A machine learning-based desulfurization system parameter optimization control method

Abstract

This invention discloses a machine learning-based parameter optimization and control method for desulfurization systems, relating to the field of desulfurization systems. The method includes: collecting multi-dimensional real-time parameters of the entire desulfurization system operation process; simultaneously collecting desulfurization efficiency and system energy consumption data under corresponding operating conditions; and constructing a time-labeled full-condition parameter-efficiency correlation dataset. Based on this dataset, a multi-objective coupled machine learning prediction model is constructed, with desulfurization efficiency reaching the target and system energy consumption being minimized as dual constraints, outputting the efficiency response range corresponding to parameter changes. This invention, by constructing a time-labeled full-condition parameter-efficiency correlation dataset, achieves the integration and correction of multi-dimensional parameters of the desulfurization system, laying a reliable data foundation for model construction. Furthermore, by combining a hybrid model of deep residual networks and attention mechanisms, it can accurately predict the efficiency response range of parameter changes, improving the accuracy of parameter-efficiency correlation analysis.

Description

Technical Field

[0001] This invention relates to the field of desulfurization system technology, specifically to a desulfurization system parameter optimization and control method based on machine learning. Background Technology

[0002] Desulfurization systems are core devices for industrial flue gas purification, primarily removing sulfur dioxide from flue gas from industries such as thermal power and steel. The mainstream approach uses the limestone-gypsum wet process, where limestone slurry acts as an absorbent, reacting with sulfur dioxide to produce gypsum. This efficiently removes sulfur pollutants, reduces acid rain damage, and achieves compliant flue gas emissions while also enabling the resource utilization of byproducts.

[0003] Patent application No. 202410638877.7 discloses a machine learning-based parameter optimization and control method for flue gas desulfurization systems. This application aims to address the following problems: "Different algorithms have varying effects on parameter optimization and control, requiring selection and adjustment based on specific circumstances; due to potential changes in flue gas desulfurization system parameters and environmental conditions, the collected data may only reflect performance under specific conditions, posing a challenge to the model's generalization ability. In practical applications, the model may not adapt well to new parameter combinations or environmental conditions, leading to inaccurate or unreliable predictions; during parameter optimization, optimization algorithms are often used to search the parameter space to find the optimal parameter combination for predicting desulfurization effects. However, commonly used optimization algorithms often get trapped in local optima, finding the optimal solution within the current search region, but not necessarily the global optimum. This may result in unsatisfactory optimization results in practical applications, requiring further optimization or adjustment of algorithm parameter settings."

[0004] However, during the parameter control of the entire desulfurization system, it is often difficult to achieve the dual goals of achieving desulfurization efficiency and minimizing system energy consumption under all operating conditions. Furthermore, the parameter optimization efficiency is low and the parameter adaptation stability under boundary conditions is poor.

[0005] To address this, we propose a machine learning-based method for optimizing and controlling the parameters of a desulfurization system. Summary of the Invention

[0006] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a desulfurization system parameter optimization control method based on machine learning, which can effectively solve the problems of the existing technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions; This invention discloses a machine learning-based parameter optimization control method for desulfurization systems, comprising: collecting multi-dimensional real-time parameters of the entire desulfurization system operation process, simultaneously collecting desulfurization efficiency and system energy consumption data under corresponding operating conditions, and constructing a time-series labeled parameter-efficiency correlation dataset for the entire operating condition; based on the parameter-efficiency correlation dataset for the entire operating condition, constructing a multi-objective coupled machine learning prediction model, with desulfurization efficiency reaching the target and system energy consumption being minimized as dual constraints, and outputting the efficiency response interval corresponding to parameter changes; setting the feasible region boundary of the desulfurization system's entire operating condition parameters, and generating several sets of non-parameter parameters within the feasible region. The initial parameter optimization nodes are continuously distributed, and the independent optimization intervals corresponding to each node are defined synchronously. The machine learning prediction model is applied to perform parallel iterative optimization of the parameter combinations in each independent optimization interval, and the optimal parameter combinations and corresponding performance indicators obtained by each interval optimization are recorded synchronously. The optimal parameter combinations output by all independent optimization intervals are subjected to full performance cross-comparison to select the global optimal parameter set, and the adaptability and stability of the parameter set under boundary conditions are verified synchronously. The verified global optimal parameter set is distributed to each execution unit of the desulfurization system and executed synchronously based on the dynamic periodic refresh step.

[0008] Furthermore, the multi-dimensional real-time parameters include inlet flue gas parameters, absorbent supply parameters, in-tower flow field parameters, circulating slurry parameters, and outlet flue gas treatment parameters of the desulfurization absorption tower. Each dimension parameter carries a millisecond-level timestamp to form a time-series label. During the construction of the full-condition parameter-efficiency association dataset, missing parameter values ​​are filled in using time-series neighborhood parameter interpolation, and abnormal parameter values ​​are corrected through condition similarity matching. ; In the formula: Cosine similarity of working condition features; The feature vector of the operating condition where the abnormal parameters are located; The feature vector represents the historical normal operating conditions. These are the corrected abnormal parameter values; M is the preset number of historical normal operating conditions with high similarity. The cosine similarity values ​​of the top M highest working condition features; The parameter value in the same dimension as the abnormal parameter in the m-th high-similarity historical normal operating condition; The dataset is divided into training set, validation set and test set according to a preset working condition division ratio, and the working condition distribution characteristics of each set are consistent.

[0009] Furthermore, the multi-objective coupled machine learning prediction model is a hybrid model combining a deep residual network and an attention mechanism. The loss function of the model is the coupling loss between the desulfurization efficiency deviation loss and the system energy consumption loss, and the calculation of the coupling loss satisfies: ; In the formula: This represents the model coupling loss value. To predict desulfurization efficiency for the model; This represents the desulfurization efficiency target value. To predict system energy consumption using a model; This represents the rated basic energy consumption of the desulfurization system. During model training, an adaptive learning rate strategy is adopted, which dynamically adjusts the learning rate based on the performance response interval prediction accuracy of the validation set until the model's prediction error falls into the preset interval. The , The calculation formula is: ; When dynamically adjusting the learning rate based on the performance response interval prediction accuracy of the validation set, the following applies: ; In the formula: The baseline desulfurization efficiency of the desulfurization system; This represents the total number of dimensions of the parameters for the entire desulfurization system under all operating conditions. Let be the attention weight for the i-th parameter; The desulfurization efficiency feature mapping function is extracted for the i-th parameter from the deep residual network. The activation coefficient is a parameter characteristic of desulfurization efficiency; This refers to the i-th full-condition parameter; The historical average value of the i-th parameter under operating conditions; This is the system energy consumption feature mapping function for the i-th parameter extracted by the deep residual network; The activation coefficient is a parameter characteristic of system energy consumption; Let be the model learning rate for the t-th iteration; The initial learning rate for the model; The preset maximum number of iterations for model training; Let be the overall prediction error of the model in the t-th iteration; Let be the overall prediction error of the model in the (t-1)th iteration; This is the preset minimum value correction coefficient.

[0010] Furthermore, the feasible domain boundary of the full-condition parameters of the desulfurization system is determined by the following formula: ; In the formula: This refers to the i-th full-condition parameter; represents the historical extreme value of the i-th parameter; k is the parameter feasible region correction coefficient; The , This represents the joint standard deviation of desulfurization efficiency versus system energy consumption. The desulfurization efficiency is the average efficiency under historical operating conditions. The average system energy consumption under historical operating conditions; Once the feasible domain boundary is determined, invalid domains corresponding to parameter combinations that have process conflicts are eliminated, and only valid feasible domains that satisfy the desulfurization process interlocking logic are retained.

[0011] Furthermore, in the initial parameter optimization node generation stage, a non-continuously distributed set of nodes is generated within the effective feasible region, and the independent optimization interval corresponding to each node is adaptively divided according to the node density. The interval partitioning logic is as follows: The Euclidean distance between the parameters of any two adjacent optimization nodes is greater than a preset threshold. The parameter combinations within each independent optimization interval satisfy the process coupling constraints of the desulfurization system. The number of independent optimization intervals is positively correlated with the parameter dimensions of the desulfurization system; The parameter search ranges of each interval do not overlap.

[0012] Furthermore, in the parallel iterative optimization phase, each independent optimization interval is configured with an independent agent. The performance evaluation of the parameter combinations within the interval by the agent is completed by a machine learning prediction model. The termination condition for iterative optimization is that within a preset number of consecutive iterations, the rate of change of the performance index is less than a preset threshold, and the parameter update amount in each iteration satisfies the following: ; In the formula: This represents the update amount of the i-th parameter in the (n+1)-th iteration; This represents the update amount of the i-th parameter in the n-th iteration; The predicted desulfurization efficiencies for the (n+1)th and nth iterations are respectively. This represents the desulfurization efficiency target value. These are the predicted system energy consumption for the (n+1)th and nth iterations, respectively. This represents the minimum energy consumption within the current optimization interval. In this process, each agent interacts with optimization information in real time during the iteration process, eliminating invalid parameter combinations that have already been searched, in order to avoid redundant calculations in global optimization.

[0013] Furthermore, the full-scale performance cross-comparison uses a comprehensive performance evaluation index to quantitatively evaluate each optimal parameter combination, wherein the comprehensive performance evaluation index is calculated using the following formula: ; In the formula: As a comprehensive performance evaluation index; Predicted desulfurization efficiency for parameter combinations; This represents the desulfurization efficiency target value. This represents the theoretical minimum system energy consumption across the entire feasible region. Predict the energy consumption of the system based on the parameter combination; The boundary condition adaptability coefficients for the parameter combination. ∈(0,1]; The preset corrected standard deviation of the comprehensive performance evaluation index; The full-scale performance cross-comparison process is as follows: First, calculate the comprehensive performance evaluation index of the optimal parameter combination output for each independent optimization interval. Directly remove For parameter combinations below a preset threshold, a full-condition, full-coverage simulation verification covering both conventional and boundary conditions of the desulfurization system is conducted for the remaining parameter combinations. During the simulation verification process, the real-time comprehensive performance evaluation index of each parameter combination under each type of condition is calculated. Then, the average real-time comprehensive performance evaluation index of a single parameter combination under all conditions is calculated, and this average value is used as the final comprehensive performance evaluation index of that parameter combination. One or more parameter combinations with the highest final comprehensive performance evaluation index are selected as candidate sets for the globally optimal parameter set.

[0014] Furthermore, the adaptation stability verification under the boundary conditions includes verification of all types of boundary conditions, such as extreme flue gas volume conditions, extreme sulfur dioxide concentration conditions, absorbent quality deterioration conditions, and system low load conditions. During the verification process, the performance fluctuation coefficient of the global optimal parameter set under each boundary condition is calculated. When the performance fluctuation coefficient under all boundary conditions is less than the preset threshold, the parameter set is determined to be stable. If the performance fluctuation coefficient of any boundary condition exceeds the preset threshold, the boundary condition is included in the optimization condition set and parameter optimization is re-executed. When the global optimal parameter set is sent to each execution unit of the desulfurization system, a hierarchical parameter sending strategy is adopted. Parameter instructions are sent sequentially according to the process response priority of the execution unit. After each execution unit receives the parameter instruction, it sends back a parameter execution confirmation signal. After all execution units have sent back confirmation signals, the execution parameters of the desulfurization system are switched.

[0015] Furthermore, the duration of the dynamic periodic refresh is dynamically adjusted according to the rate of change of the desulfurization system's operating conditions: ; In the formula: C represents the degree of change in the real-time operating conditions of the desulfurization system; For L2 norm operation, j is the dimension number of the core operating parameters of the desulfurization system; The difference between the real-time value and the steady-state value of the j-th core operating condition parameter; Let be the steady-state value of the j-th core operating condition parameter; T is the dynamic refresh cycle. This serves as the baseline refresh cycle for the desulfurization system. The degree of variation of the desulfurization system's baseline operating conditions; The operating condition sensitivity coefficient is adapted to the desulfurization system process.

[0016] Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects: This invention integrates and corrects multi-dimensional parameters of the desulfurization system by constructing a time-labeled full-condition parameter-efficiency correlation dataset, laying a reliable data foundation for model construction. Furthermore, by combining a hybrid model of deep residual networks and attention mechanisms, it can accurately predict the efficiency response range of parameter changes, improving the accuracy of parameter-efficiency correlation analysis. Simultaneously, by dividing independent optimization intervals and conducting parallel iterative optimization, the efficiency of parameter optimization is effectively improved, avoiding redundant calculations. After full-scale efficiency cross-comparison and boundary condition adaptation stability verification, the globally optimal parameter set selected balances desulfurization efficiency targets with the lowest system energy consumption, exhibiting stronger adaptability and stability. In addition, the dynamic periodic refresh mechanism can adjust the refresh duration according to the rate of change in operating conditions, ensuring that parameter control always aligns with the actual operating state of the system, making the desulfurization system more intelligent and precise overall. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0018] Figure 1 This is a flowchart illustrating a desulfurization system parameter optimization control method based on machine learning. Figure 2 This is a schematic diagram illustrating the basic structure of the desulfurization system in this invention; Figure 3 This is a flowchart illustrating the method in this invention. The diagram represents: 1. Flue gas duct; 2. Absorption tower; 3. Slurry circulation pump; 4. Gypsum dewatering system; 5. Oxidation fan. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] The present invention will be further described below with reference to embodiments.

[0021] Example: This embodiment presents a desulfurization system parameter optimization control method based on machine learning, such as... Figure 1 As shown, it includes: Collect multi-dimensional real-time parameters of the entire desulfurization system operation process, and simultaneously collect desulfurization efficiency and system energy consumption data under corresponding operating conditions to construct a full-condition parameter-efficiency correlation dataset with time-series labels. Multi-dimensional real-time parameters include inlet flue gas parameters, absorbent supply parameters, internal flow field parameters, circulating slurry parameters, and outlet flue gas treatment parameters of the desulfurization absorption tower. Each dimension parameter carries a millisecond-level timestamp to form a time-series label. During the construction of the full-condition parameter-efficiency association dataset, missing parameter values ​​are filled in using time-series neighborhood parameter interpolation, and abnormal parameter values ​​are corrected through operating condition similarity matching. ; In the formula: Cosine similarity of working condition features; The feature vector of the operating condition where the abnormal parameters are located; The feature vector represents the historical normal operating conditions. and All of them are composed of parameter characteristics of the desulfurization system under all operating conditions; These are the corrected abnormal parameter values; M is the preset number of historical normal operating conditions with high similarity. The cosine similarity values ​​of the top M highest working condition features; The parameter value in the same dimension as the abnormal parameter in the m-th high-similarity historical normal operating condition; The above formula uses cosine similarity to quantify the feature matching degree between abnormal operating conditions and historical normal operating conditions, which fits the operating condition feature correlation characteristics of multi-dimensional parameters of the desulfurization system. Then, the abnormal parameter values ​​are corrected by weighting the first M historical normal operating condition parameters with the highest similarity, so that the abnormal parameter correction results are more in line with the actual operating condition rules. The dataset is divided into training set, validation set and test set according to the preset working condition division ratio, and the working condition distribution characteristics of each set are consistent. Among them, the judgment condition for maintaining consistent operating condition distribution characteristics is: the deviation between the proportion of various desulfurization process operating conditions in each set and the corresponding operating condition proportion in the original full operating condition parameter-efficiency association dataset is less than the preset threshold, and the relative deviation between the mean and variance of the two key efficiency parameters, desulfurization efficiency and system energy consumption, in each set and the mean and variance of the corresponding parameters in the original dataset is less than the preset threshold. Based on the full-condition parameter-efficiency correlation dataset, a multi-objective coupled machine learning prediction model is constructed. With desulfurization efficiency meeting the standard and system energy consumption being minimized as the dual constraints, the output efficiency response range corresponding to parameter changes is determined. The multi-objective coupled machine learning prediction model is a hybrid model combining a deep residual network and an attention mechanism. The model's loss function is the coupling loss between the desulfurization efficiency deviation loss and the system energy consumption loss. The calculation of the coupling loss satisfies the following: ; In the formula: This represents the model coupling loss value. To predict desulfurization efficiency for the model; This represents the desulfurization efficiency target value. To predict system energy consumption using a model; This represents the rated basic energy consumption of the desulfurization system. This formula couples the desulfurization efficiency deviation with the system energy consumption loss. It quantifies the deviation of the predicted desulfurization efficiency value relative to the target value and the predicted energy consumption value relative to the rated basic energy consumption in logarithmic form. It positively constrains the desulfurization efficiency to approach the target value and negatively constrains the energy consumption to change to a lower value, accurately adapting to the dual-objective optimization requirements of achieving the desulfurization efficiency target and minimizing energy consumption. During model training, an adaptive learning rate strategy is adopted, which dynamically adjusts the learning rate based on the performance response interval prediction accuracy of the validation set until the model's prediction error falls into the preset interval. The model assigns differentiated attention weights to the input parameters of each dimension through an attention mechanism. After extracting the parameter features through a deep residual network, it outputs the predicted values ​​of desulfurization efficiency and system energy consumption. Based on the gradient variation range of the parameters under all operating conditions, combined with the predicted efficiency values ​​corresponding to each parameter, it outputs the efficiency response range of desulfurization efficiency and system energy consumption under different parameter variation ranges. , The calculation formula is: ; Based on the benchmark efficiency value of the desulfurization system, this formula incorporates differentiated parameter attention weights and feature mapping functions extracted by deep residual networks. It then uses an S-shaped activation term to characterize the nonlinear effects of parameters deviating from historical averages. At the same time, it sets exclusive activation coefficients for desulfurization efficiency and energy consumption, accurately reflecting the differentiated effects of different parameters on the two types of efficiency indicators, thereby achieving accurate prediction of efficiency indicators by multi-dimensional parameters. When dynamically adjusting the learning rate based on the performance response interval prediction accuracy of the validation set, the following applies: ; The above formula dynamically adjusts the learning rate based on the exponential decay of the initial learning rate and the combined prediction error change rate of adjacent iterations. It introduces a minimum correction coefficient to avoid the case where the denominator is zero in the calculation, so that the learning rate can be adaptively adjusted according to the prediction accuracy of the model validation set. This ensures the convergence speed of model training and improves the accuracy of model prediction. In the formula: The baseline desulfurization efficiency of the desulfurization system; This represents the total number of dimensions of the parameters for the entire desulfurization system under all operating conditions. Let be the attention weight for the i-th parameter; The desulfurization efficiency feature mapping function is extracted for the i-th parameter from the deep residual network. The activation coefficient is a parameter characteristic of desulfurization efficiency; This refers to the i-th full-condition parameter; The historical average value of the i-th parameter under operating conditions; This is the system energy consumption feature mapping function for the i-th parameter extracted by the deep residual network; The activation coefficient is a parameter characteristic of system energy consumption; Let be the model learning rate for the t-th iteration; The initial learning rate for the model; The preset maximum number of iterations for model training; Let be the overall prediction error of the model in the t-th iteration; Let be the overall prediction error of the model in the (t-1)th iteration; This is a preset minimum value correction coefficient; in, ∈ (0,1), and the sum of all weights is 1. The higher the comprehensive influence of the parameter on desulfurization efficiency and system energy consumption, the larger the value; conversely, the smaller the value. ∈(0,+∞), the larger the value is when a small change in the parameter has a more significant nonlinear effect on the desulfurization efficiency, and the smaller the value is when the change is less significant. ∈(0,+∞), the larger the value is when the small change of the parameter has a more significant nonlinear effect on the system's energy consumption, and the smaller the value is when the small change of the parameter has a less significant effect on the system's energy consumption. , All of these are comprehensive prediction error values ​​calculated by combining the deviations between the predicted and actual values ​​of desulfurization efficiency and the deviations between the predicted and actual values ​​of system energy consumption for each sample in the validation set with the weights of model coupling loss. , Based on the full-condition parameter-efficiency correlation dataset, the deep residual network is obtained by multi-layer iterative learning and nonlinear fitting of the correlation features between parameters and desulfurization efficiency and parameters and system energy consumption. The two complete independent feature mapping according to their respective parameter-efficiency correlation rules. Set the feasible domain boundary of the full-condition parameters of the desulfurization system, generate several sets of non-continuously distributed initial parameter optimization nodes within the feasible domain, and simultaneously define the independent optimization intervals corresponding to each node. The feasible domain boundary of the desulfurization system's full-condition parameters is determined by the following formula: ; In the formula: This refers to the i-th full-condition parameter; represents the historical extreme value of the i-th parameter; k is the parameter feasible region correction coefficient; , This represents the joint standard deviation of desulfurization efficiency versus system energy consumption. The desulfurization efficiency is the average efficiency under historical operating conditions. The average system energy consumption under historical operating conditions; The above formula determines the basic range of the feasible region based on the extreme values ​​of the parameters under historical operating conditions. It combines the joint standard deviation of desulfurization efficiency and system energy consumption with the historical average efficiency value to design a correction coefficient, so that the range of the feasible region can adapt to the overall fluctuation characteristics of the system efficiency. At the same time, it eliminates invalid regions with process conflicts, so that the parameter feasible region not only conforms to the historical operating rules, but also conforms to the process interlocking logic of the desulfurization system. After the feasible region boundary is determined, invalid regions corresponding to parameter combinations that have process conflicts are eliminated, and only valid feasible regions that satisfy the desulfurization process interlocking logic are retained. The machine learning prediction model is applied to perform parallel iterative optimization of parameter combinations in each independent optimization interval, and the optimal parameter combination and corresponding performance index obtained in each interval are recorded simultaneously. In the initial parameter optimization node generation stage, a non-continuously distributed set of nodes is generated within the effective feasible region, and the independent optimization interval corresponding to each node is adaptively divided according to the node density. The interval partitioning logic is as follows: The Euclidean distance between the parameters of any two adjacent optimization nodes is greater than a preset threshold. The parameter combinations within each independent optimization interval satisfy the process coupling constraints of the desulfurization system. The number of independent optimization intervals is positively correlated with the parameter dimensions of the desulfurization system; The parameter search ranges of each interval do not overlap; In the parallel iterative optimization phase, an independent agent is configured for each independent optimization interval. The performance evaluation of the parameter combinations within the interval by the agent is completed by a machine learning prediction model. The termination condition for iterative optimization is that the rate of change of the performance index is less than a preset threshold within a preset number of consecutive iterations, and the parameter update amount in each iteration satisfies the following: ; In the formula: This represents the update amount of the i-th parameter in the (n+1)-th iteration; This represents the update amount of the i-th parameter in the n-th iteration; The predicted desulfurization efficiencies for the (n+1)th and nth iterations are respectively. This represents the desulfurization efficiency target value. These are the predicted system energy consumption for the (n+1)th and nth iterations, respectively. This represents the minimum energy consumption within the current optimization interval. This formula is based on the previous update amount of the parameters in the iterative process. It adjusts the update range by combining the ratio of the deviation of desulfurization efficiency from the target value and the deviation of energy consumption from the minimum value of the range in adjacent iterations. This allows the parameter update amount to change dynamically with the desulfurization efficiency target value and the energy consumption optimization degree, making the parameter adjustment in the optimization process more targeted, thereby improving the efficiency and accuracy of iterative optimization. In this process, each agent interacts with optimization information in real time during the iteration process, eliminating invalid parameter combinations that have been searched, so as to avoid redundant calculations in global optimization. A full performance cross-comparison is performed on the optimal parameter combinations output from all independent optimization intervals to select the globally optimal parameter set, and the adaptability and stability of the parameter set under boundary conditions are verified simultaneously. The full-scale performance cross-comparison uses a comprehensive performance evaluation index to quantitatively assess each optimal parameter combination. The formula for calculating the comprehensive performance evaluation index is as follows: ; In the formula: This is a comprehensive performance evaluation index; the larger the value, the better the overall performance of the parameter combination. Predicted desulfurization efficiency for parameter combinations; This represents the desulfurization efficiency target value. This represents the theoretical minimum system energy consumption across the entire feasible region. Predict the energy consumption of the system based on the parameter combination; The boundary condition adaptability coefficient of the parameter combination is taken as the average efficiency compliance rate of the parameter combination under each preset boundary condition, and... ∈(0,1]; The preset correction standard deviation of the comprehensive performance evaluation index is used to adjust the influence weight of boundary condition adaptability according to the system performance fluctuation characteristics. The above formula integrates the desulfurization efficiency compliance ratio and the system energy consumption optimization ratio, which intuitively reflects the core performance of the parameter combination. At the same time, it introduces the correction standard deviation power term of the boundary condition adaptability coefficient and takes into account the adaptability stability of the parameter combination under boundary conditions. Through a single index, it achieves a comprehensive quantification of the performance of the parameter combination under normal operating conditions and the adaptability under boundary conditions, so as to facilitate the rapid selection of high-quality parameter combinations. The full performance cross-comparison process is as follows: First, calculate the comprehensive performance evaluation index of the optimal parameter combination output for each independent optimization interval. Directly remove For parameter combinations below a preset threshold, a full-condition, full-coverage simulation verification covering both conventional and boundary conditions of the desulfurization system is conducted for the remaining parameter combinations. During the simulation verification process, the real-time comprehensive performance evaluation index of each parameter combination under each type of condition is calculated. Then, the average real-time comprehensive performance evaluation index of a single parameter combination under all conditions is calculated, and this average is used as the final comprehensive performance evaluation index of that parameter combination. One or more parameter combinations with the highest final comprehensive performance evaluation index are selected as candidate sets of the globally optimal parameter set. The adaptation stability verification under boundary conditions includes verification of all types of boundary conditions, such as extreme flue gas volume conditions, extreme sulfur dioxide concentration conditions, absorbent quality deterioration conditions, and system low load conditions. During the verification process, the performance fluctuation coefficient of the global optimal parameter set under each boundary condition is calculated. When the performance fluctuation coefficient under all boundary conditions is less than the preset threshold, the parameter set is determined to be stable. If the performance fluctuation coefficient of any boundary condition exceeds the preset threshold, the boundary condition is included in the optimization condition set and parameter optimization is re-executed. When the global optimal parameter set is sent to each execution unit of the desulfurization system, a hierarchical parameter sending strategy is adopted. The parameter instructions are sent sequentially according to the process response priority of the execution unit. After each execution unit receives the parameter instruction, it sends back a parameter execution confirmation signal. After all execution units send back confirmation signals, the execution parameters of the desulfurization system are switched. During parameter switching, a gradual parameter adjustment method is adopted to avoid operating condition oscillations in the desulfurization system due to sudden parameter changes. The verified global optimal parameter set is distributed to each execution unit of the desulfurization system and executed synchronously based on the dynamic periodic refresh step. The cycle duration of the dynamic refresh is dynamically adjusted according to the rate of change of the desulfurization system's operating conditions. ; In the formula: C represents the degree of change in the real-time operating conditions of the desulfurization system; For L2 norm operation, j is the dimension number of the core operating parameters of the desulfurization system; The difference between the real-time value and the steady-state value of the j-th core operating condition parameter; Let be the steady-state value of the j-th core operating condition parameter; T is the dynamic refresh cycle. This serves as the baseline refresh cycle for the desulfurization system. The degree of variation of the desulfurization system's baseline operating conditions; The operating condition sensitivity coefficient adapted to the desulfurization system process; This cycle formula calculates the overall degree of change of the core operating parameters of the desulfurization system relative to the steady-state value using the L2 norm, accurately quantifies the degree of fluctuation of the system's real-time operating conditions, and then adjusts the refresh cycle based on the benchmark refresh cycle, combined with the ratio of the degree of change of operating conditions and the operating condition sensitivity coefficient of process adaptation, so that the refresh frequency of parameter optimization matches the system operating condition fluctuation rate. When the operating condition fluctuation is large, the refresh is accelerated, and when the fluctuation is small, the refresh frequency is reduced to ensure the dynamic adaptation and optimization of parameters. in, The value is greater than 0, and the more sensitive the desulfurization system process is to changes in operating conditions, the greater the value is when fluctuations in operating conditions easily cause large fluctuations in desulfurization efficiency. The stronger the desulfurization system process's ability to resist fluctuations in operating conditions, the smaller the value is when changes in operating conditions have a smaller impact on desulfurization efficiency.

[0022] In this embodiment, the above method can accurately correlate the operating parameters and operational efficiency of the desulfurization system, minimize system energy consumption while ensuring that the desulfurization efficiency meets the standards, and select the optimal parameters that are suitable for all operating conditions through scientific optimization. It can also maintain stable operation under various boundary conditions, and the parameters can be dynamically refreshed as the operating conditions change, which greatly improves the intelligence level of the desulfurization system operation and realizes efficient, stable and low-consumption operation of the entire process, thereby effectively optimizing resource utilization efficiency.

[0023] See Figure 2 As shown, the desulfurization system mainly consists of flue gas ducts, an absorption tower, a slurry circulation pump, a gypsum dewatering system, and an oxidation blower. Sulfur-containing flue gas enters the absorption tower through the flue gas ducts, where it comes into counter-current contact with the circulating desulfurization slurry, ensuring the full absorption of sulfur dioxide in the flue gas. The slurry circulation pump provides power for the slurry circulation within the absorption tower, ensuring efficient gas-liquid mass transfer. The oxidation blower blows air into the tower, oxidizing sulfites to sulfates, generating gypsum slurry. The gypsum slurry is treated by the gypsum dewatering system to obtain gypsum byproducts that can be utilized as resources, while the purified, compliant flue gas is discharged from the top of the tower.

[0024] Application example: The method in Example 1 was applied to the limestone-gypsum wet desulfurization system of a 300MW unit in a coal-fired power plant to achieve parameter optimization and intelligent control of the desulfurization system. The specific application process and results are as follows: In this desulfurization system, real-time parameters from multiple dimensions are first collected, including inlet flue gas parameters, absorbent supply parameters, internal flow field parameters, circulating slurry parameters, and outlet flue gas treatment parameters. Millisecond-level timestamps are added to all parameters. Simultaneously, desulfurization efficiency and system energy consumption data under corresponding operating conditions are collected to construct a time-labeled full-condition parameter-efficiency correlation dataset. For missing parameter values ​​in the dataset, time-series neighborhood parameter interpolation is used to complete them. For abnormal parameter values, cosine similarity matching of operating condition features is used for correction. The corrected value for abnormal parameters is the weighted average of the same dimension parameters from the top 10 highly similar historical normal operating conditions. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. Verification shows that the deviation of the proportion of various desulfurization process conditions in each set from the original dataset is less than 5%, and the relative deviation of the mean and variance of desulfurization efficiency and system energy consumption is less than 3%, meeting the requirement of consistent operating condition distribution characteristics.

[0025] Based on the constructed dataset, a multi-objective coupled machine learning prediction model combining a deep residual network and an attention mechanism was built. With 98% desulfurization efficiency and minimum system energy consumption as dual constraints, the model's final convergence loss value was calculated using a coupled loss function to be 0.02. The model training employed an adaptive learning rate strategy, which, after dynamic adjustment, stabilized at 0.001. The model's prediction error fell within a preset range of ±1%. Simultaneously, the model outputs the efficiency response range for different parameter variations. For example, when the inlet flue gas velocity varied by ±5%, the desulfurization efficiency response range was 98.2%-99.5%, and the system energy consumption response range was 120kW-135kW.

[0026] Subsequently, the feasible region boundary of the desulfurization system's full-condition parameters was determined. By calculating the historical extreme values ​​and the feasible region correction coefficient, the correction coefficient k was found to be 0.08. This was used to define the feasible region range of each parameter, and parameter combinations with process conflicts, such as conflicts between absorbent supply and circulating slurry flow, were eliminated, retaining only the effective feasible region that satisfies the desulfurization process's interlocking logic. Within the effective feasible region, 20 discontinuously distributed initial parameter optimization nodes were generated. The Euclidean distance between adjacent node parameters was greater than 0.5. Independent optimization intervals positively correlated with parameter dimensions were adaptively divided according to node density. All interval parameter search ranges had no overlap and all satisfied the desulfurization system's process coupling constraints.

[0027] A parallel iterative optimization method is employed using a machine learning prediction model. An independent agent is assigned to each independent optimization interval, and this agent evaluates the performance of parameter combinations within that interval. During the iteration process, the parameter updates gradually converge after calculation. The iteration terminates when the rate of change of the performance index is less than 0.01% for 10 consecutive iterations. The agents interact with each other in real time during the iteration, eliminating invalid parameter combinations. Ultimately, each independent optimization interval yields an optimal set of parameter combinations and corresponding performance indices.

[0028] A full-scale performance cross-comparison was conducted on the optimal parameter combinations output across all intervals. The index value of each combination was calculated using the comprehensive performance evaluation index formula. Parameter combinations with an index value lower than 1.1 were directly eliminated. Full-condition, full-coverage simulation verification was carried out on the remaining combinations under normal and boundary conditions. The mean real-time comprehensive performance evaluation index of single parameter combinations under all conditions was calculated. Finally, parameter combinations with a mean of 1.35 were selected as the candidate set of the global optimal parameter set.

[0029] The candidate set was subjected to boundary condition adaptation stability verification. Verification tests were completed for four types of boundary conditions: extreme flue gas volume, extreme sulfur dioxide concentration, absorbent quality deterioration, and system low load. The calculation showed that the efficiency fluctuation coefficient of the candidate set under each boundary condition was 0.03, which is less than the preset threshold of 0.05. Therefore, it was determined to be stable and was identified as the final global optimal parameter set.

[0030] The validated global optimal parameter set was distributed to each execution unit of the desulfurization system. A hierarchical parameter distribution strategy was adopted, issuing parameter commands sequentially according to the process response priority of the absorbent supply unit, circulating slurry unit, and flue gas conditioning unit. After all execution units provided confirmation signals, a gradual parameter adjustment method was used to complete the system parameter switching, without any oscillations in the desulfurization system's operating conditions. Simultaneously, parameter optimization and updates were performed based on a dynamic periodic refresh step. Calculations showed that when the desulfurization system operating conditions were stable, the real-time operating condition variation C was 0.04, and the dynamic refresh cycle was 30 minutes; when the system operating conditions fluctuated, the real-time operating condition variation C was 0.2, and the dynamic refresh cycle was shortened to 10 minutes, achieving dynamic intelligent optimization control of the desulfurization system parameters. After applying this method, the desulfurization efficiency of the desulfurization system remained stably above 98.5%, and the overall system energy consumption decreased by 8% compared to before optimization, achieving the dual objectives of achieving desulfurization efficiency targets and minimizing energy consumption.

[0031] In summary, the methods described in the above embodiments integrate and correct multi-dimensional parameters of the desulfurization system by constructing a time-labeled full-condition parameter-efficiency correlation dataset, laying a reliable data foundation for model construction. Furthermore, by combining a hybrid model of deep residual networks and attention mechanisms, the efficiency of parameter response intervals can be accurately predicted, improving the accuracy of parameter-efficiency correlation analysis. At the same time, by dividing independent optimization intervals and conducting parallel iterative optimization, the efficiency of parameter optimization is effectively improved, avoiding the problem of redundant calculations. Moreover, after full efficiency cross-comparison and boundary condition adaptation stability verification, the globally optimal parameter set selected balances desulfurization efficiency and minimum system energy consumption, exhibiting stronger adaptability and stability. In addition, the configuration of a dynamic periodic refresh mechanism can adjust the refresh duration according to the rate of change of operating conditions, ensuring that parameter control always conforms to the actual operating state of the system, making the desulfurization system more intelligent and precise overall.

[0032] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing and controlling parameters of a desulfurization system based on machine learning, characterized in that, include: Collect multi-dimensional real-time parameters of the entire desulfurization system operation process, and simultaneously collect desulfurization efficiency and system energy consumption data under corresponding operating conditions to construct a full-condition parameter-efficiency correlation dataset with time-series labels. Based on the full-condition parameter-efficiency correlation dataset, a multi-objective coupled machine learning prediction model is constructed. With desulfurization efficiency meeting the standard and system energy consumption being minimized as the dual constraints, the output efficiency response range corresponding to parameter changes is determined. Set the feasible domain boundary of the full-condition parameters of the desulfurization system, generate several sets of non-continuously distributed initial parameter optimization nodes within the feasible domain, and simultaneously define the independent optimization intervals corresponding to each node. The machine learning prediction model is applied to perform parallel iterative optimization of parameter combinations in each independent optimization interval, and the optimal parameter combination and corresponding performance index obtained in each interval are recorded simultaneously. A full performance cross-comparison is performed on the optimal parameter combinations output from all independent optimization intervals to select the globally optimal parameter set, and the adaptability and stability of the parameter set under boundary conditions are verified simultaneously. The verified global optimal parameter set is distributed to each execution unit of the desulfurization system and executed synchronously based on the dynamic periodic refresh step.

2. The desulfurization system parameter optimization control method based on machine learning according to claim 1, characterized in that, The multi-dimensional real-time parameters include inlet flue gas parameters, absorbent supply parameters, in-tower flow field parameters, circulating slurry parameters, and outlet flue gas treatment parameters of the desulfurization absorption tower. Each dimension parameter carries a millisecond-level timestamp to form a time-series label. During the construction of the full-condition parameter-efficiency association dataset, missing parameter values ​​are filled in using time-series neighborhood parameter interpolation, and abnormal parameter values ​​are corrected through condition similarity matching. ； In the formula: Cosine similarity of working condition features; The feature vector of the operating condition where the abnormal parameters are located; The feature vector represents the historical normal operating conditions. These are the corrected abnormal parameter values; M is the preset number of historical normal operating conditions with high similarity. The cosine similarity values ​​of the top M highest working condition features; The parameter value in the same dimension as the abnormal parameter in the m-th high-similarity historical normal operating condition; The dataset is divided into training set, validation set and test set according to a preset working condition division ratio, and the working condition distribution characteristics of each set are consistent.

3. The desulfurization system parameter optimization control method based on machine learning according to claim 1, characterized in that, The multi-objective coupled machine learning prediction model is a hybrid model combining a deep residual network and an attention mechanism. The loss function of the model is the coupling loss between the desulfurization efficiency deviation loss and the system energy consumption loss. The calculation of the coupling loss satisfies the following: ； In the formula: This represents the model coupling loss value. To predict desulfurization efficiency for the model; This represents the desulfurization efficiency target value. To predict system energy consumption using a model; This represents the rated basic energy consumption of the desulfurization system. An adaptive learning rate strategy is adopted during model training. The learning rate is dynamically adjusted according to the prediction accuracy of the performance response interval of the validation set until the prediction error of the model falls into the preset interval.

4. The desulfurization system parameter optimization control method based on machine learning according to claim 3, characterized in that, The , The calculation formula is: ； When dynamically adjusting the learning rate based on the performance response interval prediction accuracy of the validation set, the following applies: ； In the formula: The baseline desulfurization efficiency of the desulfurization system; This represents the total number of dimensions of the parameters for the entire desulfurization system under all operating conditions. Let be the attention weight for the i-th parameter; The desulfurization efficiency feature mapping function is extracted for the i-th parameter from the deep residual network. The activation coefficient is a parameter characteristic of desulfurization efficiency; This refers to the i-th full-condition parameter; The historical average value of the i-th parameter under operating conditions; This is the system energy consumption feature mapping function for the i-th parameter extracted by the deep residual network; The activation coefficient is a parameter characteristic of system energy consumption; Let be the model learning rate for the t-th iteration; The initial learning rate for the model; The preset maximum number of iterations for model training; Let be the overall prediction error of the model in the t-th iteration; Let be the overall prediction error of the model in the (t-1)th iteration; This is the preset minimum value correction coefficient.

5. The desulfurization system parameter optimization control method based on machine learning according to claim 1, characterized in that, The feasible domain boundary of the full-condition parameters of the desulfurization system is determined by the following formula: ； In the formula: This refers to the i-th full-condition parameter; represents the historical extreme value of the i-th parameter; k is the parameter feasible region correction coefficient; The , This represents the joint standard deviation of desulfurization efficiency versus system energy consumption. The desulfurization efficiency is the average efficiency under historical operating conditions. The average system energy consumption under historical operating conditions; Once the feasible domain boundary is determined, invalid domains corresponding to parameter combinations that have process conflicts are eliminated, and only valid feasible domains that satisfy the desulfurization process interlocking logic are retained.

6. The desulfurization system parameter optimization control method based on machine learning according to claim 1, characterized in that, In the initial parameter optimization node generation stage, a non-continuously distributed set of nodes is generated within the effective feasible region, and the independent optimization interval corresponding to each node is adaptively divided according to the node density. The interval partitioning logic is as follows: The Euclidean distance between the parameters of any two adjacent optimization nodes is greater than a preset threshold. The parameter combinations within each independent optimization interval satisfy the process coupling constraints of the desulfurization system. The number of independent optimization intervals is positively correlated with the parameter dimensions of the desulfurization system; The parameter search ranges of each interval do not overlap.

7. The desulfurization system parameter optimization control method based on machine learning according to claim 1, characterized in that, In the parallel iterative optimization phase, each independent optimization interval is configured with an independent agent. The efficiency evaluation of the parameter combinations within the interval by the agent is completed by a machine learning prediction model. The termination condition for iterative optimization is that within a preset number of consecutive iterations, the rate of change of the efficiency index is less than a preset threshold, and the parameter update amount in each iteration satisfies the following: ； In the formula: This represents the update amount of the i-th parameter in the (n+1)-th iteration; This represents the update amount of the i-th parameter in the n-th iteration; The predicted desulfurization efficiencies for the (n+1)th and nth iterations are respectively. This represents the desulfurization efficiency target value. These are the predicted system energy consumption for the (n+1)th and nth iterations, respectively. This represents the minimum energy consumption within the current optimization interval. In this process, each agent interacts with optimization information in real time during the iteration process, eliminating invalid parameter combinations that have already been searched, in order to avoid redundant calculations in global optimization.

8. The desulfurization system parameter optimization control method based on machine learning according to claim 1, characterized in that, The full-scale performance cross-comparison uses a comprehensive performance evaluation index to quantitatively evaluate each optimal parameter combination, wherein the formula for calculating the comprehensive performance evaluation index is: ； In the formula: As a comprehensive performance evaluation index; Predicted desulfurization efficiency for parameter combinations; This represents the desulfurization efficiency target value. This represents the theoretical minimum system energy consumption across the entire feasible region. Predict the energy consumption of the system based on the parameter combination; The boundary condition adaptability coefficients for the parameter combination. ∈(0,1]; The preset corrected standard deviation of the comprehensive performance evaluation index; The full-scale performance cross-comparison process is as follows: First, calculate the comprehensive performance evaluation index of the optimal parameter combination output for each independent optimization interval. Directly remove For parameter combinations below a preset threshold, a full-condition, full-coverage simulation verification covering both conventional and boundary conditions of the desulfurization system is conducted for the remaining parameter combinations. During the simulation verification process, the real-time comprehensive performance evaluation index of each parameter combination under each type of condition is calculated. Then, the average real-time comprehensive performance evaluation index of a single parameter combination under all conditions is calculated, and this average value is used as the final comprehensive performance evaluation index of that parameter combination. One or more parameter combinations with the highest final comprehensive performance evaluation index are selected as candidate sets for the globally optimal parameter set.

9. The desulfurization system parameter optimization control method based on machine learning according to claim 1, characterized in that, The adaptation stability verification under the boundary conditions includes verification of all types of boundary conditions, such as extreme flue gas volume conditions, extreme sulfur dioxide concentration conditions, absorbent quality deterioration conditions, and system low load conditions. During the verification process, the performance fluctuation coefficient of the global optimal parameter set under each boundary condition is calculated. When the performance fluctuation coefficient under all boundary conditions is less than the preset threshold, the parameter set is determined to be stable. If the performance fluctuation coefficient of any boundary condition exceeds the preset threshold, the boundary condition is included in the optimization condition set and parameter optimization is re-executed. When the global optimal parameter set is sent to each execution unit of the desulfurization system, a hierarchical parameter sending strategy is adopted. Parameter instructions are sent sequentially according to the process response priority of the execution unit. After each execution unit receives the parameter instruction, it sends back a parameter execution confirmation signal. After all execution units have sent back confirmation signals, the execution parameters of the desulfurization system are switched.

10. The desulfurization system parameter optimization control method based on machine learning according to claim 1, characterized in that, The duration of the dynamic periodic refresh is dynamically adjusted according to the rate of change of the desulfurization system's operating conditions. ； In the formula: C represents the degree of change in the real-time operating conditions of the desulfurization system; For L2 norm operation, j is the dimension number of the core operating parameters of the desulfurization system; The difference between the real-time value and the steady-state value of the j-th core operating condition parameter; Let be the steady-state value of the j-th core operating condition parameter; T is the dynamic refresh cycle. This serves as the baseline refresh cycle for the desulfurization system. The degree of variation of the desulfurization system's baseline operating conditions; The operating condition sensitivity coefficient is adapted to the desulfurization system process.

Images