Multi-objective optimization control method and system for BFB biomass boiler

Abstract

The invention discloses a BFB biomass boiler multi-objective optimization control method and system, and relates to the technical field of industrial process automatic control, and the method comprises the steps: collecting a boiler process amount and an execution amount, and generating a steady standardized residual error; calculating injection disturbance quantity according to the residual covariance intensity and mapping the injection disturbance quantity into a continuous risk weight; determining a local sensitivity matrix on line based on the input and output increments; adaptively reconstructing a multi-target weight matrix and an input penalty matrix in combination with the risk weight and the robust scale; and a control increment is obtained by solving the optimization cost function, and an instruction is generated. According to the method, the disturbance risk is continuously quantified, and the control weight is adaptively reconstructed, so that multi-target evaluation distortion and control over-regulation under a complex fuel working condition are effectively inhibited, and the combustion stability and the comprehensive operation performance of the boiler are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of industrial process automatic control technology, and in particular to a multi-objective optimization control method and system for BFB biomass boilers. Background Technology

[0002] Biomass boilers (BFB) are widely used in industrial heating and power generation. They typically feed biomass fuels such as sawdust, bark, and straw into the furnace through a feeding system, and regulate bed fluidization and combustion through air distribution. In practical engineering applications, biomass fuel sources are complex and their moisture content fluctuates greatly. Bridging or discontinuous feeding of fuel in the silo can easily occur, resulting in pulsed or non-uniform fuel entering the furnace. This requires the control system to simultaneously consider multiple mutually constraining objectives under dynamic operating conditions. It must respond quickly to changes in external load demand, maintain stable bed temperature, ensure complete combustion to reduce carbon monoxide emissions, and maintain the bed pressure differential within a reasonable range to ensure fluidization safety.

[0003] Existing multi-objective control methods typically employ fixed mechanistic models or static logic rules to balance optimization objectives. However, when high-moisture fuel agglomerates enter a high-temperature bed, they undergo a unique physical process of rapid surface drying and crusting followed by internal pressure accumulation and sudden release. This process induces strong disturbances in the gas-solid flow state of the bed, leading to short-term and severe nonlinear fluctuations in process signals such as oxygen, carbon monoxide, and bed pressure. Traditional control strategies struggle to effectively identify these unique disturbances caused by abrupt changes in fuel physical morphology, often misinterpreting them as normal load fluctuations or combustion deterioration, resulting in over-correction and mis-adjustment of control quantities. For example, the controller may erroneously and drastically increase or decrease airflow or feed when emission indicators transiently rise, causing the boiler operating point to deviate from the optimal state. This manifests as amplified load fluctuations, decreased bed temperature stability, and increased fly ash carryover risk, making it impossible to achieve effective coordination and stable operation of various control objectives under complex fuel conditions. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies, such as the tendency for multi-objective evaluation distortion and excessive control correction to occur when facing strong bed disturbances caused by sudden changes in fuel release patterns, leading to amplified boiler load fluctuations and combustion instability. Therefore, this invention proposes a multi-objective optimization control method and system for BFB biomass boilers.

[0005] To address the problems existing in the prior art, the present invention adopts the following technical solution: A multi-objective optimization control method for BFB biomass boilers includes: S1. Collect the process quantities of the boiler and the execution quantities of the actuators; S2. Calculate the robust position and robust scale of process variables and execution variables, and generate standardized residuals of process variables; S3. Calculate the injection disturbance amount based on the covariance strength of the standardized residuals of the process quantity, and map the injection disturbance amount to continuous risk weights. S4. Construct the output vector and execution vector, and determine the local sensitivity matrix based on the output vector and execution vector; S5. Construct a multi-objective weight matrix and input penalty matrix based on robust metrics of process and execution quantities and continuous risk weights; S6. Based on the multi-objective weight matrix, input penalty matrix, and local sensitivity matrix, construct an optimized cost function, solve the cost function to obtain the control increment, and generate execution instructions.

[0006] Preferably, the process quantities include flue gas oxygen content, flue gas carbon monoxide content, bed pressure difference, dust signal, load representative quantity, and bed temperature; the execution quantities include: feeding execution quantity, primary air execution quantity, and secondary air execution quantity.

[0007] Preferably, the robust position and robustness scale of the process quantity and the execution quantity are calculated, and the standardized residuals of the process quantity are generated, including: Obtain a sliding sample set. For each process variable and execution variable, calculate the median on the sliding sample set as a robust position, and calculate the median of the absolute deviation as a robust metric. Use the median of all robust metrics as the global self-sufficient metric; The process quantities are standardized based on their robust location, robust scale, and global self-sufficient scale to obtain standardized residuals; among them, the negative of the standardized residual of flue gas oxygen content is used as the positive injection residual.

[0008] Preferably, the jet disturbance is calculated based on the covariance intensity of the standardized residuals of the process quantity, including: Calculate the absolute values ​​of the covariances of the positive injection residuals of flue gas oxygen content, the standardized residuals of flue gas carbon monoxide content, the standardized residuals of bed pressure difference, and the standardized residuals of dust signal with the standardized residuals of load representative quantities. The absolute value of the covariance is normalized to obtain the normalized intensity coefficient of each process quantity; The jet disturbance is obtained by weighted summation of the absolute values ​​of the standardized residuals of each process quantity based on the normalized intensity coefficient.

[0009] Preferably, determining the local sensitivity matrix based on the output vector and the execution vector includes: The output vector is constructed, which includes load representative quantity, bed temperature, flue gas carbon monoxide content, flue gas oxygen content, and bed pressure difference. Construct an execution vector that includes the feed execution amount, primary air execution amount, and secondary air execution amount; Calculate the increment matrix of the output vector and the increment matrix of the execution vector within the current window; The vectorized median of the input increment matrix is ​​calculated as the sum of the squares of the global self-scaling, and used as a regularization term. The local sensitivity matrix is ​​calculated using the least squares method based on the increment matrix of the execution vector, the increment matrix of the output vector, and the regularization term.

[0010] Preferably, constructing a multi-objective weight matrix includes: The benchmark weighting factor is calculated based on the robust scale of each process quantity and the global self-sufficient scale. The sum of the continuous risk weight and the value of one is used as the first coefficient; Construct a multi-objective weight matrix in the form of a diagonal matrix, where the diagonal elements corresponding to the load representative quantity are the ratio of the baseline weight factor to the first coefficient, and the diagonal elements corresponding to the bed temperature, flue gas carbon monoxide content, flue gas oxygen content, and bed pressure difference are the product of the baseline weight factor and the first coefficient.

[0011] Preferably, constructing the input penalty matrix includes: Based on the robustness scale of each execution quantity and the global self-sufficiency scale, calculate the input change penalty factor; Construct an input penalty matrix in diagonal form, where the diagonal elements of the input penalty matrix consist of input change penalty factors.

[0012] Preferably, based on the multi-objective weight matrix, the input penalty matrix, and the local sensitivity matrix, an optimized cost function is constructed, the control increment is obtained by solving the cost function, and execution instructions are generated, including: Construct reference trajectories for process quantities, where the reference value for the load representative quantity is the external load command, and the reference values ​​for bed temperature, flue gas carbon monoxide content, flue gas oxygen content, and bed pressure difference are the corresponding robust positions. A one-step prediction model is established based on the local sensitivity matrix. The one-step prediction model is as follows: In the formula, This is the predicted output vector for the next time step. This is the output vector at the current time. For the control increment to be determined, This is the local sensitivity matrix; Construct an optimization cost function, wherein the optimization cost function is: In the formula, To optimize the cost function, As the reference trajectory vector, For multi-objective weight matrix, Given the input penalty matrix, For the control increment to be determined, This is the predicted output vector for the next time step; The optimal closed-loop control increment is obtained by minimizing the cost function. The reciprocal of the first coefficient is used as the feed attenuation coefficient; The feed component in the optimal control increment is attenuated and corrected according to the feed attenuation coefficient to obtain the final feed execution amount; The primary air execution volume, secondary air execution volume, and final feeding execution volume in the optimal control increment are sent to the actuator.

[0013] To address the above problems, the present invention also provides a multi-objective optimization control system for a BFB biomass boiler, the system comprising: The data acquisition module is used to periodically collect the process quantities of the boiler and the execution quantities of the actuators. The robust residual module is used to calculate the robust position and robust scale of process and execution quantities, and generate the standardized residuals of process and execution quantities. The risk weight module is used to calculate the injection disturbance based on the covariance strength of the standardized residuals of the process quantity, and to map the injection disturbance to continuous risk weights. The sensitivity matrix module is used to construct the output vector and the execution vector, and to determine the local sensitivity matrix based on the output vector and the execution vector. The weight matrix module is used to construct multi-objective weight matrices and input penalty matrices based on robust scales of process and execution quantities and continuous risk weights. The instruction generation module is used to construct an optimized cost function based on the multi-objective weight matrix, the input penalty matrix, and the local sensitivity matrix, solve the cost function to obtain the control increment, and generate execution instructions.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention achieves continuous quantitative characterization of the intensity of water-locked coke shell bursting injection disturbance by constructing a standardized residual and covariance intensity analysis based on robust statistics, and adaptively reconstructs the multi-objective weight matrix accordingly. When an increase in injection risk is detected, the tracking weight of the load representative quantity is automatically reduced while the constraint weights of bed temperature, burnout, and bed pressure difference are simultaneously enhanced. This enables accurate identification and adaptation to strong gas-solid flow disturbances caused by sudden changes in fuel release patterns, avoiding misjudging such nonlinear changes as normal load fluctuations or combustion deterioration. It effectively suppresses multi-objective evaluation distortion and excessive correction of control quantities, and improves combustion stability and emission control levels under complex fuel conditions.

[0015] 2. This invention utilizes the input-output increments within a sliding window to estimate the local sensitivity matrix online, and updates the marginal effects of feed and air distribution on process quantities in real time to match the changing dynamic characteristics of the system during jet disturbances. Combined with the input penalty matrix and a feed attenuation correction mechanism based on risk weights, the adjustment amplitude and frequency of the actuator are automatically limited during the optimization process, especially the feed increment is attenuated in a targeted manner. This eliminates secondary flow disturbances and mechanical shocks caused by model mismatch or aggressive adjustments, ensuring that load response and stable system operation are optimized in a coordinated manner while guaranteeing bed fluidization safety and burnout sufficiency. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a multi-objective optimization control method for a BFB biomass boiler according to an embodiment of the present invention. Figure 2 This is a functional block diagram of a multi-objective optimization control system for a BFB biomass boiler provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Example 1: This example provides a multi-objective optimization control method for a BFB biomass boiler. See [link to example]. Figure 1 Specifically, including: S1. Periodically collect the process quantities of the boiler and the execution quantities of the actuators; In embodiments of the present invention, periodically collecting process quantities of the boiler and execution quantities of the actuator includes: The process parameters collected include flue gas oxygen content, flue gas carbon monoxide content, bed pressure difference, dust signal, load representative quantity, and bed temperature. Collect execution data, which includes material feeding execution data, primary air execution data, and secondary air execution data; Specifically, flue gas oxygen content refers to the volume fraction or equivalent concentration of oxygen in the flue gas after combustion in the furnace, reflecting the degree of surplus between the supplied air and fuel reaction consumption, as well as the level of excess air; flue gas carbon monoxide content refers to the volume fraction or equivalent concentration of carbon monoxide in the flue gas. Carbon monoxide originates from the incomplete oxidation process of fuel volatiles and carbon particles under oxygen-deficient or insufficiently mixed conditions, thus characterizing the completeness of combustion and the degree of local oxygen deficiency; bed pressure difference refers to the pressure difference between the upper and lower pressure measurement points of the bed, which is determined by the weight of the bed material particles, gas... The fluidity intensity, filling state, and gas-solid distribution stability of the bed are determined by the resistance between particles and the flow impedance caused by bubble and particle agglomeration. The flue gas dust signal refers to the response of particulate matter in the flue gas to optical transmission or scattering, or the equivalent concentration output by a particulate matter measuring device. This signal reflects the carrying level and transient rise of fine ash and unburned carbon particles in the flue. The load representative quantity refers to a value that characterizes the boiler's output capacity, preferably using main steam flow rate, equivalent thermal power, or power generation capacity. Measurement parameters reflect the actual level of steam and heat production and load tracking status of the boiler; bed temperature refers to the average or representative temperature of the gas-solid two-phase system within the bed, which is determined by fuel heat release, particle and gas convective heat transfer, and moisture evaporation endothermic heat transfer, and is an important state parameter reflecting reaction rate, burnout trend, and slagging and sintering risk; feed rate refers to the mass flow rate of fuel fed into the furnace by the feeding device per unit time, or its equivalent feed mechanism speed or frequency, which determines the input intensity of fuel chemical energy and directly affects volatile matter release and heat release within the bed. The primary air flow rate refers to the airflow rate or equivalent fan frequency delivered into the bottom air chamber of the bed and through the bed material. This flow rate mainly provides the momentum required for bed fluidization and some of the oxygen required for combustion, thereby affecting the bed pressure difference, the suspension state of the bed material, and the mixing within the bed. The secondary air flow rate refers to the airflow rate or equivalent fan frequency injected into the upper part of the furnace or above the bed surface. This flow rate is mainly used to supplement the oxygen required for upper combustion and enhance gas mixing, further oxidizing volatiles and combustible gases, thereby affecting the changes in the oxygen content, carbon monoxide content, and dust-carrying state of the flue gas.

[0019] In detail, during boiler operation, process quantities and execution quantities are synchronously collected at a fixed sampling period to form a time-series data stream. The sampling period is preferably one to two seconds to cover the main dynamic range of gas-solid flow and combustion changes in the BFB bed, while also taking into account signal jitter and computational load in the industrial setting. The oxygen content of the flue gas is obtained through the output signal of the flue gas oxygen analyzer via fieldbus; the carbon monoxide content of the flue gas is obtained through the output signal of the flue gas carbon monoxide analyzer; the differential pressure output of the bed differential pressure transmitter is obtained through the differential pressure output of the bed differential pressure transmitter; the dust signal is obtained through the output of the flue gas opacity meter or particulate matter online monitoring device; the load representative quantity is obtained through the output of the steam flow meter or power meter; and the bed temperature is obtained through the output of the bed thermocouple or resistance temperature measuring device. At the end of each sampling period, the instantaneous values ​​of each channel are read, and range consistency and engineering unit conversion are performed to obtain the volume fraction or mass fraction. Standard values ​​in units of concentration or Pa are used, and each channel is timestamped and identified at the same time to form a process vector for subsequent calculations. Simultaneously, execution quantities are collected at the same sampling period: the current feed execution quantity is read from the feed inverter or feed metering device, the current primary air execution quantity is read from the primary fan inverter or airflow measurement circuit, and the current secondary air execution quantity is read from the secondary fan inverter or airflow measurement circuit. The execution quantities are preferably uniformly expressed as frequency percentage or standard volumetric flow rate to match the input format for subsequent optimization calculations. During acquisition, the execution quantity signal is subjected to amplitude limiting checks and compared with the equipment's rated boundary to eliminate abnormal jump points. Finally, within each sampling period, the process vector and execution quantity vector are written to the cache queue and historical database with the same timestamp, and the sliding sample window is updated to ensure the data continuity and traceability required for subsequent robust benchmark calculations and multi-objective optimization control solutions.

[0020] S2. Calculate the robust position and robust scale of process and execution quantities, and generate the standardized residuals of process and execution quantities; In embodiments of the present invention, calculating the robust position and robustness scale of process variables and execution variables, and generating standardized residuals of process variables, includes: Obtain a sliding sample set. For each process variable and execution variable, calculate the median on the sliding sample set as a robust position, and calculate the median of the absolute deviation as a robust metric. Use the median of all robust metrics as the global self-sufficient metric; Standardized residuals are obtained by standardizing process quantities based on their robust location, robust scale, and global self-sufficient scale. Standardized residuals refer to the dimensionless quantity obtained by differentiating the instantaneous value of a process quantity at the current sampling time from its robust position on the sliding sample set, and normalizing the difference result by the sum of the robust scale of the process quantity on the sliding sample set and the global self-sufficient scale. It is used to characterize the degree of deviation of the process quantity from the typical level of the current operating condition and to make process quantities of different dimensions comparable. The robust position is preferably the median of the process quantity sample sequence to reduce the influence of a small number of outliers on the center position. The robust scale is preferably the median of the absolute deviation of the process quantity sample sequence from the robust position to reflect the typical fluctuation amplitude and avoid amplification by extreme values. The global self-sufficient scale is the median of the robust scales of all process quantities and execution quantities to provide a unified numerical stability base to prevent the residual from being abnormally amplified due to the denominator being too small. Therefore, when the bursting and spraying of the water-locking coke shell causes a sudden deviation of certain process quantities, the standardized residual of the process quantity will produce a significant increase or decrease in response. Among them, water-locking coke shell bursting injection refers to the physical phenomenon in which compacted high-moisture biomass fuel enters a high-temperature bed, and its surface rapidly dehydrates and carbonizes to form a low-permeability coke shell, thereby locking the moisture inside. As the internal moisture violently vaporizes, the internal pressure rises and eventually breaks through the coke shell, instantly and concentratedly ejecting a large number of volatile clusters and fine carbon particles. This phenomenon will cause strong disturbances in the gas-solid flow state of the bed, resulting in brief and counterintuitive violent fluctuations in signals such as CO, bed pressure, and oxygen content. It is the core interference source causing unstable combustion in biomass boilers.

[0021] In detail, after periodically collecting process and execution quantities, the process and execution vectors formed in each sampling period are written into a buffer queue in chronological order. A sliding sample set is formed by the most recent continuous samples in the buffer queue. The window length of the sliding sample set is preferably 60 to 120 sampling points to cover the typical duration of disturbances related to water-locking coke shell bursting and injection, while also taking into account the real-time computation. When the sampling period is one to two seconds, a statistical window of one to four minutes can be formed to maintain robustness when fuel fluctuations and bed flow fluctuations coexist. For each process and execution quantity, the corresponding time series is extracted from the sliding sample set, and the median of the time series is calculated as the robust position, denoted as the robust position. The robust position is used to characterize the typical level of the quantity under the current operating conditions and is not sensitive to occasional spikes. Then, the difference between each sample in the time series and the robust position is calculated, and the absolute value is taken to form an absolute deviation sequence. The median of the absolute deviation sequence is calculated as the robust scale, denoted as the robust scale. The robust scale is used to characterize the typical fluctuation amplitude of the quantity under the current operating conditions and avoids being amplified by a few outliers.

[0022] Within the same sampling period, robust scales corresponding to all process and execution quantities are aggregated, and the median of these robust scales is calculated as the global self-sufficient scale. The global self-sufficient scale is used to provide a unified numerical stability base to avoid excessively small standardized denominators when some quantities fluctuate very little in a short period of time. After completing the calculation of robust positions, robust scales, and global self-sufficient scales, each process quantity is processed according to the following standardization formula to obtain the standardized residual. Let the instantaneous value of a process quantity in the current sampling period be its current value, the corresponding robust position be its robust position, and the corresponding robust scale be its robust scale. Then, the standardized residual is equal to the current value of the process quantity minus the difference between its robust positions, divided by the sum of its robust scale and the global self-sufficient scale. This converts process quantities with different dimensions into comparable dimensionless quantities. The residuals generate corresponding residual responses when local abnormal rises or suppressions are caused by the bursting of the water-locked coke shell. For the standardized residuals of flue gas oxygen content, the negative value is taken and defined as the positive injection residual. This ensures that the decrease in oxygen content caused by the consumption of oxygen by local reducing gas clouds when the perturbation of the bursting of the water-locked coke shell is enhanced can be represented by an increase in the value of the positive injection residual. This allows for a quantitative characterization that is in the same direction as the changes in the injection correlation, such as the increase in carbon monoxide content, the enhancement of bed pressure pulsation, and the rise in dust signal, and facilitates the construction of subsequent continuous risk weights. Finally, the standardized residuals of each process quantity are arranged in the same order as the process quantity vector to form a residual vector, which, together with the robust position, robust scale, and global self-sufficient scale, is written into the internal state cache for direct use in subsequent risk weight calculations and multi-objective optimization solutions.

[0023] S3. Calculate the injection disturbance amount based on the covariance strength of the standardized residuals of the process quantity, and map the injection disturbance amount to continuous risk weights. In an embodiment of the present invention, the injection disturbance is calculated based on the covariance strength of the standardized residuals of the process quantity, and the injection disturbance is mapped to a continuous risk weight, including: Calculate the absolute values ​​of the covariances of the positive injection residuals of flue gas oxygen content, the standardized residuals of flue gas carbon monoxide content, the standardized residuals of bed pressure difference, and the standardized residuals of dust signal with the standardized residuals of load representative quantities. The absolute value of the covariance is normalized to obtain the normalized intensity coefficient of each process quantity; Specifically, the normalized intensity coefficient is a dimensionless weighting coefficient used to characterize the proportion of the intensity of the common change of a certain injection-related process quantity to the load representative quantity fluctuation. It is obtained by normalizing the absolute value of the covariance between the residual sequence of the process quantity and the residual sequence of the load representative quantity in the sliding sample set, so that the sum of the corresponding coefficients of each process quantity is one. This reflects the relative contribution of the positive residual of oxygen content injection, the standardized residual of carbon monoxide, the standardized residual of bed pressure difference, and the standardized residual of dust signal to the intensity of load fluctuation correlation under the current operating conditions, and is used as the weight in the weighted calculation.

[0024] In detail, after obtaining the standardized residuals of each process quantity within the sliding sample set, four residual sequences associated with the water-locking coke shell bursting injection disturbance and a standardized residual sequence of the load representative quantity are extracted from the sliding sample set. These four residual sequences are, respectively, the injection positive residual sequence of flue gas oxygen content, the standardized residual sequence of flue gas carbon monoxide content, the standardized residual sequence of bed pressure difference, and the standardized residual sequence of dust signal. The standardized residual sequence of the load representative quantity is used to characterize the fluctuation amplitude and direction on the boiler output side. For each time point index within the sliding sample set, the above four residuals and the load representative quantity residual are paired according to the same timestamp. Four sets of residual pairs are formed, and the covariance is calculated for each set of residual pairs. The covariance is calculated using the sample mean in the moving sample set as the mean-reduction benchmark to reflect the common fluctuation intensity among the residuals. Specifically, the residual sequence of a certain process quantity is denoted as Sequence 1, and the residual sequence of the load representative quantity is denoted as Sequence 2. The mean of Sequence 1 and the mean of Sequence 2 are calculated respectively. Then, the product of the deviations of Sequence 1 and Sequence 2 is calculated for all sample indices in the moving sample set, and the sum is divided by the number of samples to obtain the covariance value. Then, the absolute value of the covariance value is taken to obtain the absolute value of the covariance, so as to eliminate the difference in positive and negative correlation directions and make the obtained quantity only represent the common fluctuation intensity.

[0025] The absolute covariance values ​​of the positive residual of flue gas oxygen content injection and the residual of the load representative quantity are obtained, as are the absolute covariance values ​​of the residual of carbon monoxide, the residual of bed pressure difference and the residual of the load representative quantity, and the absolute covariance values ​​of the residual of dust signal and the residual of the load representative quantity. These four absolute covariance values ​​are then summed to form an intensity set. To avoid the normalization denominator being zero and to maintain numerical stability, a global self-sufficient scale is added as a stability term when summing the intensity sets. The global self-sufficient scale is preferably the median of all the aforementioned robust scales to reflect the overall fluctuation scale of the system. Subsequently, each absolute covariance value is divided by the sum of the four absolute covariance values ​​and the sum of the global self-sufficient scale to obtain a normalized intensity coefficient. Thus, the sum of the four normalized intensity coefficients is used to characterize the contribution ratio of each process quantity to the load fluctuation in the current sliding sample set.

[0026] The jet disturbance is obtained by weighted summation of the absolute values ​​of the standardized residuals of each process quantity based on the normalized intensity coefficient; The jet disturbance is mapped to a continuous risk weight; It should be noted that the jet disturbance amount refers to a dimensionless comprehensive index that continuously quantifies the strength of the jet disturbance caused by the bursting of the water-locked coke shell. It is obtained by taking the absolute value of the standardized residuals of the above-mentioned process quantities at the current moment and weighting them according to the normalized intensity coefficient. The larger the jet disturbance amount, the greater the deviation of the process quantity from the robust benchmark, and the more the distribution of the deviation among the process quantities conforms to the intensity structure related to load fluctuations. Thus, it characterizes the significance of the comprehensive disturbances caused by the jet, such as the enhanced bed pulsation, the lifting of dust, the decrease in oxygen content due to the reducing gas cloud, and the increase in carbon monoxide due to incomplete combustion.

[0027] In detail, after obtaining the normalized intensity coefficients of each process quantity, the absolute values ​​of the injection positive residual of flue gas oxygen content, the normalized residual of flue gas carbon monoxide content, the normalized residual of bed pressure difference, and the normalized residual of dust signal are taken at the same sampling time to eliminate the influence of the sign of the residual on the quantification of disturbance intensity and to ensure that the obtained quantity only represents the magnitude of deviation from the robust benchmark. Subsequently, the above four absolute values ​​are multiplied by the corresponding normalized intensity coefficients and weighted to obtain the injection disturbance quantity. Specifically, the normalized intensity coefficients are denoted as oxygen content weight, carbon monoxide weight, pressure difference weight, and dust weight, respectively, and the corresponding absolute values ​​of the residuals are denoted as oxygen content residual amplitude, carbon monoxide residual amplitude, pressure difference residual amplitude, and dust residual amplitude, respectively. The difference amplitude value indicates that the injection disturbance is the sum of the products of the four weights and the four residual amplitude values. After obtaining the injection disturbance, the injection disturbance is mapped to a continuous risk weight for subsequent multi-objective weight adaptive reconstruction. The mapping adopts a bounded monotonically increasing fractional function to ensure that the risk weight is between zero and one and increases smoothly with the increase of the injection disturbance. Preferably, the ratio of the square of the injection disturbance to the square plus one of the injection disturbance is used as the continuous risk weight. That is, the continuous risk weight is obtained by dividing the square of the injection disturbance by the square plus one of the injection disturbance. This ensures that the risk weight is insensitive to small disturbances when the injection disturbance is small and gradually approaches one when the injection disturbance increases, so as to highlight the influence of water-locking coke shell bursting injection on the control trade-off.

[0028] S4. Construct the output vector and execution vector, and determine the local sensitivity matrix based on the output vector and execution vector; In embodiments of the present invention, constructing an output vector and an execution vector, and determining a local sensitivity matrix based on the output vector and the execution vector, includes: Construct the output vector and execution vector; Calculate the increment matrix of the output vector and the increment matrix of the execution vector within the current window; In detail, after collecting process and execution quantities and updating the sliding sample set in each sampling period, the process quantities used for multi-objective optimization are arranged into an output vector in a predetermined order, and the controllable execution quantities are arranged into an execution vector in a predetermined order. The output vector preferably includes load representative quantities, bed temperature, flue gas carbon monoxide content, flue gas oxygen content, and bed pressure difference, so that the output vector covers both load-side objectives and burnout and bed state objectives related to water-locked coke shell bursting injection. The execution vector preferably includes feed execution quantities, primary air execution quantities, and secondary air execution quantities to correspond to the adjustable quantities of fuel input intensity and air distribution conditions. After constructing the output and execution vectors, the output vector sample and execution vector sample for each moment are read sequentially from the current window, and the difference between vector samples from adjacent sampling moments is calculated to obtain... The vector increment is specifically defined as follows: the output vector at the k-th sampling time within the window is denoted as the output vector sample, and the execution vector at the k-th sampling time within the window is denoted as the execution vector sample. The vector increment of the output vector is the current output vector sample minus the output vector sample at the previous time, and the vector increment of the execution vector is the current execution vector sample minus the execution vector sample at the previous time. The output vector increments of all adjacent time points within the window are concatenated column-wise to form the output vector increment matrix, and the execution vector increments of all adjacent time points within the window are concatenated column-wise to form the execution vector increment matrix. Thus, the output vector increment matrix represents the change trajectory of output states such as load, bed temperature, discharge, and bed pressure difference within the window, and the execution vector increment matrix represents the change trajectory of control quantities such as feed and primary and secondary air within the window.

[0029] The vectorized median of the input increment matrix is ​​calculated as the sum of the squares of the global self-scaling, and used as a regularization term. The local sensitivity matrix is ​​calculated using the least squares method based on the increment matrix of the execution vector, the increment matrix of the output vector, and the regularization term. It should be noted that the regularization term is a numerical stability quantity introduced when performing least squares calculations based on the execution vector increment matrix and the output vector increment matrix. It is obtained by adding the square of the vectorized median of the input increment matrix and the square of the global self-sufficiency scale. The former reflects the typical amplitude of the changes in the control quantities such as feed and air distribution within the current window, while the latter reflects the overall fluctuation scale of the system and is used to prevent numerical amplification caused by the excessively small robustness scale of certain quantities. The sum of the squares of the two is used to improve the invertibility of the matrix and suppress the excessive amplification of the solution when the input autocorrelation matrix is ​​close to singular or the sample excitation is insufficient, so that the estimation results are more robust to accidental small sample jumps or measurement noise. The local sensitivity matrix refers to the matrix that describes the marginal mapping relationship between the small changes of each component of the execution vector and the changes of each component of the output vector under the actual operating conditions represented by the current sliding window. It is obtained by regularized least squares fitting of the output vector increment matrix to the execution vector increment matrix. The elements in the matrix represent the corresponding changes in the output quantities such as bed temperature, flue gas carbon monoxide content, flue gas oxygen content, and bed pressure difference when the feed execution quantity, primary air execution quantity, or secondary air execution quantity changes by a unit. Therefore, it can characterize the instantaneous characteristics of the control effect relationship after disturbances such as water-locking coke shell bursting and injection change the fuel release pattern and bed flow state.

[0030] In detail, after obtaining the increment matrix of the execution vector and the increment matrix of the output vector within the current window, the increment matrix of the execution vector is first vectorized to form a one-dimensional sample set. This vectorization is achieved by expanding the increment matrix column by column and concatenating it into a vector sequence, so that the magnitude of each execution change within the window is used as a sample for robust statistics. Then, the absolute value of the vector sequence is taken and the median is calculated. The median is used as the vectorized median of the input increment matrix to reflect the typical magnitude of the execution change within the current window without being affected by a few abnormal jump samples. After obtaining the vectorized median of the input increment matrix, the square of the median is summed with the square of the global self-given scale, and the sum is defined as a regularization term. This allows the regularization term to simultaneously reflect the typical magnitude of the manipulation and the overall fluctuation scale of the system, and to provide stable numerical support when the magnitude of the execution change is too small or the increment matrix is ​​approximately singular. The local sensitivity matrix is ​​calculated using the least squares method. Specifically, the increment matrix of the output vector and the increment matrix of the execution vector are used as the response matrix and input matrix of the regression, respectively. An input autocorrelation matrix is ​​constructed by multiplying the increment matrix of the execution vector by its transpose. The regularization term is multiplied by the identity matrix and then added to the input autocorrelation matrix to form a regularized invertible matrix. The increment matrix of the output vector is then multiplied by the transpose of the increment matrix of the execution vector to obtain a cross matrix. Finally, the cross matrix is ​​multiplied by the regularized invertible matrix to obtain the local sensitivity matrix. Thus, the local sensitivity matrix represents the marginal effect of changes in the feed execution rate, primary air execution rate, and secondary air execution rate on changes in load representative quantities such as bed temperature, carbon monoxide content, oxygen content, and bed pressure difference under the current window operating conditions. It is then used for the closed-form solution of subsequent multi-objective optimization control increments.

[0031] S5. Construct a multi-objective weight matrix and input penalty matrix based on robust metrics of process and execution quantities and continuous risk weights; In embodiments of the present invention, a multi-objective weight matrix and an input penalty matrix are constructed based on robust metrics of process quantities and execution quantities, as well as continuous risk weights, including: The benchmark weighting factor is calculated based on the robust scale of each process quantity and the global self-sufficient scale. The sum of the continuous risk weight and the value of one is used as the first coefficient; Construct a multi-objective weight matrix in the form of a diagonal matrix, where the diagonal elements corresponding to the load representative quantity are the ratio of the baseline weight factor to the first coefficient, and the diagonal elements corresponding to the bed temperature, flue gas carbon monoxide content, flue gas oxygen content, and bed pressure difference are the product of the baseline weight factor and the first coefficient. It should be noted that the benchmark weighting factor is a dimensionless basic weight used to unify the deviation penalties of different process quantities to a comparable scale. It is jointly determined by the robust scale of the corresponding process quantity in the sliding sample set and the global self-sufficient scale. The robust scale represents the typical fluctuation amplitude of the process quantity under the current operating conditions, while the global self-sufficient scale represents the overall fluctuation level of the system and provides numerical stability. The benchmark weighting factor is formed by adding the two together to obtain the comprehensive fluctuation scale and taking the reciprocal of its square. This ensures that process quantities with large fluctuation amplitudes will not naturally dominate in optimization due to their numerical amplitude, and process quantities with small fluctuation amplitudes will not be ignored due to differences in dimensions. This guarantees the penalty for each process quantity in multi-objective optimization. The intensity matches its actual fluctuation characteristics; the multi-objective weight matrix refers to the diagonal matrix used to allocate penalty intensity to different output deviation terms in the multi-objective optimization cost function. Its diagonal elements correspond to the weight coefficients of the load representative quantity, bed temperature, flue gas carbon monoxide content, flue gas oxygen content, and bed pressure difference, respectively. Among them, the diagonal element of the load representative quantity takes the ratio of the benchmark weight factor to the first coefficient to reduce the weight of the load term when the risk increases. The diagonal elements of other process quantities take the product of the benchmark weight factor and the first coefficient to increase the weight of the bed temperature, burnout oxygen content, and bed stability terms when the risk increases. Thus, the trade-off direction of multi-objective optimization under the water-locking coke shell bursting injection condition is consistent with the physical risk.

[0032] In detail, after updating the sliding sample set and obtaining the robust scale and global self-sufficient scale for each process quantity in each sampling period, a benchmark weight factor is first calculated for each process quantity in the output vector. This ensures that process quantities with different dimensions and fluctuation amplitudes have comparable penalty scales in multi-objective optimization. The benchmark weight factor is obtained by adding the robust scale and the global self-sufficient scale of the corresponding process quantity, taking the reciprocal, and then squaring the result. Specifically, the sum of the robust scale and the global self-sufficient scale of a certain process quantity is used as the comprehensive fluctuation scale of that process quantity, and the square reciprocal of the comprehensive fluctuation scale is used as the benchmark weight factor. This ensures that the benchmark weight factor is smaller for process quantities with greater fluctuations to avoid that process quantity dominating the optimization solution and to give process quantities with smaller fluctuations higher relative constraint strength. After obtaining the continuous risk weight, the continuous risk weight is summed with a numerical value to obtain the first coefficient, thereby transforming the continuous risk weight from a risk factor... The strength and weakness quantities are transformed into a stable modulation coefficient, ensuring that the control law automatically degenerates to the normal operating condition when there is no injection risk, without making the weights zero or infinitely large. Subsequently, a multi-objective weight matrix is ​​constructed. The multi-objective weight matrix adopts a diagonal matrix form to ensure that the weights of each objective item are independently interpretable. The diagonal elements corresponding to the load representative quantity are set as the ratio of the baseline weight factor of the load representative quantity to the first coefficient. This automatically reduces the penalty weight of aggressive load tracking and suppresses the fluctuation amplification caused by excessive load tracking when the injection risk increases. At the same time, the diagonal elements corresponding to bed temperature, flue gas carbon monoxide content, flue gas oxygen content, and bed pressure difference are respectively set as the product of their respective baseline weight factors and the first coefficient. This automatically strengthens the constraint intensity on bed temperature stability, complete combustion, oxygen level, and bed flow stability when the injection risk increases. Finally, a multi-objective weight matrix is ​​obtained for subsequent cost function construction and control increment solution.

[0033] Based on the robustness scale of each execution quantity and the global self-sufficiency scale, calculate the input change penalty factor; Construct an input penalty matrix in the form of a diagonal matrix, where the diagonal elements of the input penalty matrix consist of input change penalty factors; It should be noted that the input penalty matrix refers to the diagonal matrix used in the multi-objective optimization cost function to assign penalty intensity to the execution quantity change term. Its diagonal elements correspond to the input change penalty factors of the feed execution quantity, primary air execution quantity, and secondary air execution quantity, respectively. By writing each input change penalty factor into the diagonal elements, independent constraints are achieved on the different execution quantity change amplitudes, thereby suppressing frequent and large-scale adjustments to feed and air distribution and reducing secondary disturbances introduced to the bed flow state and combustion process, while maintaining the numerical good state and interpretability of the cost function.

[0034] In detail, after updating the sliding sample set and obtaining the robust scale and global self-sufficient scale for each execution quantity in each sampling period, in order to suppress frequent and large fluctuations in the feed and air distribution control quantities and avoid mechanical shock or secondary disturbances in the gas-solid flow state caused by excessively rapid changes in the actuator, an input change penalty factor is first calculated for the feed execution quantity, primary air execution quantity, and secondary air execution quantity respectively. The input change penalty factor is used to unify the change amplitude of different execution quantities to a comparable penalty scale and ensure the numerical stability of the cost function with respect to the input change term. Specifically, for any execution quantity, its robust scale in the sliding sample set is taken as the typical fluctuation amplitude of the execution quantity, and the robust scale is added to the global self-sufficient scale to obtain the comprehensive fluctuation scale of the execution quantity. Then, the reciprocal of the comprehensive fluctuation scale is taken and the square is taken to obtain the corresponding input change penalty factor, so that when When a certain execution quantity naturally fluctuates significantly under the current operating conditions, its penalty factor is relatively reduced to avoid the change term of that execution quantity excessively dominating the optimization solution. Conversely, when a certain execution quantity fluctuates less, its penalty factor is relatively increased to suppress unnecessary minor fluctuations. After obtaining the primary input change penalty factor of the feed execution quantity, the primary air execution quantity, and the secondary air execution quantity, an input penalty matrix is ​​constructed. The input penalty matrix adopts a diagonal matrix form to ensure that the penalties for each execution quantity change are independent and easy to interpret. The diagonal elements of the input penalty matrix are sequentially set as the primary input change penalty factor of the feed execution quantity, the primary air execution quantity, and the secondary air execution quantity, thereby obtaining the input penalty matrix used for subsequent cost function construction and control incremental closed-form solution.

[0035] S6. Based on the multi-objective weight matrix, input penalty matrix, and local sensitivity matrix, construct an optimized cost function, solve the cost function to obtain the control increment, and generate execution instructions; In an embodiment of the present invention, an optimized cost function is constructed based on a multi-objective weight matrix, an input penalty matrix, and a local sensitivity matrix. The cost function is solved to obtain the control increment, and execution instructions are generated, including: Construct reference trajectories for process quantities, where the reference value for the load representative quantity is the external load command, and the reference values ​​for bed temperature, flue gas carbon monoxide content, flue gas oxygen content, and bed pressure difference are the corresponding robust positions. A one-step prediction model is established based on the local sensitivity matrix; In detail, after obtaining the output vector and completing the robust position calculation in each sampling period, a reference trajectory of the process quantity is first constructed as the expected target of multi-objective optimization. The reference value of the load representative quantity is taken from the external load command, which comes from the upper-level dispatch system or the user-side load demand signal and is expressed in engineering units as the expected output level of the boiler. The reference value of the bed temperature is taken as the robust position of the bed temperature in the sliding sample set. The reference value of the flue gas carbon monoxide content is taken as the robust position of the carbon monoxide content in the sliding sample set. The reference value of the flue gas oxygen content is taken as the robust position of the oxygen content in the sliding sample set. The reference value of the bed pressure difference is taken as the robust position of the bed pressure difference in the sliding sample set. The reason for using the robust position as the reference value is that the robust position represents the typical stable operating center of each process quantity under the current operating condition and is not sensitive to a few abnormal fluctuations. Thus, during the water-locking coke shell bursting injection disturbance, the reference value can be avoided from drifting with the transient peak and the optimization target can be guaranteed to be achievable and stable.

[0036] After constructing the reference trajectory, a one-step prediction model is established based on the local sensitivity matrix. This model is used to predict the trend of the output vector in the next sampling period after a small change in the execution quantity under the current operating conditions, and is used for subsequent cost function calculation and control increment solution. Its form is as follows: In the formula, This represents the output vector for the current sampling period. The output vector is composed of load representative quantity, bed temperature, flue gas carbon monoxide content, flue gas oxygen content, and bed pressure difference in a predetermined order. This represents the control increment vector to be solved in the current sampling cycle. The control increment vector is composed of the feed execution increment, the primary air execution increment, and the secondary air execution increment in a predetermined order. This represents the local sensitivity matrix, which signifies the marginal mapping relationship between small changes in the executed quantity and the changes in each component of the output vector under the current operating condition. This represents the output vector for the next sampling period, predicted based on the current output and control increment. The reason for adopting this model is that approximating the system response as a linear superposition relationship within a short prediction step can maintain computability and real-time performance even when the marginal effect changes due to the bursting and injection of water-locked coke. Furthermore, the local sensitivity matrix is ​​adaptively updated by the current window data, allowing the prediction relationship to be updated with changes in operating conditions, thus providing a prediction basis consistent with actual operation for multi-objective optimization.

[0037] Construct an optimal cost function; The optimal closed-loop control increment is obtained by minimizing the cost function. In detail, after obtaining the reference trajectory, multi-objective weight matrix, input penalty matrix, and one-step prediction model, an optimization cost function is constructed for solving the multi-objective optimization problem. This optimization cost function consists of a weighted square term representing the deviation of the predicted output from the reference trajectory and a weighted square term representing the control increment. This simultaneously tracks and stabilizes representative load parameters such as bed temperature, flue gas carbon monoxide content, flue gas oxygen content, and bed pressure difference, while suppressing excessively frequent adjustments to the primary and secondary air feeds. The optimization cost function is as follows: In the formula, This represents the numerical value of the cost function with the control increment vector as the independent variable. This represents the control increment vector to be solved. This represents the predicted output vector for the next sampling period obtained from the one-step prediction model. Represents the reference trajectory vector. Represents a multi-objective weight matrix. The input penalty matrix is ​​represented by the first term of the above equation, which measures the deviation between the predicted output and the reference trajectory and assigns the penalty intensity according to the multi-objective weight matrix, so that the bed temperature burnout and bed stability related objectives are subject to higher constraints when the injection risk increases. The second term measures the magnitude of the control increment and assigns the penalty intensity according to the input penalty matrix to limit the magnitude and frequency of the execution quantity change, thereby reducing the secondary disturbances introduced to the bed flow state and combustion process.

[0038] To obtain the closed-form optimal control increment, the one-step prediction model is substituted into the aforementioned cost function, and the derivative of the cost function with respect to the control increment vector is taken and set to zero to achieve minimization. The closed-form solution for the optimal control increment is then obtained as follows: In the formula, The optimal control increment vector is used to minimize the cost function. This is the local sensitivity matrix. This is the current output vector. Represents a multi-objective weight matrix. Represents the reference trajectory vector. The input penalty matrix is ​​represented in the above formula; the matrix is... It combines the marginal effect of the execution quantity on the output and the penalty intensity for changes in the input under the current operating conditions. Taking the inverse and multiplying it by the deviation term yields the control correction quantity that should be applied under the current trade-off conditions. The reason for using a closed-form solution is that it can achieve real-time solution under the constraints of industrial computing resources and avoid the time delay and convergence uncertainty caused by iterative optimization, thereby meeting the rapid response requirements of BFB biomass boiler injection disturbance conditions.

[0039] The reciprocal of the first coefficient is used as the feed attenuation coefficient; The feed component in the optimal control increment is attenuated and corrected according to the feed attenuation coefficient to obtain the final feed execution amount; The primary air execution volume, secondary air execution volume, and final feeding execution volume in the optimal control increment are sent to the actuator; In detail, after obtaining the optimal control increment in the closed loop, to suppress further abrupt changes in heat release caused by excessive feed adjustments when the risk of water-locked coke shell bursting and injection increases, the reciprocal of the first coefficient is defined as the feed attenuation coefficient. The first coefficient is composed of the sum of the continuous risk weight and a value of 1, so that the first coefficient increases and the feed attenuation coefficient decreases as the continuous risk weight increases, thereby automatically reducing the feed adjustment amplitude during the enhanced injection disturbance stage. Subsequently, the feed component is extracted from the optimal control increment vector and multiplied by the feed attenuation coefficient to obtain the attenuated feed increment. Then, the current feed execution amount is added to the attenuated feed increment to obtain the final feed execution amount. Simultaneously, the primary air component and the secondary air component are extracted from the optimal control increment vector and multiplied by the current primary air execution amount and the secondary air component, respectively. The secondary air execution quantities are added together to obtain the primary air execution quantity and the secondary air execution quantity to be issued. To ensure execution feasibility and equipment safety, after the final feeding execution quantity and the primary and secondary air execution quantities to be issued are formed, they are constrained within the rated boundary range of the corresponding actuators. The rated boundary is derived from the allowable output range of the feeding device and the fan frequency converter, and preferably based on the read equipment parameter table to avoid manually setting thresholds. After the boundary constraints are completed, the final feeding execution quantity and the primary and secondary air execution quantities to be issued are written into the output register and sent to the feeding actuator, the primary air actuator, and the secondary air actuator via the fieldbus. This allows the actuators to adjust according to the new execution quantities and enter the next round of acquisition and optimization calculation in the next sampling cycle, thereby forming a closed-loop multi-objective optimization control.

[0040] like Figure 2 The diagram shown is a functional block diagram of a multi-objective optimization control system for a BFB biomass boiler provided in an embodiment of the present invention.

[0041] In this embodiment, the functions of each module / unit are as follows: The data acquisition module is used to periodically collect the process quantities of the boiler and the execution quantities of the actuators. The robust residual module is used to calculate the robust location and robust scale of process quantities and execution quantities, and to generate standardized residuals of process quantities; The risk weight module is used to calculate the injection disturbance based on the covariance strength of the standardized residuals of the process quantity, and to map the injection disturbance to continuous risk weights. The sensitivity matrix module is used to construct the output vector and the execution vector, and to determine the local sensitivity matrix based on the output vector and the execution vector. The weight matrix module is used to construct multi-objective weight matrices and input penalty matrices based on robust scales of process and execution quantities and continuous risk weights. The instruction generation module is used to construct an optimized cost function based on the multi-objective weight matrix, the input penalty matrix, and the local sensitivity matrix, solve the cost function to obtain the control increment, and generate execution instructions.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for multi-objective optimization control of a BFB biomass boiler, characterized in that, The method comprises the following steps: S1, collecting process quantities and execution quantities of a boiler; S2, calculating robust positions and robust scales of the process quantities and the execution quantities, and generating standardized residuals of the process quantities; S3, calculating a jet disturbance quantity according to covariance intensity of the standardized residuals of the process quantities, mapping the jet disturbance quantity to a continuous risk weight; S4, constructing an output vector and an execution quantity vector, and determining a local sensitivity matrix based on the output vector and the execution quantity vector; S5, constructing a multi-objective weight matrix and an input penalty matrix based on the robust scales of the process quantities and the execution quantities and the continuous risk weight; S6, constructing an optimization cost function based on the multi-objective weight matrix, the input penalty matrix, and the local sensitivity matrix, solving the cost function to obtain a control increment, and generating an execution instruction.

2. The multi-objective optimization control method of a BFB biomass boiler according to claim 1, characterized in that, The process quantities include flue gas oxygen content, flue gas carbon monoxide content, bed pressure difference, soot signal, load representative quantity, and bed temperature; and the execution quantities include feeding execution quantity, primary air execution quantity, and secondary air execution quantity.

3. The BFB biomass boiler multi-objective optimization control method according to claim 2, characterized in that, The method for calculating the robust positions and the robust scales of the process quantities and the execution quantities, and generating the standardized residuals of the process quantities comprises the following steps: obtaining a sliding sample set, calculating a median as a robust position and calculating an absolute deviation median as a robust scale for each process quantity and execution quantity on the sliding sample set; taking a median of all the robust scales as a global self-provision scale; performing standardization processing on the process quantities based on the robust positions, the robust scales, and the global self-provision scale to obtain standardized residuals; wherein an inverse of the standardized residual of the flue gas oxygen content is taken as a jet positive residual.

4. The BFB biomass boiler multi-objective optimization control method according to claim 3, characterized in that, The method for calculating a jet disturbance quantity according to the covariance intensity of the standardized residuals of the process quantities comprises the following steps: calculating covariance absolute values of the jet positive residual of the flue gas oxygen content, the standardized residual of the flue gas carbon monoxide content, the standardized residual of the bed pressure difference, the standardized residual of the soot signal, and the standardized residual of the load representative quantity; performing normalization processing on the covariance absolute values to obtain normalized intensity coefficients of the process quantities; performing weighted summation on absolute values of the standardized residuals of the process quantities according to the normalized intensity coefficients to obtain the jet disturbance quantity.

5. The BFB biomass boiler multi-objective optimization control method according to claim 4, characterized in that, The method for determining a local sensitivity matrix based on the output vector and the execution quantity vector comprises the following steps: constructing an output vector comprising the load representative quantity, the bed temperature, the flue gas carbon monoxide content, the flue gas oxygen content, and the bed pressure difference; constructing an execution quantity vector comprising the feeding execution quantity, the primary air execution quantity, and the secondary air execution quantity; calculating an increment matrix of the output vector and an increment matrix of the execution quantity vector in a current window; calculating a sum of a vectorized median of the input increment matrix and a global self-provision scale square as a regularization term; calculating the local sensitivity matrix by a least square method based on the increment matrix of the execution quantity vector, the increment matrix of the output vector, and the regularization term.

6. The BFB biomass boiler multi-objective optimization control method according to claim 5, characterized in that, The method for constructing a multi-objective weight matrix comprises the following steps: calculating a benchmark weight factor based on the robust scales of the process quantities and the global self-provision scale; taking a sum of the continuous risk weight and a numerical value one as a first coefficient; The multi-objective weight matrix in the form of a diagonal matrix is constructed, wherein the diagonal element corresponding to the load representative quantity is the ratio of the reference weight factor to the first coefficient, and the diagonal element corresponding to the bed temperature, the flue gas carbon monoxide content, the flue gas oxygen content and the bed pressure difference is the product of the reference weight factor and the first coefficient.

7. The BFB biomass boiler multi-objective optimization control method according to claim 6, characterized in that, The input penalty matrix is constructed, including: The input variation penalty factor is calculated based on the robust scale and the global self-sufficient scale of each execution quantity; The input penalty matrix in the form of a diagonal matrix is constructed, wherein the diagonal element of the input penalty matrix is composed of the input variation penalty factor.

8. The BFB biomass boiler multi-objective optimization control method according to claim 7, characterized in that, The optimization cost function is constructed based on the multi-objective weight matrix, the input penalty matrix and the local sensitivity matrix, the control increment is obtained by solving the cost function, and the execution instruction is generated, including: The reference trajectory of the process quantity is constructed, wherein the reference value of the load representative quantity is the external load instruction, and the reference value of the bed temperature, the flue gas carbon monoxide content, the flue gas oxygen content and the bed pressure difference is the corresponding robust position; The one-step prediction model is established based on the local sensitivity matrix, and the one-step prediction model is: wherein is the predicted output vector of the next time instant, is the output vector of the current time instant, is the control increment to be determined, is the local sensitivity matrix; The optimization cost function is constructed, and the optimization cost function is: wherein is the cost function to be optimized, is the reference trajectory vector, is the multi-objective weight matrix, is the input penalty matrix, is the control increment to be found, is the predicted output vector at the next time instant; The optimal control increment in closed form is obtained by minimizing the optimization cost function; The reciprocal of the first coefficient is taken as the feed attenuation coefficient; The feed component in the optimal control increment is attenuated and corrected according to the feed attenuation coefficient to obtain the final feed execution quantity; The primary air execution quantity, the secondary air execution quantity and the final feed execution quantity in the optimal control increment are issued to the execution mechanism.

9. A BFB biomass boiler multi-objective optimization control system, characterized by, The system includes: The data acquisition module is used for periodically acquiring the process quantity of the boiler and the execution quantity of the execution mechanism; The robust residual module is used for calculating the robust position and the robust scale of the process quantity and the execution quantity, and generating the standardized residual of the process quantity and the standardized residual of the execution quantity; The risk weight module is used for calculating the injection disturbance quantity according to the covariance intensity of the standardized residual of the process quantity, and mapping the injection disturbance quantity into continuous risk weight; The sensitivity matrix module is used for constructing the output vector and the execution quantity vector, and determining the local sensitivity matrix based on the output vector and the execution quantity vector; The weight matrix module is used for constructing the multi-objective weight matrix and the input penalty matrix based on the robust scale of the process quantity and the execution quantity and the continuous risk weight; The instruction generation module is used for constructing the optimization cost function based on the multi-objective weight matrix, the input penalty matrix and the local sensitivity matrix, obtaining the control increment by solving the cost function, and generating the execution instruction.

Images