjustment range, avoiding over-temperature or over-current situations during load changes. Figure 3 As shown, the specific steps are as follows:

[0053] 1> From the historical training sample database The steady-state operation samples are selected from the historical training samples when the coal mill is in a relatively stable output state.

[0054] In one implementation, during screening, the variation ranges of coal feed rate, mill outlet temperature, and mill current are calculated on a continuous time period basis. If the variation range of coal feed rate is less than a preset stable coal feed rate threshold, the variation range of mill outlet temperature is less than a preset stable temperature threshold, and the variation range of mill current is less than a preset stable current threshold within the continuous time period, then the historical training samples within that continuous time period are determined to be steady-state operation samples. The preset stable coal feed rate threshold can be taken as 1% to 3% of the rated coal feed rate, and the preset stable temperature threshold can be taken as... arrive The preset current stability threshold can be taken as 1% to 3% of the rated current. Based on this, the influence of load increase, load decrease and disturbance recovery process on the fitting of steady-state safety boundary can be reduced, so that the steady-state safety boundary mainly reflects the long-term stable relationship between coal feed and safety indicators.

[0055] In one embodiment, for example, if the coal feed rate varies within a certain continuous time period, the variation range is... (threshold is) The temperature variation range at the coal mill outlet is as follows: (threshold is) The current variation range of the coal mill is (threshold is) If ), then all historical training samples within that time period are determined to be steady-state operating samples.

[0056] 2> Using the coal feed rate in the steady-state operation sample as the independent variable and the pulverizer outlet temperature as the target variable, the upper limit of the steady-state pulverizer outlet temperature is obtained by kernel smoothing quantile boundary fitting. , This refers to a coal feed rate of The upper boundary of the safe coal mill outlet temperature is obtained by fitting the steady-state operation sample.

[0057] In one implementation, firstly, the range of coal feed rates is divided into multiple coal feed rate grid points. For any given coal feed rate grid point, a steady-state operating sample near that grid point is selected, and a kernel weight is calculated based on the distance between the coal feed rate of the steady-state operating sample and the grid point. The closer the distance, the larger the kernel weight; the farther the distance, the smaller the kernel weight. The kernel function can be the Epanechnikov kernel function, and the kernel window width can be determined through cross-validation or historical sample density. Then, the weighted quantile of the mill outlet temperature is calculated in the steady-state operating samples with kernel weights. The quantile can be set to 0.95, yielding the upper limit of the steady-state mill outlet temperature corresponding to that coal feed rate grid point. Finally, linear interpolation is performed on the upper limits of the steady-state mill outlet temperature between adjacent coal feed rate grid points to obtain the temperature that continuously varies with the coal feed rate. Based on this, it is possible to avoid abrupt changes in the safety boundary with the amount of coal fed, while retaining the constraint effect of historical high-temperature operating samples on the safety boundary.

[0058] In one embodiment, as an example, for the coal feed grid point Select nearby arrive The steady-state operation sample, based on the The distances were weighted using the Epanechnikov kernel function, and the 0.95 quantile of the mill outlet temperature was calculated among these weighted samples. The upper limit of the steady-state coal mill outlet temperature .

[0059] 3> Using the coal feed rate in the steady-state operation sample as the independent variable and the mill current as the target variable, the upper limit of the steady-state mill current is obtained by using the same kernel smoothing quantile boundary fitting method as the upper limit of the steady-state mill outlet temperature. , This refers to a coal feed rate of The upper boundary of the coal mill current safety is obtained by fitting the steady-state operation sample.

[0060] In one implementation, for each coal feed rate grid point, the kernel weight of the steady-state operating sample is first calculated based on the coal feed rate distance. Then, the weighted 0.95 quantile of the mill current is calculated to obtain the upper limit of the steady-state mill current corresponding to that coal feed rate grid point. Finally, the continuously changing [current value] is obtained through interpolation. The upper limit of steady-state coal mill current is used to describe the allowable range of coal mill motor current under different output levels, and can provide a safety constraint for subsequent optimization when under high load or when there is a mismatch between air and coal.

[0061] In one embodiment, for example, if the coal feed rate The upper limit of the steady-state current of the coal mill at the weighted 0.95 quantile is Coal feed The upper limit of the steady-state current of the coal mill at the weighted 0.95 quantile is Coal feed rate, after linear interpolation The upper limit of the steady-state current of the coal mill is .

[0062] 4> Based on the sensitivity of the upper limit of the steady-state coal mill outlet temperature to the change of coal feed rate, the upper limit of the steady-state coal mill outlet temperature is reduced by transient margin to obtain the upper limit of the dynamic coal mill outlet temperature.

[0063] Specifically, the current actual coal supply is recorded as follows: , This represents the actual coal feed rate of the coal mill during the current control cycle; the candidate coal feed rate is denoted as... , This represents the coal feed rate in the combination of control parameters to be evaluated during subsequent optimization; the change in coal feed rate is denoted as... , It is used to characterize the adjustment range of the candidate coal feed rate relative to the current actual coal feed rate within the current control cycle.

[0064] In practical implementation, we first base it on the upper limit curve of the steady-state coal mill outlet temperature. Calculate the difference between adjacent feed rate grid points. The ratio of the difference value to the corresponding feed rate interval is the slope within that grid point interval. After taking the absolute value of the slope, perform linear interpolation between adjacent grid points to obtain the absolute value of the sensitivity of the upper limit of the steady-state coal mill outlet temperature to the feed rate. , This value is used to characterize the rate at which the upper limit of the steady-state mill outlet temperature changes with the coal feed rate near that feed rate. A larger value indicates that the mill outlet temperature is more sensitive to feed rate adjustments within that feed rate range. Then, it is combined with the change in coal feed rate... and outlet temperature transient margin coefficient By narrowing the safety boundary, the upper limit of the dynamic coal mill outlet temperature is obtained, and the calculation method is expressed as follows: ; in, Indicates the coal feed rate as And the change in coal feed rate is The upper limit of the dynamic coal mill outlet temperature at that time; This represents the transient margin coefficient for outlet temperature, used to control the contraction of the upper limit of the outlet temperature of a dynamic coal mill, and can be taken as 0.5 to 1.2. This represents the absolute value of the change in coal feed rate. If a certain range of coal feed rates... If the change in coal feed rate is large, the same change in coal feed rate will result in a larger contraction of the safety margin; if within a certain range of coal feed rates... If the size is smaller, the safety boundary will shrink less.

[0065] In one embodiment, for example, if the candidate coal feed rate Current actual coal supply The change in coal feed rate , , outlet temperature transient margin coefficient The upper limit of the outlet temperature of the dynamic coal mill is Therefore, when the load increase is large, the upper limit of the dynamic coal mill outlet temperature will shrink earlier than the steady-state boundary, thereby improving the safety margin of the optimization process.

[0066] 5> Based on the sensitivity of the steady-state upper limit of the coal mill current to the change in coal feed rate, the transient margin of the steady-state upper limit of the coal mill current is reduced to obtain the dynamic upper limit of the coal mill current.

[0067] In the specific implementation, we first base it on the upper limit curve of the steady-state coal mill current. Calculate the difference between adjacent feed rate grid points. The ratio of the difference value to the corresponding feed rate interval is the slope within that grid point interval. After taking the absolute value of the slope, perform linear interpolation between adjacent grid points to obtain the absolute value of the sensitivity of the steady-state mill current upper limit to the feed rate. , This value is used to characterize how quickly the upper limit of the steady-state mill current changes with the coal feed rate near that feed rate. A larger value indicates that the mill current is more sensitive to feed rate adjustments within that feed rate range. Then, it is combined with the change in coal feed rate... and current transient margin factor By narrowing the safety boundary, the upper limit of the dynamic coal mill current is obtained, and the calculation method is expressed as follows: ; in, Indicates the coal feed rate as And the change in coal feed rate is The upper limit of dynamic coal mill current at that time; This represents the transient margin coefficient for current, used to control the reduction range of the upper limit of the current in a dynamic coal mill, and can be taken as 0.5 to 1.2. If within a certain coal feed rate range... If the change in coal feed rate is large, the same change in coal feed rate will result in a larger contraction of the safety margin; if within a certain range of coal feed rates... If the size is smaller, the safety boundary will shrink less.

[0068] It should be noted that by simultaneously constructing dynamic upper limits for the coal mill outlet temperature and dynamic upper limits for the coal mill current, the safety boundary of the coal mill can be adaptively constrained during load increases, load decreases, and rapid load adjustments. This effectively reduces the risk of over-temperature and over-current caused by delayed response, ensuring equipment operation safety while pursuing low energy consumption.

[0069] S203, Working Condition Pattern Recognition and Prior Guidance Center Extraction Coal mills exhibit different unit grinding energy consumption characteristics under different coal feed rates. Even with the same coal feed rate, different efficiency modes may arise due to variations in the primary air volume and separator speed. To prioritize searching historically efficient operating conditions in subsequent optimizations and avoid recommending control parameter combinations lacking historical data support, this step identifies operating mode based on coal feed rate, unit grinding energy consumption, and trend coupling degree. The mean control parameters corresponding to low-energy-consumption samples are extracted from each operating mode as prior guidance centers. This avoids recommending control parameter combinations lacking historical data support, providing an efficient and feasible search direction for the particle swarm optimization algorithm, accelerating convergence, and improving the engineering credibility of the recommended results. Figure 4 As shown, the specific steps are as follows: 1> Historical training sample database Each historical training sample in the dataset is used to construct a feature vector for identifying operating conditions.

[0070] Specifically, the first The feature vector for identifying the working conditions of each historical training sample is denoted as . , can be represented as ,in, Indicates the first The trend coupling degree corresponding to the sliding window containing each historical training sample; This represents the transpose operation of a vector.

[0071] In one implementation, since coal feed rate, unit pulverizing energy consumption, and trend coupling degree have different dimensions and value ranges, normalization processing needs to be performed on the three feature channels before inputting them into the working condition pattern recognition model. Coal feed rate and unit pulverizing energy consumption can be normalized using the minimum and maximum values ​​in the historical training sample library; the trend coupling degree is in the range of 0 to 1 and can be used directly. The working condition recognition feature vector simultaneously uses coal feed rate, unit pulverizing energy consumption, and trend coupling degree, which can incorporate output level, economy, and dynamic coordination into the same clustering analysis framework.

[0072] 2> A Gaussian mixture model is used to cluster the normalized working condition identification feature vectors to obtain multiple working condition patterns.

[0073] In practical implementation, the number of working condition modes is denoted as: , This refers to the total number of operating modes. You can start by... The number of operating conditions is set to 2 to 5 and Gaussian Mixture Models (GMMs) are trained respectively. Then, a reasonable number of operating conditions can be selected using the Bayesian information criterion. To simplify the calculation, this invention takes... The operating modes are named according to the average unit coal grinding energy consumption within each operating mode. The operating mode with the lowest average unit coal grinding energy consumption is defined as the high-efficiency mode, the operating mode with the middle average unit coal grinding energy consumption is defined as the transition mode, and the operating mode with the highest average unit coal grinding energy consumption is defined as the inefficient mode. The prior probability of each working condition pattern is denoted as: , This refers to the historical training samples belonging to the first... The overall probability of each working condition mode satisfies ,in, Indicates the operating condition mode index. .

[0074] It should be noted that a Gaussian Mixture Model (GMM) is a probabilistic clustering model that uses a weighted sum of multiple Gaussian distributions to fit the overall distribution of the data and estimates the parameters using the expectation-maximization algorithm. The Gaussian Mixture Model uses... A Gaussian distribution is used to fit the distribution of historical training samples in the working condition recognition feature space, with each Gaussian distribution corresponding to a working condition mode. During training, the expectation-maximization algorithm can be used. First, the probability of each historical training sample belonging to each working condition mode is calculated based on the current Gaussian distribution parameters. Then, the mean, covariance, and prior probability of the Gaussian distribution are updated based on the sample belonging probability, so that the overall likelihood gradually increases.

[0075] 3> Based on the clustering results, each historical training sample is assigned to the corresponding working condition mode, and within each working condition mode, it is sorted from low to high according to the unit coal grinding energy consumption. The top 20% of historical training samples with the lowest unit coal grinding energy consumption are selected to form a low energy consumption sample set.

[0076] Specifically, the first The low-energy consumption sample set for each operating mode is denoted as . , It refers to the first The set of historical training samples with the lowest unit coal grinding energy consumption in each working mode; This indicates the number of samples in the low-energy-consumption sample set.

[0077] Furthermore, from the low-energy-consumption sample set Extracting control parameter vectors and for low-energy sample sets The mean of all control parameter vectors is calculated to obtain the first... The prior guidance center for each working condition mode is calculated as follows: ; in, Indicates the first The prior guidance center for each working mode is a size of The control parameter vector includes reference values ​​for coal feed rate, primary air volume, and separator speed; Indicates that the object to be summed is the first... The historical training samples in the low-energy sample set of each operating mode. The prior guidance center refers to the average combination of control parameters corresponding to the low-energy historical training samples of this operating mode, which is used to guide particles to move towards the historically efficient region during particle swarm initialization and velocity update.

[0078] 4> Calculate the covariance matrix of each operating mode in the control parameter space.

[0079] Specifically, the first The covariance matrix of each working condition mode is denoted as . , The size is The matrix, by the first The control parameter vectors of historical training samples within each operating mode are calculated to characterize the distribution range and interrelationships of coal feed rate, primary air volume, and separator rotation speed within that operating mode. The covariance matrix is ​​used both for initial particle sampling and to calculate the degree to which candidate control parameters deviate from the known operating mode.

[0080] In specific implementation, for the first The mean of the control parameter vectors of all historical training samples within a given operating condition is calculated, and then the joint fluctuation of each control parameter component relative to the mean is calculated to obtain the covariance matrix. The diagonal elements of the covariance matrix characterize the fluctuation ranges of coal feed rate, primary air volume, and separator speed, while the off-diagonal elements characterize the linkage between different control parameters. For example, if in a certain operating mode, the primary air volume usually increases synchronously when the coal feed rate increases, then the covariance between the coal feed rate and the primary air volume is positive.

[0081] It should be noted that the operating condition mode recognition uses... As clustering inputs, prior guiding centers and covariance matrices are used... This is expressed as a control parameter. Based on this, the operating mode can be determined by output, energy consumption, and dynamic coordination, while the optimization process is directly executed within the controllable range of coal feed, primary air volume, and separator speed, avoiding the use of unit coal grinding energy consumption, which cannot be directly set, as an optimization variable.

[0082] S3. Improved Particle Swarm Optimization Control Parameter Optimization Integrating Prior Operating Conditions and Dynamic Safety Constraints After obtaining the trend coupling degree, dynamic safety boundary, and prior operating conditions, this step performs optimization within the control parameter space comprised of coal feed rate, primary air volume, and separator rotation speed. Conventional particle swarm optimization typically initializes particles randomly and relies solely on the update rate of individual historical best positions and the global historical best position. This can easily lead to problems such as initial particles falling into inefficient regions, the search process violating safety constraints, and the final recommended parameters lacking historical operating condition support.

[0083] This step achieves safe and energy-saving optimization of control parameters through probabilistic initialization of operating conditions, local composite weighted proxy model, fitness function of safety constraints, mode-guided speed update, and dynamic safety constraint repair. It minimizes unit pulverizing energy consumption while ensuring that the pulverizer outlet temperature and pulverizer current do not exceed limits, and ensures that the recommended control parameters have sufficient historical operating condition support. The specific steps are as follows: S301. Particle swarm initialization based on the probability distribution of operating conditions. To improve the initial search quality of the particle swarm, this step does not generate initial particles completely randomly within the control parameter space. Instead, it generates initial particles based on the prior probability of the operating mode, the prior guiding center, and the covariance matrix. This ensures that the particles are concentrated in historically efficient or common operating mode regions during the initialization phase, while retaining the ability to explore other feasible regions. The specific steps are as follows: 1> Set the particle swarm size and upper and lower bounds of the control parameters.

[0084] Specifically, the particle swarm size is denoted as , This represents the number of particles participating in the optimization process, which can be between 30 and 60. The lower bound of the control parameter is denoted as... , ,in, This indicates the lower limit of the coal feed rate. This indicates the lower limit of the primary air volume. Indicates the lower limit of the separator speed. The transpose operation of a vector is represented; the upper bound of the control parameter is denoted as . , ,in, This indicates the upper limit of the coal feed rate. This indicates the upper limit of the primary air volume. This indicates the upper limit of the separator speed. This represents the transpose operation of a vector.

[0085] In practical implementation, the upper and lower limits of each control parameter can be determined jointly by the physical limitations of the equipment, the allowable setting range of the control system, and historical safe operating extreme values. For example, the upper and lower limits of the coal feed rate are determined by the allowable output range of the coal feeder; the upper and lower limits of the primary air volume are determined by the ventilation capacity of the primary air fan and the coal mill; and the upper and lower limits of the separator speed are determined by the allowable frequency range of the frequency converter and the coal powder fineness requirements. By clearly defining the upper and lower limits of the control parameters, it can be ensured that the particle position is always within the executable range of the control system.

[0086] 2> Regarding the first Individual particle extraction working condition mode labels.

[0087] Specifically, the first The i-th particle represents the i-th particle in the particle swarm. There are 10 candidate search individuals, among which... Indicates particle index, . No. The working mode label for each particle is denoted as , Extracted from the prior probability distribution of the working condition mode, i.e., the th The probability of each working condition mode being selected is: Based on this, the operating conditions with a higher probability of occurrence or more stable operation in the past can obtain more initial particles, while other operating conditions still have a probability of being selected, thus avoiding all particles being concentrated near a single mode.

[0088] In one implementation, if it is necessary to further enhance the exploration of efficient modes, the sampling probability of efficient modes can be set to no less than a preset proportion (for example, 50%), and the remaining probabilities can be allocated according to the historical prior probabilities of transitional modes and inefficient modes. This can both increase the probability of particles approaching low-energy regions and not completely abandon the search for potential feasible improvement regions in transitional and inefficient modes.

[0089] 3> with the first Using the prior guidance center of the operating mode to which each particle belongs as the sampling mean, and the covariance matrix of that operating mode as the sampling scattering scale, the initial position is generated within the upper and lower bounds of the control parameters, so that the optimization starting point is close to the historical efficient operating area, rather than falling into the inefficient or dangerous area completely randomly.

[0090] In one implementation, the first The initial position of each particle is denoted as . , The size is The control parameter vector represents the first... The initial position of a particle at the 0th iteration, considering the combination of coal feed rate, primary air volume, and separator rotation speed, can be represented as:

[0091] ;

[0092] in, Indicates For the mean, The covariance matrix is ​​a truncated multivariate normal distribution, and in the case of... and The sampling points within the defined boundary are truncated, and only those points within the boundary are retained. This indicates a truncated normal distribution, used for random sampling within the upper and lower bounds of control parameters; Indicates the first The prior guidance center for the operating mode to which each particle belongs; Indicates the first The covariance matrix of the operating mode to which each particle belongs; This represents the dispersion scaling factor, used to control the degree of dispersion of initial particles around the prior guide center, and can be taken as 0.2 to 0.5; The smaller the size, the more concentrated the initial particles are in the low-energy-consumption historical region; The larger the value, the wider the initial particle coverage area.

[0093] In practical implementation, we can start with the mean value. Covariance is Sampling is performed in a multivariate normal distribution. If any control parameter in the sampling result exceeds the corresponding upper or lower bound, resampling is performed. If continuous resampling exceeds a preset number (e.g., 20 times), the out-of-bounds component can be truncated to the corresponding boundary value. The initial particles obtained in this way satisfy the equipment control range and are close to the historical high-efficiency operating condition region.

[0094] 4> Initialize particle velocity.

[0095] In specific implementation, the first The velocity of the particle at the 0th iteration is denoted as . , This is used to characterize the initial direction and magnitude of the particle's movement in the control parameter space. The initial velocity can be randomly generated within a preset proportion of the allowable range of each control parameter (e.g., 10% of the upper and lower bounds of each dimension), thereby preventing the particle from jumping out of the safe operating area due to excessive movement during the initial iteration phase.

[0096] S302, Construction of Local Composite Weighted Proxy Model

[0097] During the particle swarm optimization iteration process, it is necessary to frequently evaluate the unit pulverizing energy consumption, pulverizer outlet temperature, and pulverizer current corresponding to the candidate control parameters. This step constructs a local composite weighted surrogate model for each candidate control parameter, giving greater weight to historical training samples that are more similar to the candidate control parameter in terms of control parameter spatial location, trend coupling, and operating conditions. The specific steps are as follows:

[0098] 1> Establish historical regression inputs and target values ​​for surrogate modeling.

[0099] Specifically, the first The control parameter vector of each historical training sample is: , No. The unit coal grinding energy consumption, coal mill outlet temperature, and coal mill current corresponding to the historical training samples are as follows: , and For candidate control parameters, the query point is denoted as... , represented as ,in, This indicates the amount of coal supplied at the query point. This indicates the primary air volume at the query point. Indicates the speed of the separator at the query point. This represents the transpose operation of a vector.

[0100] Furthermore, the control parameter vector of the historical training samples is expanded into a historical regression input containing an intercept term, denoted as […]. , represented as In this context, the constant term 1 is used to fit the linear regression intercept; the query point control parameter vector is expanded into a query regression input containing the intercept term, and the expanded query regression input is denoted as... , represented as By adding an intercept term, the local linear regression model can fit the local bias near the query point.

[0101] 2> Calculate the control parameter spatial distance weight between the query point and the historical training samples based on the normalized control parameter spatial distance.

[0102] Specifically, since the coal feed rate, primary air volume, and separator rotation speed have different dimensions, before calculating the distance, the three control parameters are first normalized using the minimum and maximum values ​​from the historical training sample database. The spatial distance of the normalized control parameters is denoted as... , It refers to the first Historical training samples and query points Euclidean distance in the normalized control parameter space.

[0103] In one implementation, the spatial distance weight of the control parameter is denoted as... The Gaussian kernel function can be used for calculation, and the calculation method is expressed as follows:

[0104] ;

[0105] in, Indicates the first Each historical training sample relative to the query point Control parameters spatial distance weighting; This represents the kernel width of the control parameter space, used to control the decay rate of the weights of neighboring samples in the space. It can be taken as 1 to 2 times the average neighbor distance of historical training samples in the normalized control parameter space (the average neighbor distance is the arithmetic mean of the Euclidean distances between all historical training samples and their nearest neighbors). This represents the natural exponential function. The smaller, The larger the value, the closer the historical training sample is to the query point in terms of the combination of control parameters; The larger, The smaller the value, the more significant the difference between the historical training sample and the query point's operating area.

[0106] 3> Calculate the similarity weight of trend coupling between the query point and the historical training samples based on the difference in trend coupling.

[0107] Specifically, the first The trend coupling degree of each historical training sample is The trend coupling degree corresponding to the query point is calculated by S201 based on the sliding window containing the historical training sample; When searching for the best online, Take the current trend coupling degree calculated by S201 for the current control cycle.

[0108] In one implementation, the similarity weight of trend coupling is denoted as... The calculation method is expressed as follows:

[0109] ;

[0110] in, Indicates the first Each historical training sample relative to the query point The trend coupling degree and similarity weights; This represents the width of the trend coupling kernel, used to control the degree of influence of the trend coupling difference on the sample weights, and can be taken from 0.1 to 0.3; This represents the natural exponential function. and The smaller the difference, the more similar the historical training samples are to the current operating conditions in terms of the degree of coordination of control variables. The larger the value, the more significant the difference between the historical training samples and the current operating conditions. The smaller.

[0111] 4> Calculate the posterior probability of the query point belonging to each working condition mode in the control parameter space based on the position of the query point in the control parameter space, and use the posterior probability as the mode similarity weight.

[0112] Specifically, since the unit coal grinding energy consumption of the candidate control parameters is unknown before the evaluation, the Gaussian mixture model cannot be directly used to calculate the pattern probability of the query point. Therefore, in the optimization and evaluation stage, this step uses the prior guiding center and covariance matrix of each working condition mode in S203 in the control parameter space to calculate the posterior probability of the query point pattern.

[0113] In one implementation, the query point By substituting the Gaussian distribution of the control parameters corresponding to each operating mode, the likelihood value of the query point belonging to each operating mode is obtained. This is then combined with the prior probability of the operating mode. Normalization is performed to obtain the query point. Belongs to the Posterior probability of each working condition mode The calculation method is expressed as follows:

[0114] ;

[0115] in, Indicates As the mean, with The probability density function of the multivariate Gaussian distribution of the covariance matrix at the query point The value at; Indicates the index for summing operating conditions. ; Indicates the first Prior probabilities of each working condition mode; Indicates the first A priori guidance center for each working condition mode; Indicates the first Covariance matrix of each operating condition mode; Indicates As the mean, with The probability density function of the multivariate Gaussian distribution of the covariance matrix at the query point The value at that position. The operating mode to which each historical training sample belongs is denoted as... Then the first The pattern similarity weights of each historical training sample relative to the query point are denoted as follows: , The larger the value, the closer the query point is to the first... The more consistent the operating conditions of a historical training sample, the greater its influence on local fitting.

[0116] 5> Multiply the control parameter spatial distance weight, trend coupling degree similarity weight, and pattern similarity weight together to obtain the composite weight.

[0117] In one implementation, the first Historical training samples for query points The composite weight is denoted as , Used to characterize the The degree of influence of each historical training sample on the local regression fitting of the query point is calculated as follows: This composite weighting can simultaneously consider the proximity of control parameters, the coordination of dynamic trends, and the consistency of operating conditions, so that the surrogate model remains smooth at the mode boundary, while improving the accuracy of local predictions.

[0118] In practical implementation, if the composite weights of all historical training samples are too small (e.g., less than...), (This can be selected by selecting the point closest to the query point in the normalized control parameter space.) Each historical training sample participates in local fitting. This represents the number of nearest neighbor samples participating in the local fitting, which can be taken from 30 to 100. This helps to prevent the local model from becoming unusable when the query point is located in a sparse area of ​​historical samples, and to ensure that the optimization process maintains a reasonable predictive ability for areas with insufficient historical data coverage.

[0119] 6> Three locally weighted linear regression models were constructed with unit coal grinding energy consumption, coal mill outlet temperature and coal mill current as target values, respectively.

[0120] Specifically, for any target value, the target value is denoted as... , The unit coal grinding energy consumption, coal mill outlet temperature, or coal mill current can be taken separately. The locally weighted linear regression coefficient is denoted as... , This refers to the query point relative to target value The regression coefficient vector obtained by performing local linear regression has a size of The vector contains the intercept coefficient, coal feed coefficient, primary air volume coefficient, and separator speed coefficient.

[0121] In the specific implementation, all extended historical regression inputs are used. Construct the historical regression input matrix by row. The corresponding target values ​​are arranged in the same order to form a target value vector. Composite weights Form a diagonal weight matrix The locally weighted linear regression coefficients are solved using the weighted least squares method, and a ridge regression stabilization term can be added. The calculation method is expressed as follows:

[0122] ;

[0123] in, This refers to the historical regression input matrix, with a size of ; express transpose; This refers to the diagonal weight matrix corresponding to the query point, with a size of ; This refers to the target value vector, with a size of ; This represents the ridge regression stability coefficient, used to reduce the risk of instability in matrix inversion. A suitable value is... ; Represents a 4th-order identity matrix; This represents the matrix inversion operation.

[0124] Furthermore, after obtaining the locally weighted linear regression coefficients, the predicted value of the query point is denoted as... The calculation method is expressed as ;in, This refers to the proxy model for query points. The target value prediction result; when the target value is the unit coal grinding energy consumption, the prediction result is recorded as , Indicates query point The unit coal mill energy consumption predicted by the proxy model; when the target value is the coal mill outlet temperature, the prediction result is denoted as... , Indicates query point The surrogate model predicts the coal mill outlet temperature; when the target value is the coal mill current, the prediction result is denoted as... , Indicates query point The surrogate model predicts the coal mill current. By recalculating the composite weights for each query point and solving the local model, the surrogate model can be adaptively updated as the particle position changes, avoiding the problem that a single global model cannot describe the multi-mode nonlinear characteristics.

[0125] S303, Construction of a fitness function that integrates dynamic safety boundaries and pattern deviation penalties

[0126] The goal of optimizing coal mill control parameters is to find a combination of control parameters that reduces unit coal grinding energy consumption while minimizing the need to meet safety constraints on the coal mill outlet temperature and current, and also avoids combinations with insufficient coverage of recommended historical data. This step uses the predicted unit coal grinding energy consumption, the penalty for exceeding the dynamic safety boundary, and the penalty for deviation from the operating mode to construct a fitness function. This function comprehensively reflects the economy, safety, and mode reliability of candidate control parameters, guiding the particle swarm optimization to search for operating conditions that are both supported by historical data and energy-efficient and safe. The specific steps are as follows:

[0127] 1> Based on particle position Construct query points and call the surrogate model to calculate the surrogate model prediction results for each particle position.

[0128] In the specific implementation, the particle position is denoted as... , This refers to the combination of candidate control parameters, i.e. ,in, Indicates the candidate coal feed amount. Indicates the candidate primary air volume. Indicates the rotational speed of the candidate separator. This represents the transpose operation of a vector. It sets the particle position... Inputting the query point into the local composite weighted proxy model yields the predicted unit coal grinding energy consumption. Predicted values ​​of coal mill outlet temperature Predicted current value of coal mill , This refers to the particle position. The unit coal grinding energy consumption predicted by the proxy model; This refers to the particle position. The coal mill outlet temperature predicted by the proxy model; This refers to the particle position. The surrogate model predicts the mill current. The predicted unit mill energy consumption is used to evaluate economic efficiency, while the predicted mill outlet temperature and mill current are used to evaluate safety.

[0129] 2> Calculate the change in coal feed rate based on the candidate coal feed rate and the current actual coal feed rate, and substitute it into S202 to obtain the upper limit of the dynamic coal mill outlet temperature and the upper limit of the dynamic coal mill current.

[0130] In specific implementation, based on and The upper limit of the outlet temperature of the dynamic coal mill was calculated. and the upper limit of dynamic coal mill current When the candidate coal feed rate changes significantly relative to the current actual coal feed rate, the dynamic safety boundary automatically shrinks, enabling the optimization process to maintain a higher safety margin for rapid output adjustments.

[0131] 3> Calculate the penalties for exceeding the limits of the coal mill outlet temperature and the coal mill current based on the dynamic safety boundary and the surrogate model prediction results, respectively.

[0132] In practical implementation, if the predicted value of the coal mill outlet temperature is... Not exceeding the upper limit of the outlet temperature of the dynamic coal mill If the mill outlet temperature exceeds the limit, the penalty is 0; if the predicted mill outlet temperature is... Temperatures exceeding the upper limit of the dynamic coal mill outlet temperature The square of the excess amount will be used as the penalty for exceeding the mill outlet temperature limit. The penalty for exceeding the mill current limit is calculated in the same way, i.e., only when the predicted mill current value... Exceeding the upper limit of dynamic coal mill current Penalties are applied only when necessary. By employing a squared penalty, minor overshoots result in smaller penalties, while severe overshoots result in rapidly increasing penalties, thus guiding particles away from the outer regions of the safety boundary.

[0133] 4> Calculate the deviation penalty of the working mode based on the particle position and the Mahalanobis distance of each working mode.

[0134] In the specific implementation, firstly based on the particle position Determine the most likely operating condition mode by calculating the posterior probabilities of each operating condition mode within the control parameter space. , It means to make The largest operating condition index, This refers to the particle position. Belongs to the The posterior probability of each working condition mode is calculated in the same way as the query point in step 4> of S302. The posterior probability is calculated in the same way; then the particle position is calculated. The Mahalanobis distance to the prior guidance center of the most likely operating condition mode is used to obtain the mode deviation distance. The calculation method is expressed as follows:

[0135] ;

[0136] in, This refers to the degree to which the particle's position deviates from the most likely efficient operating mode. The prior guidance center represents the most likely operating condition mode; The covariance matrix represents the most likely operating condition. The inverse matrix of the covariance matrix; This represents the transpose operation of a vector. Mahalanobis distance can take into account the scale differences and correlations of different control parameters, and is more suitable than ordinary Euclidean distance for describing whether a particle's position is in a dense region of historical operating conditions.

[0137] 5> Combine the predicted unit coal grinding energy consumption, safety constraint penalty, and mode deviation penalty into a fitness function. , Used to evaluate the overall merits of candidate control parameter combinations, the smaller the value, the better the control parameter combination. The calculation method is expressed as follows:

[0138] ;

[0139] in, This indicates that the minimum and maximum unit coal grinding energy consumption will be determined based on the historical training sample database. Dimensionless value after min-max normalization; This represents the temperature penalty normalization coefficient, used to make the squared term of the temperature deviation dimensionless. It can be taken as the typical allowable fluctuation range for safe operation of the coal mill outlet temperature. arrive ; This represents the current penalty normalization coefficient, used to make the square term of the current deviation dimensionless. It can be taken as the typical allowable fluctuation range for safe operation of the coal mill current. arrive ; This represents the penalty coefficient for exceeding the outlet temperature limit of the coal mill, used to control the weight of the penalty for exceeding the outlet temperature limit of the coal mill in the fitness function, and can be taken from 5.0 to 20.0; This represents the over-limit penalty coefficient for the coal mill current, used to control the weight of the over-limit penalty for the coal mill current in the fitness function, and can be taken from 5.0 to 20.0; The mode deviation penalty coefficient is used to control the penalty intensity when a particle deviates from the historical high-efficiency operating condition region, and can be taken from 0.2 to 1.0; This indicates that a non-negative value is taken, and a penalty is only incurred when the predicted value exceeds the corresponding dynamic safety boundary.

[0140] In practical implementation, if the predicted unit coal grinding energy consumption of a candidate control parameter combination is low, but the predicted coal mill outlet temperature exceeds the upper limit of the dynamic coal mill outlet temperature, the fitness function will increase significantly due to the square penalty for exceeding the coal mill outlet temperature limit, and this candidate control parameter combination will not be selected first. If another candidate control parameter combination satisfies the temperature and current constraints, but is far from all historical high-efficiency operating modes, the mode deviation penalty will increase, prompting the particle swarm to prioritize the safe and energy-saving region with historical operating support.

[0141] It should be noted that, The term is a normalized unit coal grinding energy consumption prediction term, used to drive the optimization process to find energy-saving control parameter combinations; This item is a temperature safety penalty item, used to penalize the portion of the predicted outlet temperature that exceeds the limit. When an exceedance occurs, a penalty value proportional to the square of the exceedance is generated, and the greater the exceedance, the greater the penalty. This item is a current safety penalty item, used to penalize the predicted current exceeding the limit. When an over-limit occurs, a penalty value proportional to the square of the over-limit is generated, and the more the limit is exceeded, the greater the penalty. The term is a pattern credibility penalty term, which is used to penalize control parameter combinations that are far from areas with a high density of historical high-efficiency operating conditions, guiding the algorithm to prioritize searching for control parameters that are supported by historical operating data and are more reliable.

[0142] S304, Particle velocity update with mode-guided acceleration and dynamic safety constraint repair.

[0143] Conventional particle swarm optimization (PSO) velocity updates primarily rely on individual historical best positions and the global historical best position. This can easily lead to local optima on the multi-mode energy consumption surface of a coal mill, and position updates may result in predicted limits for the coal mill outlet temperature or current exceeding certain ranges. This step adds a mode guidance term pointing to the prior guidance center of the operating mode during particle velocity updates, and performs control parameter boundary repair and dynamic safety constraint repair after position updates. This guides the particle search direction, prevents recommended control parameters from violating equipment safety constraints, and improves the feasibility of the optimal solution. The specific steps are as follows:

[0144] 1> Calculate the most likely operating mode for each particle based on its current position.

[0145] In specific implementation, the first The particle in the first The position of the next iteration is denoted as... , The size is The control parameter vector; the first The particle in the first Current position of the next iteration The most likely operating condition is denoted as .Will Substituting the values ​​into the posterior probability calculation process of the control parameter space working condition mode in S302, the posterior probability of the particle belonging to each working condition mode is obtained, and the working condition mode with the highest posterior probability is selected as the working condition mode. By updating in each iteration This enables particles to automatically accept new operating mode guidance after crossing mode boundaries.

[0146] 2> The particle velocity is updated jointly based on the inertia term, individual cognition term, social learning term, and pattern guidance term.

[0147] In one implementation, the first The particle in the first The velocity of the next iteration is denoted as... , No. The particle in the first The velocity of the next iteration is denoted as... The speed update method is expressed as:

[0148] ;

[0149] in, Indicates the first The inertia weight of the next iteration is used to control the inheritance ratio of the current speed; Indicates the first The optimal position of an individual particle in its historical context refers to the position of the first particle. The position with the lowest fitness currently found for each particle; The best position in the global history of the particle swarm is the position with the lowest fitness found by all particles so far. Indicates the first The prior guidance center for the most likely operating mode of each particle is determined by the operating mode prior guidance center in step 3> of S203. The calculation results, based on Indexed; This represents the individual's cognitive acceleration coefficient, which can be taken as 1.2 to 1.8. This represents the social learning acceleration coefficient, which can be taken as 1.2 to 1.8; The mode-guided acceleration coefficient can be between 0.5 and 1.0. , and They represent the first The particle in the first In the next iteration, from the interval Independent random numbers drawn from within.

[0150] In this implementation, inertia weight A linear decreasing strategy can be used, and the calculation method is as follows: .in, The initial value for the inertia weight can be 0.9; This is the final value of the inertia weight, which can be taken as 0.4; This represents the maximum number of iterations. Maintaining a larger inertia weight in the early stages of iteration is beneficial for global exploration; decreasing the inertia weight in the later stages of iteration is beneficial for local convergence.

[0151] It should be noted that the mode guidance term can guide particles towards the low-energy control parameter region of their respective operating modes, enabling the search process to not only rely on historical best particles but also utilize the efficient operating mode prior knowledge extracted by S203. When a particle gets stuck in a local inefficient region, the mode guidance term helps the particle move back towards the efficient region of the corresponding operating mode, improving its ability to escape local optima.

[0152] 3> Calculate the new particle position based on the updated velocity and perform upper and lower bound repair of control parameters.

[0153] In one implementation, the first The particle in the first The position of the next iteration is denoted as... The location update method is represented as follows:

[0154] ;

[0155] in, It refers to the first The particle in the first The combination of control parameters for each iteration.

[0156] In practical implementation, after the position update, upper and lower bounds are first corrected for the coal feed rate, primary air volume, and separator speed. If any component of the coal feed rate, primary air volume, or separator speed is lower than the corresponding lower bound, that component is set to the corresponding lower bound; if any component is higher than the corresponding upper bound, that component is set to the corresponding upper bound. Through boundary correction, it can be ensured that the particle position always falls within the allowable setting range of the coal mill control system.

[0157] Furthermore, particle velocities are subject to throttling. The maximum velocity for each control parameter dimension can be set to 10% of the span between the upper and lower bounds of that dimension. If a velocity component exceeds the corresponding maximum velocity, that velocity component is truncated to the corresponding maximum velocity. Velocity throttling prevents particles from making excessive jumps between adjacent iterations, thereby reducing the risk of crossing the safe zone.

[0158] 4> Determine the particle locations where dynamic safety constraint repair needs to be performed based on the surrogate model predictions and dynamic safety boundaries.

[0159] In practical implementation, the local composite weighted surrogate model of S302 is first used to predict the mill outlet temperature and mill current corresponding to the particle position after boundary repair. Then, S202 is used to calculate the upper limit of the dynamic mill outlet temperature and the upper limit of the dynamic mill current corresponding to the particle position. If the predicted values ​​of the mill outlet temperature and the mill current do not exceed the corresponding dynamic safety boundary, the particle position is retained; if the predicted values ​​of the mill outlet temperature or the mill current exceed the corresponding dynamic safety boundary, the particle position is constrained and repaired.

[0160] Furthermore, the coal mill outlet temperature constraint function is denoted as... The coal mill current constraint function is denoted as The calculation methods are expressed as follows:

[0161] ;

[0162] ;

[0163] in, Indicates particle position The degree to which the upper limit of the outlet temperature of the dynamic coal mill is exceeded; Indicates particle position The degree to which the upper limit of the dynamic coal mill current is exceeded. When and When, the particle position satisfies dynamic safety constraints; when or At that time, the particle position violates the dynamic safety constraints.

[0164] 5> For particle positions that violate dynamic safety constraints, perform dynamic safety constraint repair along the descent direction of the constraint function.

[0165] In one implementation, the parameters can be adjusted in the normalized control parameter space. and Calculate the finite difference gradient for each of the three dimensions: coal feed rate, primary air volume, and separator rotation speed. Apply positive and negative small perturbations to that dimension, calculate the change in the constraint function, and obtain the gradient component for that dimension based on this change. A consistent sequence is used to merge the gradient components of the three dimensions into a gradient vector, yielding the coal mill outlet temperature constraint gradient. Coal mill current constraint gradient .

[0166] In practical implementation, if only the coal mill outlet temperature exceeds the limit, then along Move the particle position in the direction; if only the coal mill current exceeds the limit, then move along... The direction of particle movement; if both the coal mill outlet temperature and the coal mill current exceed the limit, the repair direction is weighted and synthesized according to the magnitude of the exceedance, which can be expressed as:

[0167] ;

[0168] in, Indicates the direction of dynamic safety constraint repair, used to indicate the comprehensive direction that reduces constraint overrun; This indicates that the non-negative value operation is performed, and the gradient in this direction only participates in the synthesis of the repair direction when the corresponding constraint function is greater than 0.

[0169] Furthermore, according to The particle position is updated in the manner described above, where, This represents the safe repair step size, used to control the magnitude of each repair movement, and can be taken as 1% to 5% of the normalized control parameter range. After each movement, the upper and lower bound repairs are re-executed, and the surrogate prediction value and dynamic safety boundary are recalculated until the dynamic safety constraints are met or the maximum number of repairs is reached. The maximum number of repairs can be taken as 5 to 10.

[0170] In one embodiment, for example, if a particle's position During the dynamic safety constraint repair process, the predicted temperature is: The upper limit of the outlet temperature of the dynamic coal mill is , The predicted current is The upper limit of the current for dynamic coal mills is , Simultaneously, both temperature and current exceedances exist. The calculated temperature constraint gradient is... The current constraint gradient is The direction of dynamic safety constraint repair is as follows: If we take Then the particle position is updated to After moving, the boundary is re-verified until the constraints are met or the maximum number of repairs is reached.

[0171] It should be noted that if the particle position still does not meet the dynamic safety constraints after reaching the maximum number of repairs, then the particle position will not be considered a feasible recommended position, and a large safety penalty will be imposed in the fitness function, or the particle position will be reverted to its individual historical best position. Nearby. Dynamic safety constraint repair can reduce ineffective searches by the particle swarm in infeasible regions and increase the probability that the final recommended control parameters meet the constraints of the pulverizer outlet temperature and pulverizer current.

[0172] 6> Calculate the fitness of the particle position after repair, and update the individual historical best position and the global historical best position.

[0173] In practical implementation, if the first The fitness of a particle's repaired position is less than its individual best historical position. If the fitness is such that the repaired position is updated to... If the fitness of the repaired position is less than the current global historical best position. If the fitness is such that the repaired position is updated to... Repeat the process of speed update, position update, constraint repair, and optimal position update until the maximum number of iterations is reached. (For example, 50 to 200 iterations) or the global historical best fitness changes less than a preset convergence threshold within a certain number of consecutive iterations (for example, 10 iterations). ).

[0174] 7> The final global historical optimal position is used as the recommended control parameter for the coal mill.

[0175] In the actual implementation, the final global historical optimal position is... This includes recommended coal feed rate, recommended primary air volume, and recommended separator speed. Before being output to the coal mill control system, these values ​​are smoothed according to the control system's setpoint change rate limit to avoid excessive setpoint changes. After smoothing, the recommended coal feed rate, recommended primary air volume, and recommended separator speed are used as control setpoints for the next control cycle to achieve low-energy operation of the coal mill under dynamic safety constraints.

[0176] It should be noted that this step introduces a prior operating condition model in particle swarm optimization. The search direction is constrained by the prior guiding center, the posterior probability of the model, and the mode deviation penalty. This not only preserves the ability of the particle swarm to explore unknown better regions, but also avoids the optimization results from deviating significantly from the historical safe operating region, thereby improving the engineering feasibility of the recommended control parameters.

[0177] S4. Application Process of Online Low-Energy Safety Control for Coal Mills

[0178] After completing the construction of the historical sample library, trend coupling degree extraction, dynamic safety boundary modeling, operating condition pattern recognition, and particle swarm optimization model configuration, this step is deployed to the online control optimization system for the coal mill. During online operation, the system continuously receives coal mill operating data using the same data acquisition method as in the historical training phase and outputs recommended control parameters in each control cycle, such as... Figure 5 As shown, the specific steps are as follows:

[0179] 1> Extract the online sliding window corresponding to the current control cycle from the online operation data stream to obtain the current coal feed window sequence, current primary air volume window sequence, current separator speed window sequence, current pulverizer outlet temperature window sequence, and current pulverizer current window sequence. Current actual coal feed... The coal feed rate is taken from the most recent valid sampling point within the current control period, or the average coal feed rate within a short window of the current control period. Using the current online sliding window allows the optimization to reflect the current actual operating status, rather than relying solely on offline historical statistical results.

[0180] 2> Execute S201 on the current coal feed window sequence, the current primary air volume window sequence, and the current separator speed window sequence to obtain the current trend coupling degree. Current trend coupling degree This is used to characterize the degree of coordinated change in coal feed rate, primary air volume, and separator speed during the current control cycle, and serves as the trend coupling degree similarity input for the local composite weighted surrogate model in S302. If the current trend coupling degree is low, it indicates that the current regulation state is not coordinated. The local surrogate model will prioritize referencing samples that were also in a low-coordination state in the past, thereby improving the consistency between the prediction results and the current state.

[0181] 3> Read the offline-saved steady-state coal mill outlet temperature upper limit curve, steady-state coal mill current upper limit curve, dynamic safety boundary parameters, operating mode prior probability, prior guiding center, and covariance matrix. The dynamic safety boundary parameters include the outlet temperature transient margin coefficient. Current transient margin factor The absolute value of the sensitivity of the upper limit of the steady-state coal mill outlet temperature to the coal feed rate. The absolute value of the sensitivity of the steady-state coal mill current limit to the coal feed rate These parameters remain fixed during the online inference phase to avoid model parameter instability due to short-term fluctuations.

[0182] 4> Based on the current actual coal supply Current trend coupling degree and historical training sample library Using this as input, the particle swarm optimization process from S301 to S304 is executed to obtain the global historical optimal position. Global historical best position This refers to the combination of control parameters with the lowest overall adaptability within the current control cycle, including the recommended coal feed rate. Recommended primary air volume and recommended separator speed .

[0183] 5. Apply rate-limiting and smoothing processing to the recommended control parameters to obtain the final control setpoints. The maximum permissible rate of change of coal feed rate is denoted as... The maximum permissible rate of change of primary air volume is denoted as The maximum permissible rate of change of the separator speed is denoted as These are used to constrain the maximum variation in coal feed rate, primary air volume, and separator speed within adjacent control cycles. If the recommended coal feed rate... Compared with the current actual coal supply If the difference between the two exceeds the maximum allowable variation range of the coal feed rate, the set value of the coal feed rate will be limited to the allowable variation range; the primary air volume and the separator speed will be handled in the same way.

[0184] In practical implementation, a first-order smoothing method can be used to filter the recommended control parameters. The control setpoint of the previous control cycle is denoted as... The recommended control parameters obtained from the current control cycle optimization are denoted as follows: The smoothed current control cycle control setpoint is denoted as You can press Calculate, where, This represents the setpoint smoothing coefficient, which can be between 0.2 and 0.5. By limiting the rate of change and smoothing the process, it is possible to avoid the impact of sudden changes in the recommended control parameters on the operational stability of the coal mill.

[0185] 6. Perform a safety check before outputting the final control setpoint. Input the final control setpoint back into the local composite weighted surrogate model of S302 to obtain the predicted values ​​of the mill outlet temperature and mill current. Input the corresponding coal feed rate and coal feed rate change into S202 to obtain the dynamic upper limit of the mill outlet temperature and the dynamic upper limit of the mill current. If both the predicted mill outlet temperature and the predicted mill current do not exceed the corresponding dynamic safety boundaries, the final control setpoint is output to the mill control system. If either predicted value exceeds the corresponding dynamic safety boundary, the final control setpoint is not output, and the system reverts to the control setpoint of the previous control cycle or the current individual historical optimal safety position.

[0186] In one embodiment, for example, if particle swarm optimization yields recommended control parameters... However, after the rate of change is limited, the final control setpoint is The system again predicted the coal mill outlet temperature to be... The upper limit of the outlet temperature of the dynamic coal mill is The current of the coal mill is The upper limit of the current for dynamic coal mills is If the final control setpoint meets the safety verification, it can be output to the control system; if the predicted coal mill outlet temperature is... And the upper limit of the outlet temperature of the dynamic coal mill is If the final control setting is not output, a re-optimization or conservative rollback strategy will be triggered.

[0187] 7. After the control setpoint is executed, continuously monitor the pulverizer outlet temperature, pulverizer current, and trend coupling degree, and add newly generated safe operation samples to the candidate sample cache. If a new sample does not trigger over-temperature, over-current, or mill blockage alarms within a certain period and the operation response is stable, the new sample can be periodically added to the historical training sample library for model updates. Model updates can be performed offline on a daily, weekly, or maintenance cycle basis. The update content includes steady-state safety boundaries, operating mode parameters, and local proxy model sample libraries. Through online sample backfilling and offline updates, it can adapt to changes in coal quality, equipment wear, and adjustments to operating strategies.

[0188] It should be noted that this invention forms a complete closed loop through trend coupling degree extraction, multi-mode prior modeling, dynamic safety boundary contraction, local composite weighted surrogate prediction, mode-guided particle swarm optimization, and dynamic safety constraint repair. Trend coupling degree is used to characterize the synergy of current control variables; dynamic safety boundary is used to characterize the safety margin of variable load processes; operating condition mode priors are used to guide the optimization search for historically efficient regions; the local composite weighted surrogate model is used to quickly predict the energy consumption and safety response of candidate control parameters; the fitness function is used to uniformly evaluate economy, safety, and mode credibility; and dynamic safety constraint repair is used to reduce infeasible solutions. Based on this, it is possible to output a combination of coal feed rate, primary air volume, and separator speed that is closer to the historically efficient operating region, while ensuring that the coal mill outlet temperature and coal mill current do not exceed limits.

[0189] In one embodiment, as an example, a medium-speed coal mill in a coal-fired power plant uses the method of this invention for online control optimization. The online control system operates on a 60-second control cycle. The current operating condition is as follows: coal feed rate... Primary air volume Separator speed Coal mill outlet temperature Coal mill current After optimization using this method, the recommended control parameter is the coal feed rate. Primary air volume Separator speed The surrogate model predicts that under this parameter, the unit coal grinding energy consumption will decrease by 3.2%, and the coal mill outlet temperature will... (Dynamic upper limit) ), coal mill current (Dynamic upper limit) (This satisfies safety constraints, and recommended parameters are smoothed before being sent for execution.)

[0190] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A control optimization method for optimal output conditions of a coal mill, characterized in that, Includes the following steps: Multi-source operating data of the coal mill are collected and cleaned to construct a historical training sample library. The multi-source operating data includes coal feed rate sequence, primary air volume sequence, separator speed sequence, coal mill outlet temperature sequence, coal mill current sequence, and coal mill power sequence. Based on historical training sample library and current online operation data, the trend coupling degree between coal feed rate, primary air volume and separator speed is extracted, a dynamic safety boundary model is constructed that is adjusted with coal feed rate and coal feed rate change, and the operating mode is identified to obtain the operating condition prior information for subsequent particle swarm optimization. Initial particles are generated based on the prior probability of the working condition mode, the prior guidance center and the covariance matrix, so that the particles are concentrated in the historically efficient or common working condition area during the initialization phase, while retaining the ability to explore other feasible areas. A local composite weighted surrogate model is constructed separately for each candidate control parameter, so that historical training samples that are more similar to the candidate control parameters in terms of control parameter spatial location, trend coupling degree and operating mode are given greater weight. The predicted unit coal grinding energy consumption, the penalty for exceeding the dynamic safety boundary, and the penalty for deviation from the operating mode are combined to form a fitness function, which guides the particle swarm to search for an operating region that is both supported by historical data and energy-saving and safe. In the particle velocity update, a mode guidance term pointing to the prior guidance center of the working condition mode is added, and after the position update, the control parameter boundary repair and dynamic safety constraint repair are performed to guide the search direction of the particles and realize the safe and energy-saving optimization of the control parameters. The optimized setpoints for coal feed rate, primary air volume and separator speed obtained through optimization are sent to the coal mill field control station to drive the actuator and realize closed-loop dynamic adjustment of coal mill output. Specifically, a dynamic safety boundary model that adjusts with the coal feed rate and its variation is constructed as follows: Steady-state operation samples are selected from the historical training sample library. Steady-state operation samples refer to historical training samples when the coal mill is in a relatively stable output state. The upper limits of the steady-state coal mill outlet temperature and the upper limit of the steady-state coal mill current were obtained by kernel smoothing quantile boundary fitting. Finally, based on the change in coal feed rate and boundary sensitivity, transient margin shrinkage is performed to obtain the upper limit of dynamic coal mill outlet temperature and the upper limit of dynamic coal mill current. The specific operating mode identification is as follows: For each historical training sample in the historical training sample library, construct a working condition identification feature vector. The working condition identification feature vector includes coal feed rate, unit coal grinding energy consumption and trend coupling degree. A Gaussian mixture model is used to cluster the normalized working condition identification feature vectors to obtain multiple working condition patterns. The mean value of control parameters corresponding to low-energy consumption samples is extracted from each operating mode and used as a priori guiding center. The mean of the control parameter vectors of all historical training samples within the operating mode is calculated, and then the joint fluctuation of each control parameter component relative to the mean is calculated to obtain the covariance matrix. The diagonal elements of the covariance matrix are used to characterize the fluctuation range of coal feed, primary air volume and separator speed, respectively; the off-diagonal elements are used to characterize the linkage relationship between different control parameters.

2. The control optimization method for optimal output conditions of a coal mill according to claim 1, characterized in that, The trend coupling degree between coal feed rate, primary air volume, and separator speed is extracted as follows: By extracting a fixed-length sliding window from the coal feed rate sequence, primary air volume sequence, and separator speed sequence, we obtain the coal feed rate window sequence, primary air volume window sequence, and separator speed window sequence. Calculate the coal feed rate change sequence, primary air volume change sequence, and separator speed change sequence respectively within the sliding window; The coal feed rate change rate sequence, the primary air volume change rate sequence, and the separator speed change rate sequence are combined into a rate change sequence group, and the cross-correlation matrix is ​​calculated based on the rate change sequence group; The cross-correlation matrix is ​​decomposed into eigenvalues ​​to obtain three eigenvalues, which are then sorted in descending order and L1 normalized. The eigenvalue spectral entropy is calculated based on the normalized eigenvalues, and then converted into trend coupling degree.

3. The control optimization method for optimal output conditions of a coal mill according to claim 1, characterized in that, The initial particles are generated based on the prior probability of the operating mode, the prior guidance center, and the covariance matrix as follows: Set the particle swarm size and upper and lower bounds of the control parameters. The upper and lower bounds of the control parameters are jointly determined by the physical limitations of the equipment, the allowable setting range of the control system, and the historical safe operating extreme values. Extract the working condition mode label of the i-th particle from the prior probability distribution of the working condition mode. The prior probability of the working condition mode is the overall probability that the historical training sample belongs to the m-th working condition mode. The initial position is generated within the upper and lower bounds of the control parameters by using the prior guidance center of the working condition mode to which the i-th particle belongs as the sampling mean and the covariance matrix of the working condition mode as the sampling scattering scale. Initialize particle velocity. The initial velocity is randomly generated within a preset proportion of the allowable range of each control parameter.

4. The control optimization method for optimal output conditions of a coal mill according to claim 1, characterized in that, The method for constructing a local composite weighted proxy model is as follows: Establish historical regression inputs and target values ​​for agent modeling; The control parameter spatial distance weight between the query point and the historical training samples is calculated based on the normalized control parameter spatial distance. Calculate the similarity weight of trend coupling degree between the query point and the historical training samples based on the difference in trend coupling degree. The posterior probability of the query point belonging to each working condition mode in the control parameter space is calculated based on the position of the query point in the control parameter space, and the posterior probability is used as the mode similarity weight. The composite weight is obtained by multiplying the control parameter spatial distance weight, the trend coupling degree similarity weight, and the pattern similarity weight. Three locally weighted linear regression models were constructed with unit coal grinding energy consumption, coal mill outlet temperature and coal mill current as target values, respectively.

5. The control optimization method for optimal output conditions of a coal mill according to claim 1, characterized in that, The fitness function is constructed as follows: Query points are constructed based on particle positions x, and the proxy model is invoked to calculate the proxy model prediction value for each particle position. The change in coal feed rate is calculated based on the candidate coal feed rate and the current actual coal feed rate, and then input into the dynamic safety boundary model to obtain the upper limit of the dynamic coal mill outlet temperature and the upper limit of the dynamic coal mill current. Calculate the penalties for exceeding the limits of the coal mill outlet temperature and the coal mill current based on the dynamic safety boundary and the predicted values ​​of the surrogate model, respectively. The deviation penalty for each operating mode is calculated based on the particle position and the Mahalanobis distance between them. The predicted unit coal grinding energy consumption, safety constraint penalty, and mode deviation penalty are combined into a fitness function. The fitness function is used to evaluate the overall quality of candidate control parameter combinations. The smaller the value, the better the control parameter combination.

6. The control optimization method for optimal output conditions of a coal mill according to claim 5, characterized in that, After the location update, control parameter boundary repair and dynamic safety constraint repair are performed, specifically: Calculate the most likely operating mode for each particle based on its current position; The particle velocity is updated jointly based on the inertia term, individual cognitive term, social learning term, and pattern guidance term; Calculate the new particle position based on the updated velocity and perform upper and lower bound repair of the control parameters; The particle locations where dynamic safety constraint repair needs to be performed are determined based on the surrogate model predictions and dynamic safety boundaries. For particle positions that violate dynamic safety constraints, dynamic safety constraints are repaired along the descent direction of the constraint function; Calculate the fitness of the particle position after repair, and update the individual historical best position and the global historical best position; The final global historical optimal position is used as the recommended control parameters for the coal mill. The final global historical optimal position includes the recommended coal feed rate, recommended primary air volume, and recommended separator speed.

7. The control optimization method for optimal output conditions of a coal mill according to claim 1, characterized in that, The multiple operating modes include high-efficiency mode, transition mode, and low-efficiency mode. The operating mode with the lowest average unit coal grinding energy consumption is defined as high-efficiency mode, the operating mode with the middle average unit coal grinding energy consumption is defined as transition mode, and the operating mode with the highest average unit coal grinding energy consumption is defined as low-efficiency mode.

Images