A data-driven multi-objective optimization method for desulfurization system
By employing a data-driven multi-objective optimization method, combined with fuzzy clustering and evolutionary algorithms, characteristic operating state clusters of desulfurization systems are identified, and an adaptive optimization strategy library is constructed. This solves the problem of insufficient control flexibility in desulfurization systems and achieves more efficient operating condition classification and operating parameter optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- DITIAN ENVIRONMENT TECH (NANJING) CO LTD
- Filing Date
- 2026-05-26
- Publication Date
- 2026-08-04
AI Technical Summary
The existing desulfurization system has insufficient accuracy in classifying operating conditions, which makes it unable to adapt to changing field conditions, resulting in insufficient control flexibility and inability to meet the operating requirements of different operating conditions.
A data-driven multi-objective optimization method is adopted. By acquiring multi-source operating parameters, performing data cleaning and alignment processing, using an improved fuzzy clustering algorithm to identify characteristic operating state clusters, and combining a multi-objective evolutionary algorithm to search for the optimal combination of control parameters, an optimization control strategy library with adaptive operating conditions is constructed.
It improves the accuracy of operating condition classification and dynamic adaptability, enhances the flexibility of system control, better adapts to complex and ever-changing on-site operating environments, and improves the operating efficiency and energy consumption management of the desulfurization system.
Smart Images
Figure CN122260877B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of environmental desulfurization control technology, and in particular to a data-driven multi-objective optimization method for desulfurization systems. Background Technology
[0002] Industrial desulfurization systems are core facilities for flue gas pollution control. During actual operation, they are affected by multiple factors such as flue gas load, material properties, and equipment wear and tear, resulting in significant dynamic fluctuations in operating conditions. Current conventional optimization and control methods for desulfurization systems largely rely on fixed mechanism models to build a control framework, using basic clustering to simply divide operating conditions, employing unified optimization algorithms to centrally optimize parameters across all operating conditions, and relying on fixed control logic to complete system parameter adjustments.
[0003] Conventional fuzzy clustering operations use fixed rules to update cluster centers, and the calculation process only refers to the difference in the value of a single parameter to complete the classification, ignoring the dynamic coupling characteristics between various operating parameters of the desulfurization system. The classification results deviate from the actual operating state of the equipment, and complex transitional operating conditions cannot be effectively distinguished and identified. A unified global parameter optimization model cannot meet the needs of different operating conditions, and a single combination of control parameters is difficult to adapt to changes in multiple operating scenarios, resulting in insufficient system control flexibility and decreased matching degree of operating parameters.
[0004] To address the challenges of uncovering the dynamic correlation characteristics of desulfurization system parameters and the insufficient accuracy of operating condition classification, it is necessary to adjust the internal computational logic of the clustering algorithm to adapt to the dynamic changes in parameters and complete the classification of operating states. Faced with diverse on-site operating conditions, it is necessary to break away from a uniform optimization model and conduct independent parameter optimization calculations based on different operating states to meet the control requirements of differentiated operating conditions. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a data-driven multi-objective optimization method for desulfurization systems.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: a data-driven multi-objective optimization method for desulfurization systems, comprising: Obtain a set of multi-source operating parameters of the target desulfurization system within a preset time period. The set of multi-source operating parameters includes a set of inlet flue gas parameters, a set of absorbent parameters, a set of process control parameters, and a set of outlet emission parameters. Data cleaning and alignment processing is performed on the multi-source operating parameter set to generate time-aligned multi-dimensional operating status time-series data; An improved fuzzy clustering algorithm is applied to the multidimensional operating state time series data to identify several characteristic operating state clusters that characterize typical operating conditions of the system. The improved fuzzy clustering algorithm optimizes the cluster center update mechanism based on the time-varying coupling relationship of system operating parameters. Extract the centroid vector and boundary features of each characteristic operating state cluster, and combine them with the preset optimization target set of the target desulfurization system. Then, use a multi-objective evolutionary algorithm to search in parallel for the optimal combination of control parameters under each characteristic operating state cluster. Based on the optimal control parameter combination obtained by searching under each characteristic operating state cluster, an optimized control strategy library for operating conditions is constructed. The multidimensional operating status time series data is input into the trained operating condition recognition model in real time to obtain the feature operating status cluster identifier corresponding to the current operating condition. Based on the characteristic operating state cluster identifier, the system queries the adaptive optimization control strategy library for the current operating condition to obtain the optimal control parameter combination corresponding to the current operating condition, and then sends it to the underlying controller of the target desulfurization system for execution.
[0007] As a further aspect of the present invention, data cleaning and alignment processing is performed on the multi-source operating parameter set to generate time-aligned multi-dimensional operating state time-series data, including: Outlier detection and smoothing are performed on each parameter sequence in the inlet flue gas parameter set, absorbent parameter set, process control parameter set, and outlet emission parameter set, respectively. To ensure that all processed parameter sequences have the same sampling time interval; All parameter sequences are interpolated and aligned with time as the reference axis to ensure that at each identical sampling time point, there are parameter values from different sets, forming a synchronized multidimensional data point sequence; The synchronized multidimensional data point sequence is organized in chronological order to form the time-aligned multidimensional running status time series data.
[0008] As a further aspect of the present invention, an improved fuzzy clustering algorithm is applied to the multidimensional operating state time-series data to identify several characteristic operating state clusters representing typical operating conditions of the system, including: Initialize a set of random cluster centers, and assign a membership degree to each cluster center to each data point in the multidimensional running state time series data; The improved fuzzy clustering algorithm iteratively updates cluster centers and membership degrees. Each time the cluster centers are updated, the improved fuzzy clustering algorithm considers not only the spatial distribution of data points but also introduces temporal correlation weights. The calculation process of the temporal correlation weights includes: calculating the Euclidean distance between the current data point and its immediate neighbors in the time series; taking the reciprocal of the distance and normalizing it to obtain a similarity coefficient; performing a weighted summation of the similarity coefficient sequence to obtain the degree of local stationarity; and then mapping the degree of local stationarity to a temporal correlation weight using a monotonically increasing function. When updating the cluster centers, the contribution value of each data point is multiplied by its corresponding temporal correlation weight. The algorithm is considered to have converged when the change in the location of the cluster centers is less than a preset threshold. Based on the converged cluster centers and the maximum membership degree corresponding to each data point, the multidimensional operating state time series data is divided into clusters represented by the corresponding cluster centers, and each cluster is a characteristic operating state cluster.
[0009] As a further aspect of the present invention, the improved fuzzy clustering algorithm introduces temporal correlation weights each time the cluster centers are updated, and the calculation process includes: For the current data point, calculate its Euclidean distance to several adjacent data points before and after it in the multidimensional running state time series data, and obtain a distance sequence composed of multiple Euclidean distance values; The calculated distance sequence is taken inversely and normalized to obtain a set of coefficients representing the similarity between the current data point and its neighboring data points; Calculate a weighted sum of similarity coefficient sequences, whereby the weighted sum measures the degree of local stationarity of the current data point in the time series. The weighted sum is mapped through a monotonically increasing function to obtain the final temporal correlation weight, which is used to amplify or reduce the contribution of the current data point in updating the cluster center.
[0010] As a further aspect of the present invention, the step of extracting the centroid vector and boundary features of each characteristic operating state cluster, combining them with the preset optimization target set of the target desulfurization system, and using a multi-objective evolutionary algorithm to search in parallel for the optimal combination of control parameters under each characteristic operating state cluster includes: The preset set of optimization targets includes desulfurization efficiency, system energy consumption, and operating costs; For a cluster of characteristic operating states, its centroid vector is used as the typical operating condition parameter of the cluster of characteristic operating states; Using the adjustable variables in the set of process control parameters as decision variables, an initial population of candidate control parameter combinations is randomly generated within the feasible domain of the decision variables. The desulfurization efficiency, system energy consumption, and operating cost simulated by operating the target desulfurization system under the typical operating conditions are calculated using each candidate control parameter combination, and are used as the objective function value of the corresponding candidate control parameter combination. Based on the multi-objective evolutionary algorithm, the population is subjected to selection, crossover, and mutation operations to generate a new population, and the objective function value is repeatedly calculated to simultaneously optimize desulfurization efficiency, system energy consumption, and operating costs. When the evolutionary process reaches the termination condition, output the Pareto optimal solution set composed of all non-dominated solutions in the current population, and select a solution from the Pareto optimal solution set according to the preset preference as the optimal control parameter combination under the corresponding feature running state cluster. The optimal control parameter combination search process for all feature running state clusters is performed in parallel.
[0011] As a further aspect of the present invention, the step of constructing an adaptive optimization control strategy library based on the optimal control parameter combination obtained from searching under each characteristic operating state cluster includes: Create a policy record for each feature running state cluster. The policy record shall at least include the identifier, centroid vector, boundary features and corresponding optimal control parameter combination of the corresponding feature running state cluster. All policy records corresponding to feature running state clusters are indexed and organized to form a mapping data structure with feature running state cluster identifiers as keys and corresponding policy records as values. The mapping relationship data structure is persistently stored to form a queryable library of adaptive optimization control strategies for the operating conditions.
[0012] As a further aspect of the present invention, the multi-dimensional operating state time-series data is input into the trained operating condition recognition model in real time to obtain the feature operating state cluster identifier corresponding to the current operating condition, including: The operating condition identification model is a classification model trained based on historical multidimensional operating state time series data and its labeled feature operating state cluster identifiers. The latest operating parameters of the target desulfurization system are collected in real time within a certain time window. After data cleaning and alignment, a current multi-dimensional operating status data sample is formed. Input the current multidimensional operating status data sample into the operating condition identification model; The operating condition identification model outputs a probability distribution vector, where each component of the probability distribution vector represents the probability that the current multidimensional operating state data sample belongs to each feature operating state cluster. The identifier of the feature operating state cluster with the highest probability is used as the feature operating state cluster identifier corresponding to the current operating condition.
[0013] As a further aspect of the present invention, the optimal control strategy library for adaptive operation is queried based on the characteristic operating state cluster identifier to obtain the optimal control parameter combination corresponding to the current operating condition, and then sent to the underlying controller of the target desulfurization system for execution, including: Using the characteristic operating state cluster identifier as the query key, a search is performed in the operating condition adaptive optimization control strategy library; Obtain the retrieved strategy records and extract the optimal combination of control parameters from them; The optimal control parameter combination is encapsulated according to an instruction format recognizable by the underlying controller; Through a preset communication interface, the encapsulated control commands are sent to the underlying controller of the target desulfurization system. The underlying controller parses the commands and adjusts the corresponding actuators so that the system operating parameters approach the optimal combination of control parameters.
[0014] As a further aspect of the present invention, the calculation of the desulfurization efficiency, system energy consumption, and operating cost simulated by running the target desulfurization system under the typical operating parameters for each candidate control parameter combination is used as the objective function value for the corresponding candidate control parameter combination, including: Based on the uncontrollable variables in the typical operating condition parameters and the controllable variables in the current candidate control parameter combination, a complete input vector is constructed to describe the operating state of the target desulfurization system. The complete input vector is input into a pre-trained desulfurization system performance proxy model; Through forward calculation of the desulfurization system performance proxy model, the predicted desulfurization efficiency, predicted system energy consumption and predicted operating cost of the target desulfurization system are obtained under the system operating state described by the complete input vector. The values of predicted desulfurization efficiency, predicted system energy consumption, and predicted operating cost are extracted from the output of the desulfurization system performance proxy model. The extracted predicted desulfurization efficiency is negatively evaluated as the first objective function value to be minimized, the predicted system energy consumption is evaluated as the second objective function value to be minimized, and the predicted operating cost is evaluated as the third objective function value to be minimized, thus completing the calculation of the objective function value for the current candidate control parameter combination.
[0015] As a further aspect of the present invention, when the evolutionary process reaches the termination condition, outputting a Pareto optimal solution set composed of all non-dominated solutions in the current population, and selecting a solution from the Pareto optimal solution set according to a preset preference as the optimal control parameter combination under the corresponding feature running state cluster, includes: Calculate and output all non-dominated solutions in the current population at the end of the evolutionary process, forming a Pareto optimal solution set; From the Pareto optimal solution set, extract the three objective function values of desulfurization efficiency, system energy consumption and operating cost corresponding to each nondominated solution; The objective function value vector of each non-dominated solution is input into a preset preference decision model. The preset preference decision model is based on the actual operating constraints and priorities of the target desulfurization system and pre-sets preference weight vectors for the three optimization objectives of desulfurization efficiency, system energy consumption, and operating cost. The preset preference decision model calculates the weighted sum of the objective function value vector and the preset preference weight vector for each non-dominated solution, and normalizes the weighted sum into a comprehensive preference score. The comprehensive preference scores of all non-dominated solutions are sorted in descending order, and the non-dominated solution with the highest comprehensive preference score is selected as the optimal control parameter combination under the corresponding characteristic operating state cluster.
[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By combining the time-varying coupling relationship of desulfurization system operating parameters, the cluster center update mechanism within the fuzzy clustering algorithm is adjusted, moving away from the fixed computational logic of traditional clustering algorithms. The interconnected changes in multi-dimensional time-series operating data are integrated, and the dynamic correlation patterns formed by various operating parameters are combined to complete the operating condition classification, mitigating classification bias caused by differences in single parameter values. The dimensions for determining operating states are increased, and the boundary division standards for different operating state clusters are unified, improving the state differentiation level during the dynamic operation of the system. The inherent correlation patterns in time-series operating data are mined, and holistic analysis is performed on continuously changing operating data, improving the precision of operating condition identification under complex fluctuation conditions. This ensures that the operating condition classification results closely match the actual changes generated by continuous equipment operation, improving the analysis efficiency and utilization level of massive operating data.
[0017] The centroid vectors and boundary features of each characteristic operating state cluster are extracted, and independent optimization operation intervals are segmented. Parallel parameter calculations are performed within each interval using a multi-objective evolutionary algorithm. This transforms the globally unified parameter optimization mode, using independent operating state clusters as basic units to complete differentiated parameter solutions, reducing the coverage of optimization operations and improving the adjustment and adaptation dimensions of control parameters. Parameter combinations with higher matching degrees are selected based on diverse control requirements, and control parameter content corresponding to different operating states is accumulated. Corresponding parameter combinations are matched according to real-time operating conditions, adapting to dynamic state changes during operation, expanding the adaptation forms of parameter control, improving the fit between control parameters and differentiated operating conditions, enhancing the dynamic adaptability of the system's control mode, and adapting to the long-term, complex, and ever-changing field operating environment of the desulfurization system. Attached Figure Description
[0018] Figure 1 This is a flowchart of a data-driven multi-objective optimization method for a desulfurization system according to the present invention; Figure 2 A flowchart for generating time-aligned multidimensional runtime status time-series data for data cleaning and alignment; Figure 3 A flowchart for improving the fuzzy clustering algorithm to identify feature-based running state clusters. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0021] See Figure 1 A data-driven multi-objective optimization method for desulfurization systems is proposed, with the following overall implementation scheme: A set of multi-source operating parameters of the target desulfurization system within a preset time period is obtained. This set includes inlet flue gas parameters, absorbent parameters, process control parameters, and outlet emission parameters. Data cleaning and alignment are performed on the multi-source operating parameter set to generate time-aligned multi-dimensional operating state time-series data. An improved fuzzy clustering algorithm is applied to the multi-dimensional operating state time-series data. This algorithm optimizes the cluster center update mechanism based on the time-varying coupling relationship of the system operating parameters, thereby identifying several characteristic operating state clusters representing typical system operating conditions. The centroid vector and boundary features of each characteristic operating state cluster are extracted, and combined with the preset optimization objective set of the target desulfurization system, a multi-objective evolutionary algorithm is used to search in parallel for the optimal control parameter combination under each characteristic operating state cluster. Based on the optimal control parameter combination obtained under each characteristic operating state cluster, an adaptive optimization control strategy library is constructed. In actual operation, the multi-dimensional operating state time-series data is input into the trained operating condition recognition model in real time to obtain the characteristic operating state cluster identifier corresponding to the current operating condition. Based on the characteristic operating state cluster identifier, query the adaptive optimization control strategy library for the current operating condition, obtain the optimal control parameter combination corresponding to the current operating condition, and send it to the underlying controller of the target desulfurization system for execution.
[0022] In one embodiment of the present invention, see [reference] Figure 2 Outlier detection and smoothing were performed on each parameter sequence in the inlet flue gas parameter set, absorbent parameter set, process control parameter set, and outlet emission parameter set. A uniform sampling time interval was set for each processed parameter sequence. All parameter sequences were interpolated and aligned along the time axis to ensure that parameter values from different sets existed at each identical sampling time point, forming a synchronized multidimensional data point sequence. This synchronized multidimensional data point sequence was organized chronologically to constitute time-aligned multidimensional operational status time-series data.
[0023] In practical implementation, after obtaining the multi-source operating parameter set of the target desulfurization system within a preset time period, outlier detection and smoothing are performed on each parameter sequence in the inlet flue gas parameter set, absorbent parameter set, process control parameter set, and outlet emission parameter set. The processing is carried out for each independent parameter sequence. The flue gas flow rate sequence in the inlet flue gas parameter set may use an outlier identification method based on statistical quantiles to identify and correct values exceeding the preset quantile range. The slurry pH value sequence in the absorbent parameter set may use a sliding window mean filter for smoothing to suppress high-frequency fluctuations. The circulating pump frequency sequence in the process control parameter set and the sulfur dioxide concentration sequence in the outlet emission parameter set are also selected with appropriate outlier handling and smoothing algorithms based on their data characteristics. A uniform sampling time interval is set for all processed parameter sequences. This uniform sampling time interval is determined comprehensively based on the dynamic characteristics and control cycle of the desulfurization system process. For example, the sampling time interval for all parameter sequences is uniformly set to 1 minute, so that subsequent time alignment operations have a common time reference. All parameter sequences are interpolated and aligned along the time axis. At each identical sampling time point, linear interpolation or spline interpolation algorithms are used to estimate the corresponding parameter values for those parameter sequences missing data points at the original timestamps. This ensures that parameter values from the inlet flue gas parameter set, absorbent parameter set, process control parameter set, and outlet emission parameter set exist at each identical sampling time point. This process generates a multidimensional data point sequence that is strictly synchronized in the time dimension. The synchronized multidimensional data point sequence is organized in chronological order, with each sampling time point corresponding to a vector containing the values of all parameter sets. These vectors are arranged according to the order of their timestamps, forming time-aligned multidimensional operational status time-series data. Mathematically, this can be represented as a matrix, where the rows of the matrix correspond to time points, and the columns correspond to different operational parameters.
[0024] In some embodiments, outlier detection employs a method based on three standard deviations. For each data point in the parameter sequence, its deviation from the sequence mean is calculated. If the absolute value of the deviation exceeds three times the sequence standard deviation, the data point is marked as an outlier, and the outlier is replaced by the linear interpolation result of the adjacent data points. In some embodiments, smoothing uses an exponentially weighted moving average algorithm, which assigns higher weights to recent data. Its smoothing calculation relationship can be expressed as:
[0025] in: This represents the smoothed parameter value at time t. This represents the original parameter value at time t. This represents the smoothed parameter value from the previous time step. This is a smoothing coefficient between 0 and 1. Optionally, the interpolation method used in the time alignment process is piecewise linear interpolation, which constructs a linear function between any two adjacent known data points to calculate the parameter value at the intermediate time point. Optionally, when the original sampling frequency of some parameter sequences is higher than the uniformly set sampling time interval, the high-frequency sampled data is downsampled, and the parameter value corresponding to that window is obtained by calculating the average value of the original data within each uniform time window. It can be understood that the setting of the uniform sampling time interval needs to be no greater than the minimum sampling interval of all original parameter sequences to ensure that no information is lost. It can be understood that the generated multidimensional operating state time series data is the basis for subsequent operating condition clustering and optimization search, and its quality directly depends on the rigor of data cleaning and alignment processing.
[0026] In one embodiment of the present invention, see [reference] Figure 3A set of random cluster centers is initialized, and each data point in the multidimensional running state time series data is assigned a membership degree to each cluster center. The cluster centers and membership degrees are iteratively updated using an improved fuzzy clustering algorithm. This improved fuzzy clustering algorithm considers not only the spatial distribution of data points but also introduces temporal correlation weights during each cluster center update, making data points that are temporally adjacent and have similar parameter values have a stronger influence on the cluster center update. The algorithm is considered converged when the positional change of the cluster centers is less than a preset threshold. Based on the converged cluster centers and the maximum membership degree corresponding to each data point, the multidimensional running state time series data is divided into clusters represented by the corresponding cluster centers, with each cluster being a feature running state cluster. The calculation process of the temporal correlation weights includes: for the current data point, calculating its Euclidean distance to several adjacent data points in the multidimensional running state time series data, obtaining a distance sequence composed of multiple Euclidean distance values. The reciprocal of the calculated distance sequence is taken and normalized to obtain a set of coefficients representing the similarity between the current data point and its neighboring data points. The weighted sum of the similarity coefficient sequence is calculated, which measures the degree of local stationarity of the current data point in the time series. The weighted sum is then mapped through a monotonically increasing function to obtain the final temporal correlation weight, which is used to amplify or diminish the contribution of the current data point in updating the cluster centers.
[0027] In practical implementation, an improved fuzzy clustering algorithm is applied to the time-aligned multidimensional operational status time-series data. A set of random cluster centers is initialized, with the number of cluster centers pre-set based on domain knowledge. Each data point in the multidimensional operational status time-series data is assigned a membership degree to each cluster center. The membership degree is a value between 0 and 1, representing the degree to which a data point belongs to a certain cluster. The sum of the membership degrees of all data points to all cluster centers is 1. The cluster centers and membership degrees are iteratively updated according to the improved fuzzy clustering algorithm. In each update of the cluster centers, the improved fuzzy clustering algorithm considers not only the spatial distribution of data points but also introduces temporal correlation weights. In the calculation formula for updating the cluster center position, the contribution of each data point is multiplied by its corresponding temporal correlation weight. This makes data points that are temporally adjacent and have similar parameter values have a stronger influence on the update of the cluster centers. This mechanism allows the generated cluster centers to better reflect the continuous and stable operation phase of the system over time. When the change in the position of the cluster centers is less than a preset threshold, the algorithm is considered to have converged. The preset threshold is a small positive number close to zero. Based on the converged cluster centers and the maximum membership degree corresponding to each data point, the multidimensional operating state time series data are divided into the clusters represented by the corresponding cluster centers. Each data point is classified into the cluster where the cluster center with the maximum membership degree is located. Each cluster is a characteristic operating state cluster, representing a typical operating condition mode of the desulfurization system.
[0028] In some embodiments, the calculation process of temporal correlation weights includes: for the current data point in the multidimensional running state time series data, calculating its Euclidean distance with several adjacent data points in the multidimensional running state time series data. The number of neighboring data points can be set to a fixed value, resulting in a distance sequence composed of multiple Euclidean distance values. The calculated distance sequence is then normalized by taking its reciprocal. The normalization operation makes the sum of all reciprocals equal to 1, thus obtaining a set of coefficients with a sum of 1, representing the similarity between the current data point and its neighboring data points. The higher the similarity, the larger the coefficient value. A weighted sum of the similarity coefficient sequence is calculated. The specific calculation method for the weighted sum is to multiply each similarity coefficient by a decreasing weight related to temporal proximity. This weighted sum is used to measure the degree of local stationarity of the current data point in the time series. A larger weighted sum indicates that the parameter fluctuations within the local time segment where the current data point is located are smaller and more stable. The weighted sum is mapped through a monotonically increasing function, which can be either a sigmoid function or a linear scaling function, to obtain the final temporal correlation weight. This temporal correlation weight is directly used to amplify or reduce the contribution of the current data point in updating the cluster centers. In some embodiments, the temporal correlation weight is calculated using the following formula:
[0029] in: It is the temporal correlation weight of data point i. It is a monotonically increasing scaling function. It is a set of neighboring indices of data point i in the time series, containing the indices of data points within a certain time window before and after point i. It is a decay factor between 0 and 1. This represents the absolute value of the time index difference between neighboring point j and the current point i. This is the normalized similarity coefficient between data point i and its neighbor j. Optionally, the membership update rule in the improved fuzzy clustering algorithm has also been adjusted. When calculating the spatial distance between a data point and the cluster center, the temporal correlation weight factor of that data point is also incorporated, making the influence of data points within a stable time segment on the membership distribution greater. Optionally, the preset threshold is set based on observations from multiple experiments to ensure that the positional change of the cluster center is sufficiently small in continuous iterations, thus considering the clustering result to be stable. It can be understood that introducing temporal correlation weights is the core difference between the improved fuzzy clustering algorithm and the traditional fuzzy clustering algorithm. It enables the clustering process to perceive the temporal continuity of the data, thus giving the resulting feature operating state clusters a more explicit physical meaning in the time dimension, and more accurately corresponding to the steady-state or quasi-steady-state operating conditions of the system. It can be understood that although the initial position of the cluster center is random, through the iterative update process of introducing temporal correlation weights, the final converged cluster center position can effectively represent the typical stable operating point of the system in the multi-dimensional parameter space.
[0030] In one embodiment of the present invention, the preset optimization target set includes desulfurization efficiency, system energy consumption, and operating cost. For a characteristic operating state cluster, its centroid vector is used as the typical operating condition parameter of the characteristic operating state cluster. Using the adjustable variables in the process control parameter set as decision variables, an initial population of candidate control parameter combinations is randomly generated within the feasible region of the decision variables. The desulfurization efficiency, system energy consumption, and operating cost simulated by running the target desulfurization system under typical operating conditions with each candidate control parameter combination are calculated and used as the objective function value of the corresponding candidate control parameter combination. According to a multi-objective evolutionary algorithm, selection, crossover, and mutation operations are performed on the population to generate a new population, and the objective function value is repeatedly calculated to simultaneously optimize desulfurization efficiency, system energy consumption, and operating cost. When the evolutionary process reaches the termination condition, the Pareto optimal solution set consisting of all non-dominated solutions in the current population is output. From the Pareto optimal solution set, a solution is selected as the optimal control parameter combination under the corresponding characteristic operating state cluster according to a preset preference. This search process is executed in parallel for all characteristic operating state clusters.
[0031] The process of calculating the objective function value includes: constructing a complete input vector describing the operating state of the target desulfurization system based on uncontrollable variables in typical operating condition parameters and controllable variables in the current candidate control parameter combination. This complete input vector is then input into a pre-trained desulfurization system performance surrogate model. Through forward computation of the desulfurization system performance surrogate model, the predicted desulfurization efficiency, predicted system energy consumption, and predicted operating cost of the target desulfurization system under the system operating state described by the complete input vector are obtained. The values of the predicted desulfurization efficiency, predicted system energy consumption, and predicted operating cost are extracted from the output of the desulfurization system performance surrogate model. The extracted predicted desulfurization efficiency is negatively evaluated and used as the first objective function value to be minimized; the predicted system energy consumption is used as the second objective function value to be minimized; and the predicted operating cost is used as the third objective function value to be minimized, thus completing the calculation of the objective function value for the current candidate control parameter combination. The process of selecting a solution from the Pareto optimal solution set includes: calculating and outputting all non-dominated solutions in the current population at the end of the evolutionary process, constituting the Pareto optimal solution set. From the Pareto optimal solution set, the three objective function values—desulfurization efficiency, system energy consumption, and operating cost—are extracted for each non-dominated solution. The objective function value vector of each non-dominated solution is input into a pre-defined preference decision model. This model, based on the actual operating constraints and priorities of the target desulfurization system, pre-sets preference weight vectors for the three optimization objectives: desulfurization efficiency, system energy consumption, and operating cost. The pre-defined preference decision model calculates the weighted sum of the objective function value vector of each non-dominated solution and the pre-defined preference weight vector, and normalizes the weighted sum into a comprehensive preference score. All non-dominated solutions are sorted in descending order of their comprehensive preference scores, and the non-dominated solution with the highest comprehensive preference score is selected as the optimal control parameter combination under the corresponding characteristic operating state cluster.
[0032] In practical implementation, after extracting the centroid vector and boundary features of each characteristic operating state cluster, and combining them with the preset optimization objective set of the target desulfurization system, a multi-objective evolutionary algorithm is used to search in parallel for the optimal combination of control parameters under each characteristic operating state cluster. The preset optimization objective set includes desulfurization efficiency, system energy consumption, and operating cost. For a specific characteristic operating state cluster, its centroid vector is used as the typical operating condition parameter of this characteristic operating state cluster. The typical operating condition parameter includes operating parameters that cannot be directly adjusted by the controller, such as inlet sulfur dioxide concentration, flue gas flow rate, and absorbent concentration. Adjustable variables in the process control parameter set are used as decision variables. Decision variables may include the operating frequency of the circulating pump, the valve opening degree of the slurry spray layer, and the frequency of the oxidation fan. Within the feasible domain of the decision variables, an initial candidate control parameter combination population containing a certain number of individuals (e.g., 100) is randomly generated. Each individual in the population represents a specific value of a set of decision variables. The desulfurization efficiency, system energy consumption, and operating cost simulated by running the target desulfurization system under the typical operating parameters when using each candidate control parameter combination are calculated and used as the objective function value of the corresponding candidate control parameter combination individual. Based on a multi-objective evolutionary algorithm, selection, crossover, and mutation operations are performed on the population to generate new offspring populations. The objective function value of each individual in the offspring population is repeatedly calculated. This iterative process aims to simultaneously optimize desulfurization efficiency, system energy consumption, and operating costs. When the evolutionary process reaches a predetermined number of iterations or a convergence degree of the solution set, a Pareto optimal solution set consisting of all non-dominated solutions in the current population is output. From the Pareto optimal solution set, a solution is selected as the optimal control parameter combination under the corresponding characteristic operating state cluster according to a preset preference. The search process for the optimal control parameter combination of the characteristic operating state cluster is performed in parallel for all characteristic operating state clusters.
[0033] In some embodiments, the process of calculating the objective function value of an individual candidate control parameter combination includes: constructing a complete input vector describing the operating state of the target desulfurization system based on uncontrollable variables in typical operating condition parameters and controllable variables in the current candidate control parameter combination. This complete input vector is then input into a pre-trained desulfurization system performance proxy model, which is a neural network model trained based on historical data. Through forward computation of the desulfurization system performance proxy model, the predicted desulfurization efficiency, predicted system energy consumption, and predicted operating cost of the target desulfurization system under the system operating state described by the complete input vector are obtained. The values of the predicted desulfurization efficiency, predicted system energy consumption, and predicted operating cost are extracted from the output of the desulfurization system performance proxy model, respectively. The extracted predicted desulfurization efficiency is negatively evaluated and used as the first objective function value to be minimized, the predicted system energy consumption as the second objective function value to be minimized, and the predicted operating cost as the third objective function value to be minimized, thereby completing the calculation of the objective function value for the individual candidate control parameter combination. In some embodiments, the process of selecting a solution from the Pareto optimal solution set includes: calculating and outputting all non-dominated solutions in the current population at the end of the evolutionary process, constituting the Pareto optimal solution set. From the Pareto optimal solution set, extract the three objective function values (desulfurization efficiency, system energy consumption, and operating cost) corresponding to each non-dominated solution. Input the objective function value vector of each non-dominated solution into a preset preference decision model. Based on the actual operating constraints and priorities of the target desulfurization system, the preset preference decision model pre-sets preference weight vectors for the three optimization objectives: desulfurization efficiency (0.5), system energy consumption (0.3), and operating cost (0.2). The preset preference decision model calculates the weighted sum of the objective function value vector and the preset preference weight vector for each non-dominated solution, and normalizes the weighted sum to a comprehensive preference score between 0 and 1. Sort all non-dominated solutions in descending order of their comprehensive preference scores, and select the non-dominated solution with the highest comprehensive preference score as the optimal control parameter combination under the corresponding characteristic operating state cluster. Optionally, an example formula for calculating the comprehensive preference score in the preset preference decision model is as follows:
[0034] in: This represents the overall preference score of the k-th non-dominated solution. This represents the preset preference weight for the m-th optimization objective. This represents the result of minimax normalization of the m-th objective function value of the k-th non-dominated solution. Optionally, see Table 1, which shows a simplified Pareto optimal solution set containing three non-dominated solutions, their corresponding objective function values, and the calculated comprehensive preference scores.
[0035] Table 1: Pareto optimal solution set of characteristic operating state cluster A
[0036] It is understandable that multi-objective evolutionary algorithms, by simulating the selection, crossover, and mutation mechanisms in biological evolution, can efficiently search for a compromise set of solutions that simultaneously satisfy multiple optimization objectives in the multidimensional space of decision variables. It is also understandable that selecting the final optimal combination of control parameters from the Pareto optimal solution set according to preset preferences is a key step in combining multi-objective optimization problems with practical engineering decision-making needs, giving the optimization results a clear engineering application orientation.
[0037] In one embodiment of the present invention, a strategy record is created for each characteristic operating state cluster. This strategy record includes at least the identifier, centroid vector, boundary features, and corresponding optimal control parameter combination of the corresponding characteristic operating state cluster. All strategy records corresponding to all characteristic operating state clusters are indexed and organized to form a mapping data structure with the characteristic operating state cluster identifier as the key and the corresponding strategy record as the value. This mapping data structure is persistently stored to form a queryable, condition-adaptive optimization control strategy library.
[0038] In practical implementation, an adaptive optimization control strategy library is constructed based on the optimal control parameter combinations obtained from searching under each characteristic operating state cluster. A strategy record is created for each characteristic operating state cluster, containing the cluster's identifier, centroid vector, boundary features, and corresponding optimal control parameter combination. The cluster identifier can be a unique integer or string; the centroid vector is a multi-dimensional numerical vector representing the average operating parameters of the typical operating condition represented by the cluster; the boundary features can be described as the range or standard deviation of data points within the cluster across various dimensions; and the optimal control parameter combination is a specific control parameter setting selected from the Pareto optimal solution set obtained by the multi-objective evolutionary algorithm based on preset preferences. All strategy records corresponding to the characteristic operating state clusters are indexed and organized. The indexing process assigns a unique key to each strategy record, typically the cluster identifier, forming a mapping data structure with the cluster identifier as the key and the corresponding strategy record as the value. This data structure can be represented as a hash table or dictionary in a computer. The mapping relationship data structure is persistently stored. This persistent storage writes the mapping relationship data structure from memory to a non-volatile storage medium, forming a queryable, condition-adaptive optimization control strategy library. This library allows for quick retrieval of the corresponding optimal control strategy using characteristic operating state cluster identifiers. In some embodiments, strategy records are defined using a structured data format, with each strategy record containing multiple fields. See Table 2, which illustrates the possible data structures of a strategy record.
[0039] Table 2: Feature Running State Cluster Strategy Record Structure Table
[0040] In some embodiments, indexing is implemented by constructing a hash map where the key is the feature runtime cluster identifier and the value is a pointer to the memory address of the complete policy record or a serialized policy record data block. This organization supports query operations with near constant time complexity. Optionally, persistent storage saves the optimization control policy library as a table in a relational database. Each row in the table corresponds to a policy record, and the columns of the table correspond to the fields of the policy record. A primary key index is built using the feature runtime cluster identifier column to achieve efficient querying. Optionally, persistent storage saves the optimization control policy library as a structured file, such as a JSON or XML file. The file content is a nested object with the feature runtime cluster identifier as the key. The persistent relationship of the optimization control policy library can be represented as follows:
[0041] in: This represents the persisted optimization control strategy library, where C is the set of identifiers for all feature runtime state clusters. It is identified as The complete strategy record corresponding to the characteristic operating state cluster. It can be understood that the adaptive operating condition optimization control strategy library is the result of structured storage of knowledge obtained from offline optimization, serving as the core link between offline operating condition analysis and online optimization control. It can also be understood that the construction of the optimization control strategy library enables the systematic management and invocation of pre-calculated optimal control strategies for different typical operating conditions.
[0042] In one embodiment of the present invention, the operating condition identification model is a classification model trained based on historical multidimensional operating state time-series data and its labeled feature operating state cluster identifiers. Operating parameters within the latest time window are collected from the target desulfurization system in real time. After data cleaning and alignment, a current multidimensional operating state data sample is formed. This current multidimensional operating state data sample is input into the operating condition identification model. The operating condition identification model outputs a probability distribution vector, where each component represents the probability that the current multidimensional operating state data sample belongs to each feature operating state cluster. The identifier of the feature operating state cluster with the highest probability is used as the feature operating state cluster identifier corresponding to the current operating condition. The process of querying the operating condition adaptive optimization control strategy library based on the feature operating state cluster identifier, obtaining the optimal control parameter combination corresponding to the current operating condition, and sending it to the underlying controller of the target desulfurization system includes: using the feature operating state cluster identifier as the query key to search the operating condition adaptive optimization control strategy library; obtaining the retrieved strategy records and extracting the optimal control parameter combination from them; and encapsulating the optimal control parameter combination according to an instruction format recognizable by the underlying controller. Through a pre-defined communication interface, the encapsulated control commands are sent to the underlying controller of the target desulfurization system. The underlying controller parses the commands and adjusts the corresponding actuators to make the system operating parameters approach the optimal combination of control parameters.
[0043] In practical implementation, multi-dimensional operating status time-series data is input in real time into a pre-trained operating condition identification model. This model is a classification model trained based on historical multi-dimensional operating status time-series data and its labeled feature operating status cluster identifiers. The historical multi-dimensional operating status time-series data has undergone data cleaning and alignment processing, and the labeled feature operating status cluster identifiers are generated by an improved fuzzy clustering algorithm. Operating parameters within the latest time window are collected from the target desulfurization system in real time. These real-time collected parameters include inlet flue gas parameters, absorbent parameters, process control parameters, and outlet emission parameters. After undergoing the same data cleaning and alignment process as in the offline processing stage, a current multi-dimensional operating status data sample is formed. This current multi-dimensional operating status data sample is input into the operating condition identification model, which outputs a probability distribution vector. The dimension of the probability distribution vector is equal to the total number K of feature operating status clusters, and each component of the probability distribution vector represents the probability that the current multi-dimensional operating status data sample belongs to each feature operating status cluster. The identifier of the feature operating status cluster with the highest probability is used as the feature operating status cluster identifier corresponding to the current operating condition. Based on the characteristic operating state cluster identifier, query the adaptive optimization control strategy library for the current operating condition, obtain the optimal control parameter combination corresponding to the current operating condition, and send it to the underlying controller of the target desulfurization system for execution.
[0044] In some embodiments, the characteristic operating state cluster identifier is used as the query key to search the adaptive optimization control strategy library. The search process utilizes a mapping relationship data structure to achieve fast lookup. The retrieved strategy records are obtained, and the optimal control parameter combination is extracted from them. The optimal control parameter combination includes specific values such as the circulating pump frequency setpoint and the slurry spray valve opening setpoint. The optimal control parameter combination is encapsulated according to an instruction format recognizable by the underlying controller. The encapsulation process may convert floating-point numerical parameters into messages or signals of a specific protocol. Through a preset communication interface, the encapsulated control instruction is sent to the underlying controller of the target desulfurization system. The underlying controller parses the instruction and adjusts the corresponding actuators to make the system operating parameters approach the optimal control parameter combination, completing one closed-loop optimization control. In some embodiments, the operating condition identification model adopts a deep neural network structure, and its last layer uses the Softmax activation function to output a probability distribution for an input sample. The probability that it belongs to the c-th feature running state cluster. The calculation method is as follows:
[0045] in: The logits are the linear outputs (logits) of the last layer of the neural network corresponding to the values of the c-th cluster. This indicates that the model output layer corresponds to the original output value of the j-th feature running state cluster. It is the total number of characteristic running state clusters. It is an exponential function. Optionally, the preset communication interface uses an industrial fieldbus protocol, such as PROFIBUS or Modbus, and the encapsulated control commands conform to the message specifications of the corresponding protocol. Optionally, after receiving a new control command, the underlying controller will adjust the actuator to the target setpoint in a gradual, gradient manner to avoid excessive impact on the system. It can be understood that the real-time identification capability of the operating condition identification model is a prerequisite for realizing adaptive optimization control, enabling the system to perceive its current operating state. It can be understood that the entire process from querying the optimization control strategy library to issuing control commands needs to be completed within seconds or less to meet the real-time control requirements of the desulfurization system.
[0046] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A data-driven based multi-objective optimization method for desulfurization system, characterized in that, include: Obtain a set of multi-source operating parameters of the target desulfurization system within a preset time period. The set of multi-source operating parameters includes a set of inlet flue gas parameters, a set of absorbent parameters, a set of process control parameters, and a set of outlet emission parameters. Data cleaning and alignment processing is performed on the multi-source operating parameter set to generate time-aligned multi-dimensional operating status time-series data; An improved fuzzy clustering algorithm is applied to the multidimensional operating state time series data to identify several characteristic operating state clusters that characterize typical operating conditions of the system. The improved fuzzy clustering algorithm optimizes the cluster center update mechanism based on the time-varying coupling relationship of system operating parameters. Extract the centroid vector and boundary features of each characteristic operating state cluster, and combine them with the preset optimization target set of the target desulfurization system. Then, use a multi-objective evolutionary algorithm to search in parallel for the optimal combination of control parameters under each characteristic operating state cluster. Based on the optimal control parameter combination obtained by searching under each characteristic operating state cluster, an optimized control strategy library for operating conditions is constructed. The multidimensional operating status time series data is input into the trained operating condition recognition model in real time to obtain the feature operating status cluster identifier corresponding to the current operating condition. Based on the characteristic operating state cluster identifier, query the operating condition adaptive optimization control strategy library to obtain the optimal control parameter combination corresponding to the current operating condition, and send it to the underlying controller of the target desulfurization system for execution; An improved fuzzy clustering algorithm is applied to the multidimensional operating state time series data to identify several characteristic operating state clusters representing typical system operating conditions, including: Initialize a set of random cluster centers, and assign a membership degree to each cluster center to each data point in the multidimensional running state time series data; The improved fuzzy clustering algorithm iteratively updates cluster centers and membership degrees. Each time the cluster centers are updated, the improved fuzzy clustering algorithm considers not only the spatial distribution of data points but also introduces temporal correlation weights. The calculation process of the temporal correlation weights includes: calculating the Euclidean distance between the current data point and its immediate neighbors in the time series; taking the reciprocal of the distance and normalizing it to obtain a similarity coefficient; performing a weighted summation of the similarity coefficient sequence to obtain the degree of local stationarity; and then mapping the degree of local stationarity to a temporal correlation weight using a monotonically increasing function. When updating the cluster centers, the contribution value of each data point is multiplied by its corresponding temporal correlation weight. The algorithm is considered to have converged when the change in the location of the cluster centers is less than a preset threshold. Based on the converged cluster centers and the maximum membership degree corresponding to each data point, the multidimensional running state time series data is divided into clusters represented by the corresponding cluster centers, and each cluster is a feature running state cluster. The process involves extracting the centroid vector and boundary features of each characteristic operating state cluster, combining them with the preset optimization target set of the target desulfurization system, and using a multi-objective evolutionary algorithm to search in parallel for the optimal combination of control parameters under each characteristic operating state cluster, including: The preset set of optimization targets includes desulfurization efficiency, system energy consumption, and operating costs; For a cluster of characteristic operating states, its centroid vector is used as the typical operating condition parameter of the cluster of characteristic operating states; Using the adjustable variables in the set of process control parameters as decision variables, an initial population of candidate control parameter combinations is randomly generated within the feasible domain of the decision variables. The desulfurization efficiency, system energy consumption, and operating cost simulated by operating the target desulfurization system under the typical operating conditions are calculated using each candidate control parameter combination, and are used as the objective function value of the corresponding candidate control parameter combination. Based on the multi-objective evolutionary algorithm, the population is subjected to selection, crossover, and mutation operations to generate a new population, and the objective function value is repeatedly calculated to simultaneously optimize desulfurization efficiency, system energy consumption, and operating costs. When the evolutionary process reaches the termination condition, output the Pareto optimal solution set composed of all non-dominated solutions in the current population, and select a solution from the Pareto optimal solution set according to the preset preference as the optimal control parameter combination under the corresponding feature running state cluster. The optimal control parameter combination search process for all feature running state clusters is performed in parallel.
2. The data-driven multi-objective optimization method for a desulfurization system according to claim 1, characterized in that, Data cleaning and alignment processing is performed on the multi-source operating parameter set to generate time-aligned multi-dimensional operating status time-series data, including: Outlier detection and smoothing are performed on each parameter sequence in the inlet flue gas parameter set, absorbent parameter set, process control parameter set, and outlet emission parameter set, respectively. To ensure that all processed parameter sequences have the same sampling time interval; All parameter sequences are interpolated and aligned with time as the reference axis to ensure that at each identical sampling time point, there are parameter values from different sets, forming a synchronized multidimensional data point sequence; The synchronized multidimensional data point sequence is organized in chronological order to form the time-aligned multidimensional running status time series data.
3. The data-driven multi-objective optimization method for a desulfurization system according to claim 2, characterized in that, The improved fuzzy clustering algorithm introduces temporal correlation weights each time the cluster centers are updated, and the calculation process includes: For the current data point, calculate its Euclidean distance to several adjacent data points before and after it in the multidimensional running state time series data, and obtain a distance sequence composed of multiple Euclidean distance values; The calculated distance sequence is taken inversely and normalized to obtain a set of coefficients representing the similarity between the current data point and its neighboring data points; Calculate a weighted sum of similarity coefficient sequences, whereby the weighted sum measures the degree of local stationarity of the current data point in the time series. The weighted sum is mapped through a monotonically increasing function to obtain the final temporal correlation weight, which is used to amplify or reduce the contribution of the current data point in updating the cluster center.
4. The data-driven multi-objective optimization method for a desulfurization system according to claim 3, characterized in that, The optimal control parameter combination obtained by searching under each characteristic operating state cluster is used to construct an adaptive optimization control strategy library, including: Create a policy record for each feature running state cluster. The policy record shall at least include the identifier, centroid vector, boundary features and corresponding optimal control parameter combination of the corresponding feature running state cluster. All policy records corresponding to feature running state clusters are indexed and organized to form a mapping data structure with feature running state cluster identifiers as keys and corresponding policy records as values. The mapping relationship data structure is persistently stored to form a queryable library of adaptive optimization control strategies for the operating conditions.
5. The data-driven multi-objective optimization method for a desulfurization system according to claim 4, characterized in that, The multidimensional operating state time-series data is input into the trained operating condition recognition model in real time to obtain the feature operating state cluster identifiers corresponding to the current operating condition, including: The operating condition identification model is a classification model trained based on historical multidimensional operating state time series data and its labeled feature operating state cluster identifiers. The latest operating parameters of the target desulfurization system are collected in real time within a certain time window. After data cleaning and alignment, a current multi-dimensional operating status data sample is formed. Input the current multidimensional operating status data sample into the operating condition identification model; The operating condition identification model outputs a probability distribution vector, where each component of the probability distribution vector represents the probability that the current multidimensional operating state data sample belongs to each feature operating state cluster. The identifier of the feature operating state cluster with the highest probability is used as the feature operating state cluster identifier corresponding to the current operating condition.
6. The data-driven multi-objective optimization method for a desulfurization system according to claim 5, characterized in that, Based on the characteristic operating state cluster identifier, the system queries the adaptive optimization control strategy library to obtain the optimal control parameter combination corresponding to the current operating condition, and then sends it to the underlying controller of the target desulfurization system for execution, including: Using the characteristic operating state cluster identifier as the query key, a search is performed in the operating condition adaptive optimization control strategy library; Obtain the retrieved strategy records and extract the optimal combination of control parameters from them; The optimal control parameter combination is encapsulated according to an instruction format recognizable by the underlying controller; Through a preset communication interface, the encapsulated control commands are sent to the underlying controller of the target desulfurization system. The underlying controller parses the commands and adjusts the corresponding actuators so that the system operating parameters approach the optimal combination of control parameters.
7. The data-driven multi-objective optimization method for a desulfurization system according to claim 6, characterized in that, The calculation uses the desulfurization efficiency, system energy consumption, and operating cost simulated by the target desulfurization system under the typical operating parameters for each candidate control parameter combination, as the objective function value for the corresponding candidate control parameter combination, including: Based on the uncontrollable variables in the typical operating condition parameters and the controllable variables in the current candidate control parameter combination, a complete input vector is constructed to describe the operating state of the target desulfurization system. The complete input vector is input into a pre-trained desulfurization system performance proxy model; Through forward calculation of the desulfurization system performance proxy model, the predicted desulfurization efficiency, predicted system energy consumption and predicted operating cost of the target desulfurization system are obtained under the system operating state described by the complete input vector. The values of predicted desulfurization efficiency, predicted system energy consumption, and predicted operating cost are extracted from the output of the desulfurization system performance proxy model. The extracted predicted desulfurization efficiency is negatively evaluated as the first objective function value to be minimized, the predicted system energy consumption is evaluated as the second objective function value to be minimized, and the predicted operating cost is evaluated as the third objective function value to be minimized, thus completing the calculation of the objective function value for the current candidate control parameter combination.
8. The data-driven multi-objective optimization method for a desulfurization system according to claim 7, characterized in that, When the evolutionary process reaches the termination condition, the Pareto optimal solution set consisting of all non-dominated solutions in the current population is output. From this Pareto optimal solution set, a solution is selected according to a preset preference as the optimal control parameter combination under the corresponding feature running state cluster, including: Calculate and output all non-dominated solutions in the current population at the end of the evolutionary process, forming a Pareto optimal solution set; From the Pareto optimal solution set, extract the three objective function values of desulfurization efficiency, system energy consumption and operating cost corresponding to each nondominated solution; The objective function value vector of each non-dominated solution is input into a preset preference decision model. The preset preference decision model is based on the actual operating constraints and priorities of the target desulfurization system and pre-sets preference weight vectors for the three optimization objectives of desulfurization efficiency, system energy consumption, and operating cost. The preset preference decision model calculates the weighted sum of the objective function value vector and the preset preference weight vector for each non-dominated solution, and normalizes the weighted sum into a comprehensive preference score. The comprehensive preference scores of all non-dominated solutions are sorted in descending order, and the non-dominated solution with the highest comprehensive preference score is selected as the optimal control parameter combination under the corresponding characteristic operating state cluster.