A data-driven wet flue gas desulfurization pump set optimization method and system

Abstract

The application discloses a data-driven wet desulfurization pump group optimization method and system, which comprises the following steps: obtaining historical operation data and current operation data of a desulfurization system, and extracting steady-state operation data segments from the historical operation data; the operation data comprises operation parameters and pump group operation state information; based on the steady-state operation data segments, a clustering algorithm is used to divide a working condition characteristic space composed of multiple historical operation parameters, form multiple working condition regions, and assign working condition labels; the current operation data is matched with the working condition regions to identify a current working condition label; based on the current working condition label and the current pump group operation state information, a reinforcement learning model is used to determine at least one candidate pump combination action from a preset action space; the future change trend of the slurry pH value after the execution of each candidate pump combination action is predicted; each candidate pump combination action is safety checked according to the future change trend of the pH value, and a recommended pump combination action is determined based on the checking result.

Description

Technical Field

[0001] This invention relates to the field of wet desulfurization technology, and more specifically to a data-driven method and system for optimizing wet desulfurization pump sets. Background Technology

[0002] Against the backdrop of environmental upgrades and the advancement of "dual carbon" targets in the power industry, limestone-gypsum wet flue gas desulfurization (FGD) technology is the mainstream technology for tail gas desulfurization in large coal-fired power units. As the core power-consuming equipment in the desulfurization system, the slurry circulation pump accounts for more than 50% of the total power consumption of the desulfurization system, making it a key target for energy conservation and emission reduction.

[0003] Currently, the operation and management of desulfurization systems in domestic thermal power plants are still mainly based on traditional manual operation mode, with a low level of intelligence. Pump combination adjustment relies on manual experience, which has obvious lag and subjectivity. It cannot match subtle changes in operating conditions in real time, often resulting in insufficient output under high operating conditions leading to excessive emissions, or excessive operation under low operating conditions causing energy waste. Moreover, it is difficult to balance the dual goals of desulfurization compliance and energy consumption optimization. Often, energy economy is sacrificed in order to ensure environmental compliance, resulting in a large amount of energy waste. Summary of the Invention

[0004] This invention provides a data-driven method and system for optimizing wet desulfurization pump sets, in order to solve the technical problem in the prior art that it is impossible to match subtle changes in operating conditions in real time and that it is difficult to achieve both desulfurization compliance and energy consumption optimization goals.

[0005] To address the aforementioned technical problems, in a first aspect, the present invention provides a data-driven optimization method for wet desulfurization pump sets, comprising: The historical and current operating data of the desulfurization system are acquired, and the historical operating data is preprocessed and a steady-state operating data segment is extracted; wherein, the operating data includes multiple operating parameters and pump group operating status information used to reflect changes in operating conditions; Based on the steady-state operation data segment, a clustering algorithm is used to divide the operating condition feature space composed of multiple historical operating parameters into multiple operating condition regions, and a unique operating condition label is assigned to each operating condition region. The current operating data is matched with the operating condition area to identify the current operating condition label; Based on the current operating condition label and the current pump group operating status information, a pre-trained reinforcement learning model is used to determine at least one candidate pump combination action from the preset action space according to the principle of maximizing cumulative reward. Each action in the action space corresponds to a pump combination configuration. Predict the future trend of slurry pH value after each of the candidate pump combinations performs its operation; The safety of the corresponding candidate pump combination actions is verified based on the future change trend of each pH value, and a recommended pump combination action is determined from the at least one candidate pump combination action based on the verification result, and its execution is controlled.

[0006] Secondly, a data-driven wet desulfurization pump set optimization system is provided, including: The data acquisition and preprocessing module is used to acquire historical and current operating data of the desulfurization system, preprocess the historical operating data, and extract steady-state operating data segments; wherein, the operating data includes multiple operating parameters and pump group operating status information to reflect changes in operating conditions; The working condition clustering and identification module is used to divide the working condition feature space composed of multiple historical operating parameters based on the steady-state operating data segment using a clustering algorithm, forming multiple working condition regions, and assigning a unique working condition label to each working condition region. The identification and judgment module is used to match the current operating data with the operating condition area, identify the current operating condition label, and determine whether the current operating mode of the desulfurization system is a steady-state operating mode or a rapidly changing operating mode based on the current operating data. The reinforcement learning recommendation module has a built-in pre-trained reinforcement learning model, which is used to determine at least one candidate pump combination action from a preset action space based on the current working condition label and the current pump group operating status information in steady-state operation mode, according to the principle of maximizing cumulative reward. Each action in the action space corresponds to a pump combination configuration. The rapid response decision module is used to determine at least one candidate pump combination action based on the urgency of emission compliance and the trend of changing operating conditions by adopting preset risk assessment and action decision rules in rapidly changing operating modes. The pH prediction module is used to predict the future trend of slurry pH value after the execution of each candidate pump combination action. The engineering constraint and decision module is used to perform safety verification on the corresponding candidate pump combination actions according to the future change trend of each pH value, and to determine the recommended pump combination action from the at least one candidate pump combination action based on the verification result, and control its execution. The incremental training module is used to perform non-intrusive hot updates on the working condition regions divided by the clustering algorithm, the Q-table of the reinforcement learning model, and the prediction model using newly added steady-state operating data, and retains historical model versions for fault rollback during the update process.

[0007] Compared with existing technologies, the data-driven wet desulfurization pump group optimization method of this invention divides the operating condition feature space of steady-state data segments using a clustering algorithm and establishes a fixed operating condition label system. This enables reliable and accurate identification of complex dynamic operating conditions of the desulfurization system, allowing pump combination adjustments to be highly adaptable to real-time operating conditions. This fundamentally avoids emission exceedances or energy waste caused by operating condition mismatches. Furthermore, it combines a reinforcement learning model to output candidate pump combination actions in real time and predicts the future change trend of slurry pH value after the execution of each candidate pump combination action. Based on the future change trend of each pH value, it performs a forward-looking safety check on the corresponding candidate pump combination actions before execution, constructing a collaborative mechanism of decision-making-prediction-checking. By embedding pH risk constraints into the pump combination optimization process, it ensures that energy-saving and consumption-reducing actions do not sacrifice desulfurization efficiency. This achieves a synergistic balance between energy-saving goals and desulfurization stability, as well as a dual guarantee of environmental compliance and optimal energy consumption, thereby improving the economy, stability, and intelligence level of the desulfurization system. Attached Figure Description

[0008] Figure 1 This is a schematic flowchart of a data-driven wet desulfurization pump set optimization method according to the present invention.

[0009] Figure 2 yes Figure 1 The diagram shows the specific process flow of S20 in the data-driven wet desulfurization pump set optimization method.

[0010] Figure 3 yes Figure 1 The diagram shows the specific process flow of S50 in the data-driven wet desulfurization pump set optimization method.

[0011] Figure 4 yes Figure 1 The diagram shows the specific process flow of S60 in the data-driven wet desulfurization pump set optimization method.

[0012] Figure 5 yes Figure 1 The diagram shows the specific process flow of S80 in the data-driven wet desulfurization pump set optimization method.

[0013] Figure 6 This is a schematic diagram of the structure of a data-driven wet desulfurization pump group optimization system according to the present invention. Detailed Implementation

[0014] To enable those skilled in the art to more clearly understand the purpose, technical solution, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0015] Reference Figure 1 , Figure 1This is a schematic flowchart of the data-driven wet desulfurization pump set optimization method of the present invention. In the embodiment shown in the figure, the data-driven wet desulfurization pump set optimization method includes: S10. Obtain historical and current operating data of the desulfurization system, preprocess the historical operating data, and extract the steady-state operating data segment.

[0016] In this invention, the operating data includes multiple operating parameters and pump group operating status information reflecting changes in operating conditions. The operating parameters include those reflecting boiler and flue gas conditions and those reflecting the current chemical state of the absorber and APT towers. The boiler and flue gas operating parameters may include unit load, raw flue gas sulfur dioxide concentration, total coal quantity, raw flue gas flow rate, net flue gas sulfur dioxide concentration, and liquid-to-gas ratio. The tower operating parameters may include the current pH value (absorber pH value and APT tower pH value) and liquid level within the tower. The pump group operating status information includes the start / stop status and cumulative operating time of each slurry circulation pump. In this embodiment, the desulfurization system can adopt a limestone-gypsum wet flue gas desulfurization process, including a two-stage desulfurization reaction device of an absorber and an APT tower. It can be configured with 7 slurry circulation pumps, with 4 pumps configured for the absorber (numbered P1 to P4) and 3 pumps configured for the APT tower (numbered P5 to P7). Each pump has an independent start / stop status and cumulative operating time. The operating status of the seven circulating pumps is represented by a 7-bit binary string code. The first 4 bits correspond to the absorption tower pump group (P1-P4), and the last 3 bits correspond to the APT tower pump group (P5-P7). A value of 1 for each bit indicates that the pump is running, and 0 indicates that the pump is stopped. For example, if P1, P3, P4, and P6 are running, the pump group operating status code is "1-0-1-1-0-1-0".

[0017] In this embodiment, the steady-state operation data segment refers to a continuous data segment in which the pump unit's operating state remains unchanged within the sliding time window and the fluctuation of key operating parameters is within a preset threshold. The preprocessing of the historical operation data and extraction of the steady-state operation data segment may include: using the 3σ criterion to remove abnormal data points from the historical operation data; and extracting from the historical operation data after removing abnormal data a continuous data segment in which the pump unit's operating state remains unchanged within the sliding time window (i.e., the start-stop status of all slurry circulation pumps remains unchanged) and the fluctuation of key operating parameters is within a preset threshold (e.g., load fluctuation ≤ 5%, flue gas flow fluctuation ≤ 8%), as the steady-state operation data segment. Through this step, transition state and wave dynamic data are effectively filtered, providing high-quality samples for subsequent operating condition clustering and ensuring the reliability of the operating condition analysis.

[0018] Specifically, the 3σ criterion, also known as the Raida criterion, calculates the mean μ and standard deviation σ of each operating parameter, and marks data points outside the range of [μ-3σ, μ+3σ] as outliers and removes them. For example, if the mean load of the unit under a certain operating condition is approximately 600MW and the standard deviation is approximately 20MW, then data points with loads below 540MW or above 660MW will be considered outliers and removed.

[0019] S20. Based on the steady-state operation data segment, a clustering algorithm is used to divide the operating condition feature space composed of multiple historical operating parameters into multiple operating condition regions, and a unique operating condition label is assigned to each operating condition region.

[0020] like Figure 2 As shown, step S20 specifically includes the following steps S21-S24: S21. Extract the multi-dimensional operating condition feature space composed of the multiple historical operating parameters from the steady-state operating data segment.

[0021] In this step, a multi-dimensional operating condition feature space is extracted from the steady-state operation data segment, consisting of multiple operating parameter features such as unit load and raw flue gas SO2 concentration. Each operating parameter feature serves as a different dimension of the operating condition feature space, collectively forming the multi-dimensional operating condition feature space. The steady-state operation data segment comprises multiple operation data records, with each record serving as a sample.

[0022] S22. Discretize the multidimensional working condition feature space into multiple regular grids, with each grid serving as a micro-working condition unit. Statistically analyze historical operating data samples within each grid and pre-assign a data unit to each grid. Temporary tag.

[0023] In this step, the continuous operating condition feature space, composed of multiple operating parameter features in the steady-state operating data segment of historical operating data, is discretized into multiple micro-operating condition grids. In this embodiment, the value range of each dimension is divided into several intervals according to a preset step size, and the step size of each dimension is independently set to 5%~10% of the value range of that dimension. For example, if the unit load range is 0~1000MW and the step size is set to 50MW, then this dimension is divided into 20 intervals; the original flue gas SO2 concentration range is 0~5000mg / Nm³. 3 The step size was set to 250 mg / Nm 3This dimension is then divided into 20 intervals. For each operational data record in the steady-state operation data segment, based on the actual values ​​of its unit load, raw flue gas SO2 concentration, and other operational parameters, the interval index for each dimension is determined. The combination of all dimension interval indices uniquely determines a grid, and the operational data record is mapped to that grid. When the values ​​of each dimension of multiple operational data records fall within the same interval range, thus having the exact same combination of dimension interval indices, these operational data records are mapped to the same grid. For example, sample A has a unit load of 623MW and a raw flue gas SO2 concentration of 1840mg / Nm³. 3 Sample B has a unit load of 638MW and a raw flue gas SO2 concentration of 1920mg / Nm³. 3 The unit loads of both samples fell into the 13th range (600~650MW), and the SO2 concentrations in the raw flue gas both fell into the 8th range (1750~2000mg / Nm³). 3 If the dimensional interval index combination of the two samples is (13, 8), they are both mapped to the grid (13, 8). All steady-state operating data samples are mapped to the corresponding grid according to this rule. After mapping, the number of samples falling into each grid, the liquid-to-gas ratio statistics, and the dominant pump group structure format are statistically analyzed. Among them, the liquid-to-gas ratio statistics are calculated using the trimmed mean, and the dominant pump group structure format is the pump combination configuration with the highest frequency in that grid. For example, "3+2" indicates that the absorption tower has 3 pumps operating and the APT tower has 2 pumps operating. A temporary label is pre-assigned to each grid. In this way, the continuous multidimensional operating condition feature space is discretized into a finite number of regular grids. Each grid is a micro-operating condition unit, providing a basis for subsequent merging of adjacent grids and division of operating condition regions.

[0024] S23. Based on the topological adjacency relationship between grids, and utilizing preset sample size thresholds, liquid-to-gas ratio similarity, liquid-to-gas ratio zone consistency, and pump group structure consistency conditions, historical operating data samples that meet the conditions are... Adjacent grids are clustered and merged to form multiple connected working areas.

[0025] In this embodiment, based on the four-adjacent topological adjacency relationship between grids, a disjoint-set data structure is used to determine whether any two adjacent regular grids meet the following business indicator similarity rules: (a) the number of samples on both sides is not less than the preset minimum sample size threshold (e.g., 3); (b) the absolute difference between the liquid-to-gas ratio statistics of both sides does not exceed the preset threshold (e.g., 2); (c) both sides belong to the same liquid-to-gas ratio zone (the zone step size is, for example, 3); (d) the dominant pump group structure format of both sides is consistent, and their respective dominant proportion is not less than 50%. Adjacent grids that meet all the rule conditions are merged and connected through the disjoint-set method to form a connected working condition region. That is, based on the four-adjacent topological adjacency relationship between grids, for any two adjacent specification grids, under the conditions that the number of samples on both sides is not less than the preset minimum sample size threshold, the difference between the liquid-to-gas ratio statistics does not exceed the preset threshold, they belong to the same liquid-to-gas ratio zone, and the dominant pump group structure format is consistent, the two are merged, and the adjacent grids that meet the conditions are connected through the disjoint-set method to form a connected working condition region. The connected working condition region is a working condition category obtained by clustering.

[0026] S24. For each connected operating condition region, the pre-assigned temporary label corresponding to the grid with the largest number of samples in that region is used as the final operating condition label for that region.

[0027] Furthermore, in this invention, an independent grid that fails to merge with any adjacent grid, if its sample count is not zero, is retained as an independent operating condition region and its pre-assigned temporary label is used as the final operating condition label. The final operating condition label remains fixed during subsequent incremental updates, ensuring the consistency and continuity of the operating condition labels.

[0028] To address the above steps, the continuous operating condition feature space of the steady-state operating data segment in historical operating data is discretized into a micro-operating condition grid. The grids are merged by combining topological adjacency and business indicator similarity rules to form physically meaningful connected operating condition regions and assign fixed labels. This not only achieves the scientific division of complex operating states but also ensures the consistency and continuity of operating condition labels, enabling precise matching of real-time operating conditions. This ensures that the pump set optimization strategy is highly adapted to the current operating conditions, effectively avoiding the problems of unstable desulfurization efficiency and energy waste caused by operating condition mismatch, and making pump set adjustments more targeted and scientific.

[0029] S30. Match the current operating data with the operating condition area to identify the current operating condition label.

[0030] This step may include: calculating the Euclidean distance between the operating parameters in the current operating data and the cluster centers of each connected operating condition region, and taking the operating condition label with the smallest distance that is within the distribution range of the operating condition region as the current operating condition label; if no matching operating condition region is found, then selecting the operating condition label of the operating condition region with the smallest Euclidean distance as the current operating condition label.

[0031] Specifically, the operating parameters (unit load, raw flue gas SO2 concentration, flue gas flow rate, etc.) in the current operating data are mapped to the operating condition feature space. The Euclidean distance between the data point and the cluster centers of each known connected operating condition area is calculated. The operating condition label with the smallest distance and where the data point falls within the distribution range of that operating condition area is selected as the current operating condition label. If the real-time data point fails to match any known operating condition area (e.g., under an extreme load condition), the similar operating condition matching logic is activated, and the operating condition label of the operating condition area with the smallest Euclidean distance is selected as the alternative current operating condition label.

[0032] S40. Determine the current operating mode of the desulfurization system based on the current operating data. If the current operating mode is a steady state, proceed to step S50. If the current operating mode is a rapidly changing operating mode, proceed to step S60.

[0033] In this step, the judgment can be made based on the operating condition labels within multiple consecutive sliding time windows corresponding to the current operating data. If more than half of the windows are identified as the same operating condition label within 10 consecutive windows (e.g., about 10 minutes), and the current operating parameters do not show a significant trend change, then it is determined that the current operation is in a steady-state mode. If at the same time, a clear and continuous upward trend is detected in the coal quantity and raw flue gas flow rate (e.g., an increase of more than 10% within 5 consecutive minutes), or the net flue gas SO2 concentration is close to the emission exceedance threshold (e.g., less than 15% away from the exceedance limit), then it is determined that the current operation is in a rapid change mode.

[0034] In this embodiment, different recommendation strategies are adopted based on the determined current operating mode, i.e., a dual-mode decision-making mechanism is designed. This mechanism can automatically identify steady-state and rapidly changing operating conditions and adopt optimization-first and safety-first strategies respectively, avoiding the contradiction between response lag and over-adjustment in traditional control. The decision-making mechanisms in the two modes are independent of each other. The steady-state operating mode uses a reinforcement learning model, while the rapidly changing operating mode uses risk assessment and action decision rules.

[0035] S50. Based on the current operating condition label and the current pump group operating status information, using a pre-trained reinforcement learning model, at least one candidate pump combination action is determined from the preset action space according to the principle of maximizing cumulative reward. Each action in the action space corresponds to a pump combination configuration.

[0036] In this invention, the preset action space includes all feasible pump combination configurations for engineering. Each action corresponds to one pump combination configuration, and each pump combination configuration is composed of the start-stop states of each slurry circulation pump, represented by the start-stop state codes of each slurry circulation pump. For example, taking a desulfurization system with a total of 7 slurry circulation pumps as an example, the action space excludes all combinations where all pumps are stopped, resulting in a total of 127 selectable actions. Furthermore, a preset time (e.g., 5 minutes) pump start-stop cooling time constraint can be set, prohibiting the same pump from being switched on and off again during the cooling time.

[0037] like Figure 3 As shown, step S50 specifically includes the following steps S51-S53: S51. Construct the decision state of the current input to the pre-trained reinforcement learning model using the current operating condition label and the current pump group operating status information.

[0038] In this step, a decision state is constructed using the current operating condition label and the current pump group operating status information. Specifically, the current operating condition label, the start / stop status code of each slurry circulation pump in the pump group operating status, and the preset duration level after discretizing the cumulative running time of each slurry circulation pump are combined into a structured state string, which serves as the input to the decision state of the reinforcement learning model. The running time discretization rule can be: less than 2 hours is S (short time), 6 to 12 hours is M (medium), 2 to 6 hours and 12 to 24 hours is L (long time), 24 to 48 hours is K (relatively long time), and more than 48 hours is H (extremely long time). For example, if the current operating condition label is 5, the pump group status is "1-0-1-1-0-1-0" (P1, P3, P4, P6 are running), and the running time levels of each pump are S, L, M, K, L, H, and S respectively, then the decision state string is "5_1011010_S_L_M_K_L_H_S".

[0039] S52. Based on the current input decision state, query the Q value of each optional pump combination action from the Q table corresponding to the current operating condition label.

[0040] In this step, the Q table corresponding to the current working condition label is loaded, the constructed decision state is used as input, and the Q value of each optional action is queried from the Q table.

[0041] In this embodiment, the Q value represents the expected cumulative reward obtained by performing the corresponding pump combination action under the current input decision state. The Q value can be obtained by offline training based on a multi-factor weighted reward function. The reward factors of the multi-factor weighted reward function include a power factor for encouraging energy saving, a load balancing factor for encouraging load balancing, a switching factor for limiting the number of pump start-ups and shutdowns, and an emission factor for giving segmented rewards or penalties based on the emission concentration range. That is, the reward function of this reinforcement learning model is a multi-factor weighted summation structure, which includes the following four types of reward factors: power factor, which is used to characterize the proportion of the total current of the pump group after the action to the total rated current of all pumps. The lower the current proportion, the higher the reward, so as to encourage energy saving; balance factor, which is used to characterize the degree of difference in the running time of each pump. The smaller the difference in running time, the higher the reward, so as to encourage load balance and reduce the risk of equipment wear; switching factor, which is used to limit the number of pump group start-ups and shutdowns. No additional penalty is imposed within the penalty-free threshold (e.g., allowing switching of 1 pump). After exceeding the threshold, the reward is linearly reduced according to the excess part; emission factor, which is used to give segmented rewards or penalties according to the net flue gas SO2 concentration range. Positive rewards are given when the concentration meets the standard and is close to the center of the target range, and negative rewards are given when it exceeds the standard.

[0042] Based on the above design, the design of the multi-dimensional reward factor not only ensures that SO2 emissions stably meet environmental standards, but also minimizes energy consumption under the premise of desulfurization compliance. At the same time, the constraints on the balanced operation of pumps and the number of pump start-ups and shutdowns can guide the selection of a more reasonable pump combination adjustment strategy, making the running time of each pump group more balanced, effectively reducing the wear of pump groups, reducing the probability of equipment failure, thereby extending the service life of equipment, and also avoiding the current surge caused by frequent pump start-ups and shutdowns, thus reducing equipment maintenance costs.

[0043] S53. Using an ε-greedy strategy, at least one pump combination action is selected as the candidate pump combination action based on the Q value.

[0044] In this step, an ε-greedy strategy is adopted. An action is randomly selected for exploration with a preset probability ε. The top N pump combination actions are selected in descending order of their current Q values ​​with probabilities 1-ε, where N is an integer greater than or equal to 1. The pump combination configuration corresponding to the at least one pump combination action is then output as the at least one candidate pump combination action. In this embodiment, the top 3 pump combination configurations are selected as candidate pump combination actions in descending order of Q values, forming a candidate list. For example, if the current operating condition label is 5, the three pump combination configurations with the highest Q values ​​after querying the Q table are: Action A (Q=8.52) corresponding to pump combination configuration "1-0-1-1-0-1-0", Action B (Q=8.17) corresponding to pump combination configuration "1-1-0-1-0-1-0", and Action C (Q=7.93) corresponding to pump combination configuration "0-1-1-1-0-1-0".

[0045] Understandably, in this invention, the pump group optimization problem under steady-state operation mode is modeled as a Markov decision process. The decision process is based on a reinforcement learning Q-learning model, and the main optimization objectives for this operation phase are operational economy and equipment protection. Specifically, in this embodiment, the pre-trained reinforcement learning model uses the Q-learning algorithm and consists of four layers: a state space layer, an action space layer, a reward function layer, and a Q-table. The state space layer combines the current operating condition label, pump group start-stop state code, and discretized pump runtime levels into a structured state string, serving as a complete description of the desulfurization system's operating state. The action space layer defines all feasible pump combination configurations, with each action corresponding to a set of start-stop state codes for each slurry circulation pump, and includes pump start-stop cooling time constraints. The reward function layer uses a multi-factor weighted summation structure, integrating four types of reward factors: power factor, equilibrium factor, switching factor, and emission factor, to provide real-time quantitative evaluation of the action's effect from four dimensions: energy saving, equipment protection, operational stability, and emission compliance. The Q-table stores the expected cumulative reward value of each "state-action" pair in partitions according to the operating condition label. When the model is running, the decision state constructed with the current operating condition label and pump group operating status information is used as input. The Q value of each optional action is queried from the Q table corresponding to the current operating condition label. At least one action is selected in descending order of Q value using an ε-greedy strategy. The pump combination configuration corresponding to the selected action is output as the candidate pump combination action.

[0046] The reinforcement learning Q-learning model can be trained offline. It extracts state transition samples using historical steady-state operating data and iteratively updates the Q-table according to the Bellman equation until the Q-table converges. The purpose of its training is to enable the Q-value to accurately reflect the comprehensive cumulative reward that can be obtained by selecting different pump combinations under different states through continuous interactive learning with the desulfurization system's operating environment. This allows for multi-objective optimization of energy saving, balance, emission reduction, and stable operation during decision-making.

[0047] Based on the above steps, a multi-dimensional decision-making state space is constructed, including operating condition label ID, pump operating status, and cumulative pump operating time. A pump combination action space that meets engineering feasibility is designed, and a multi-dimensional reward mechanism integrating power factor, balancing factor, switching factor, and emission factor is established. Multiple objectives such as desulfurization compliance, energy consumption minimization, balanced equipment operation, and reduction of pump start-up and shutdown frequency are integrated into the reward function. Through continuous interaction with the environment, the optimal pump combination strategy is learned, and the optimal pump combination suggestions can be output in real time according to changes in operating conditions. The response speed is much faster than manual adjustment, and the decision-making process is more objective and scientific. It not only ensures that SO2 emissions stably meet environmental protection standards, but also minimizes energy consumption under the premise of desulfurization compliance. Actual operation verification shows that the unit desulfurization power consumption can be reduced by 6.4%, which significantly improves the operating economy of the desulfurization system.

[0048] S60. Using preset risk assessment and action decision rules, determine at least one candidate pump combination action based on the urgency of emission compliance and the trend of operating condition changes.

[0049] In this invention, under rapidly changing operating modes, a preset three-factor risk assessment and action decision-making rule is used for rapid response, which can prioritize ensuring emission safety.

[0050] like Figure 4 As shown, step S60 specifically includes the following steps S61-S64: S61. Evaluate the effect risk level, trend intensity level, and structural mismatch level based on the emission concentration, operating parameters, and current pump unit operating status information in the current operating data.

[0051] In this step, the effect risk level can be assessed based on the emission concentration and emission exceedance threshold in the current operating data. The effect risk level represents the urgency of achieving emission standards and can be divided into three levels: low, medium, and high. The trend intensity level can be assessed based on the rate of change of operating parameters reflecting changes in operating conditions (such as unit load and raw flue gas flow) in the current operating data within a preset time window. The trend intensity level represents the speed of deterioration of operating conditions. The structural mismatch level can be assessed based on the degree of deviation between the current pump set operating status and the historical reasonable pump combination operating status corresponding to the current operating condition label. The structural mismatch level represents the necessity of adjusting the current pump combination. Here, the historical reasonable pump combination refers to the pump combination configuration that appears most frequently in the historical steady-state data corresponding to the operating condition label.

[0052] S62. Determine the action bias based on the effect risk level and the trend intensity level, wherein the action bias is one of maintaining the current pump combination, switching the pump combination, or increasing the number of operating pumps.

[0053] In this step, the action bias is determined based on the effect risk level and trend strength level using preset decision rules. For example, when the effect risk is low and the trend strength is weak, the action bias is "maintain"; when the effect risk is high and the trend strength is strong, the action bias is "increase the pump".

[0054] S63. If the action is biased towards switching pump combinations, the switching priority is adjusted according to the structural mismatch level.

[0055] In this step, if the action is biased towards switching pump combinations, the priority of switching target pump combinations is adjusted according to the structural mismatch level; if the structural mismatch level is high, it indicates that the current pump combination action deviates significantly from the historical reasonable pump combination, and the priority of switching target pump combinations that can eliminate the structural mismatch is increased.

[0056] S64. Based on the action bias, perform action optimization in the union of the similar pump combination set and the historical preferred pump combination set, and determine at least one pump combination action as the candidate pump combination action.

[0057] In this step, if the action bias is to maintain the current pump combination or increase the number of operating pumps, then based on the action bias, action optimization is performed in the union of the similar pump combination set and the historical preferred pump combination set, and at least one pump combination is output as a candidate pump combination. If the action bias is to switch pump combinations, then based on the action bias and the corrected switching priority, at least one pump combination action is selected from the union of the similar pump combination set and the historical preferred pump combination set in descending order of priority as the candidate pump combination action.

[0058] In this step, the union of the similar pump combination set and the preferred pump combination set constitutes the candidate search space under this rapidly changing operating mode. The similar pump combination set refers to the set of all pump combination actions in the preset action space whose Hamming distance to the current pump combination action does not exceed a preset distance threshold (e.g., Hamming distance ≤ 2). The historical preferred pump combination set refers to the set of pump combination actions that have been used most frequently in the historical operating data corresponding to the current operating condition label (e.g., the top 5 pump combinations with the highest frequency).

[0059] S70. Predict the future trend of slurry pH value after the execution of each candidate pump combination action.

[0060] In this step, based on the pump combination configuration corresponding to each candidate pump combination action, the current operating parameters, and the slurry pH value at multiple time steps, a prediction model based on the XGBoost algorithm is used to predict the slurry pH value change trend within a preset time period after the execution of each candidate pump combination action. Specifically, in this embodiment, the input features of the prediction model may include unit load, raw flue gas sulfur dioxide concentration, current pH value of the absorber and APT tower, liquid-to-gas ratio, pump combination configuration corresponding to the candidate pump combination action (i.e., the start-stop status code of the 7 single pumps), temporal difference features of the pH value of the absorber and APT tower at 1, 3, and 5 time steps, as well as features such as total coal quantity, net flue gas SO2 concentration, absorber level, and APT tower level.

[0061] In this invention, the prediction model employs a regression model based on the XGBoost algorithm, consisting of three layers: an input layer, an ensemble learning layer, and an output layer. The input layer receives the aforementioned input feature data; the ensemble learning layer uses a gradient boosting strategy, employing a CART regression tree as the base learner, iteratively fitting the residuals of the previous tree to gradually build a strong learner, and introducing a regularization mechanism to control model complexity; the output layer outputs the trend of pH changes in the slurry of the absorber and APT towers within a preset future time period after the execution of the candidate pump combination. The training objective of the prediction model is to learn the mapping relationship between input features and future pH changes, enabling the model to accurately predict the impact of different candidate pump combinations on the desulfurization reaction efficiency, providing a forward-looking safety verification basis for pump combination decisions. During the training phase, a training sample set is constructed from historical steady-state operating data. Each sample uses the pump combination configuration, operating parameters, and slurry pH values ​​at multiple time steps as input, and the difference between the current pH value and the pH value after a preset future time step as the regression target. The training and validation sets are divided in an 8:2 ratio. The hyperparameter combination of XGBoost is optimized using grid search. Training is completed when the mean absolute error on the validation set falls below a preset threshold. After training, during runtime, for each candidate pump combination action, its corresponding pump combination configuration is decoded into individual pump start / stop state features. These features, along with the operating parameters from the current running data and the temporal difference features, serve as model input. Through forward inference via the ensemble learning layer, the predicted pH value for that candidate pump combination is output. Understandably, the standardized parameters (mean and standard deviation) required by the model during the inference phase are saved and reused during incremental training and online prediction to ensure the consistency of the input distribution.

[0062] S80. Perform a safety check on the corresponding candidate pump combination actions based on the future change trend of each pH value, and determine the recommended pump combination action from the at least one candidate pump combination action based on the check result, and control its execution.

[0063] In this step, the safety verification of each candidate pump combination action is performed sequentially according to the Q value of the at least one candidate pump combination action or the priority in the rapidly changing operating mode, from high to low. In this invention, pH risk constraints are embedded in the pump group optimization process to perform forward-looking safety verification, ensuring that energy-saving and consumption-reducing actions do not sacrifice desulfurization efficiency, and that a synergistic balance between energy-saving goals and desulfurization stability can be achieved.

[0064] like Figure 5 As shown, the step S80, which involves determining the recommended pump combination action from the at least one candidate pump combination action based on the verification result, specifically includes the following steps S81-S83: S81. The future trend of pH value corresponding to each candidate pump combination action is compared with the preset slurry pH value process range.

[0065] S82. If the future trend of the pH value corresponding to the candidate pump combination action being checked remains stable within the preset slurry pH value process range, then the check is passed, and the candidate pump combination action is determined as the recommended pump combination action.

[0066] Specifically, if the predicted slurry pH value will fall below the lower limit of the preset slurry pH value process range, or if the predicted slurry pH value approaches the lower limit and the safety margin is insufficient, then the candidate pump combination action fails the safety check, and the next candidate pump combination action is checked. However, if the predicted slurry pH value remains stable within the preset slurry pH value process range, then the candidate pump combination action passes the safety check, and the candidate pump combination is determined as the recommended pump combination action, and subsequent candidate pump combination actions are no longer checked. For example, in steady-state operation mode, the candidate pump combination action with the highest Q value output by the reinforcement learning model is the pump reduction scheme, which reduces the number of operating pumps in the absorber from 4 to 3. The prediction model predicts the pH value after the execution of this candidate pump combination action. If the predicted pH value of the absorber will drop to 4.8 (falling below the lower limit of 5.0), the verification fails, and the second-best candidate pump combination action is verified. If the second-best candidate combination predicts that the pH value of the absorber is 5.08, although it has not fallen below the lower limit, the safety margin is small (only 0.08 away from the lower limit), so the verification still fails, and the third candidate pump combination action is verified.

[0067] S83. If all candidate pump combination actions fail the safety check, then the current pump combination action is determined to be the recommended pump combination action.

[0068] Furthermore, in some embodiments, after determining the recommended pump combination action from the at least one candidate pump combination action based on the verification result, the method further includes: if the difference in the number of operating pumps between the recommended pump combination action and the current pump combination action exceeds a preset number threshold, enabling a step-by-step approximation strategy, determining at least one transition pump combination configuration according to the recommended pump combination action, and controlling the slurry circulation pump group to sequentially execute each of the transition pump combination configurations until the recommended pump combination action is finally executed; wherein, the change in the number of operating pumps in each pump group adjustment does not exceed the preset number threshold. Based on the above design, in order to avoid the impact of drastic changes in the pump group's operating status on the desulfurization system, a step-by-step approximation strategy is adopted to complete a smooth transition from the current pump combination action to the recommended pump combination action. For example, if the current pump combination action is "1-1-1-1-1-1-0" (6 pumps are running) and the recommended pump combination action is "1-0-1-0-1-0-0" (3 pumps are running), the difference is 3 pumps, which exceeds the threshold of 2 pumps. A transition pump combination "1-1-1-0-1-0-0" (4 pumps are running) will be output first. After execution and stabilization, the recommended pump combination action will be output.

[0069] S90. Detect whether the preset incremental update triggering condition is met. When the incremental update triggering condition is met, use the newly added steady-state operating data to perform non-intrusive hot updates on the working condition regions divided by the clustering algorithm, the Q-table of the reinforcement learning model, and the prediction model, and retain the historical model version for fault rollback during the update process.

[0070] In this step, the incremental update triggering conditions include: the time interval since the last incremental update reaches a preset update cycle, and the number of samples of newly added steady-state operating data accumulated since the last incremental update reaches a preset minimum data volume threshold.

[0071] In this invention, a periodic update triggering mode is adopted. When the update conditions are met and sufficient new steady-state operating data is accumulated, a non-intrusive hot update process is triggered: First, the samples of the new steady-state operating data are mapped to the existing grid, merged with the historical grid statistical results, and the adjacent grid merging, working condition region division, and working condition center calculation are re-executed to obtain the updated working condition labels and cluster centers; then, based on the updated working condition labels, the existing Q-table of the reinforcement learning model is incrementally corrected with small steps using the new samples, and a small fixed incremental learning rate is used in conjunction with historical Q-values ​​for smooth updates; simultaneously, the prediction model is continued to be trained using the new samples while reusing the original standardized parameters. After the update is completed, the new model is pushed to the online service through hot loading, and the backup of the previous version model is retained for fault rollback.

[0072] Furthermore, in this invention, the data-driven wet desulfurization pump group optimization method may also include: presenting the recommended pump group operation, real-time status of each pump, current operating mode, pH prediction curve, desulfurization cost statistics and / or three-factor evaluation results in a visually intuitive form in the form of charts, and if SO2 concentration exceeds the standard, pH prediction error continues to be greater than 0.1 or data acquisition is abnormal, an alarm will be triggered immediately and a log will be recorded to prompt maintenance personnel to pay attention and intervene.

[0073] As can be seen, the data-driven wet desulfurization pump group optimization method of this invention divides the operating condition feature space of the steady-state data segment through clustering algorithm and establishes a fixed operating condition label system, realizing reliable and accurate identification of complex dynamic operating conditions of the desulfurization system. This enables pump combination adjustment to be highly adaptable to real-time operating conditions, fundamentally avoiding emission exceedances or energy waste caused by operating condition mismatch. Combined with a reinforcement learning model, candidate pump combinations are output in real time, and the future change trend of slurry pH value after the execution of each candidate pump combination action is predicted. Based on the future change trend of each pH value, a forward-looking safety check is performed on the corresponding candidate pump combination action before execution, constructing a collaborative mechanism of decision-making-prediction-check. The pH risk constraint is embedded in the pump combination optimization process, ensuring that energy-saving and consumption-reducing actions do not sacrifice desulfurization efficiency. At the same time, a "initial training + incremental training" model is also constructed. The non-intrusive model update mechanism enables the optimization strategy to continuously evolve with equipment operation, overcoming the poor adaptability of traditional fixed strategies and effectively coping with new operating conditions. Even in the face of special operating conditions such as extreme loads, it can make effective control responses through similar operating condition matching logic, ensuring the stable operation of the desulfurization system under all operating conditions. It can also provide intuitive decision-making suggestions for operators, reducing the difficulty and cost of operation and maintenance. It can be seen that, through the above closed-loop process, this invention realizes the transformation of the desulfurization system from manual experience control to data-driven intelligent optimization control. Under the premise of ensuring emission compliance, it reduces energy waste, effectively balances equipment wear, and extends pump life. That is, it achieves a synergistic balance between energy saving goals and desulfurization stability, as well as a dual guarantee of environmental compliance and optimal energy consumption, which can improve the economy, stability and intelligence level of desulfurization system operation.

[0074] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0075] Reference Figure 6 , Figure 6 This is a schematic block diagram of the data-driven wet desulfurization pump set optimization system of the present invention. In the embodiment shown in the figure, the data-driven wet desulfurization pump set optimization system includes: The data acquisition and preprocessing module 110 is used to acquire historical and current operating data of the desulfurization system, preprocess the historical operating data, and extract steady-state operating data segments; wherein, the operating data includes multiple operating parameters and pump group operating status information used to reflect changes in operating conditions; The working condition clustering and identification module 120 is used to divide the working condition feature space composed of multiple historical operating parameters based on the steady-state operating data segment using a clustering algorithm to form multiple working condition regions, and assign a unique working condition label to each working condition region. The identification and judgment module 130 is used to match the current operating data with the operating condition area, identify the current operating condition label, and determine whether the current operating mode of the desulfurization system is a steady-state operating mode or a rapidly changing operating mode based on the current operating data. The reinforcement learning recommendation module 140 has a built-in pre-trained reinforcement learning model, which is used to determine at least one candidate pump combination action from a preset action space based on the current working condition label and the current pump group operating status information in steady-state operation mode, according to the principle of maximizing cumulative reward. Each action in the action space corresponds to a pump combination configuration. The rapid response decision module 150 is used to determine at least one candidate pump combination action based on the urgency of emission compliance and the trend of operating condition changes in a rapidly changing operating mode by adopting preset risk assessment and action decision rules. pH prediction module 160 is used to predict the future trend of slurry pH value after each candidate pump combination action is performed. The engineering constraint and decision module 170 is used to perform safety verification on the corresponding candidate pump combination actions according to the future change trend of each pH value, and to determine the recommended pump combination action from the at least one candidate pump combination action based on the verification result, and control its execution. The incremental training module 180 is used to perform non-intrusive hot updates on the working condition regions divided by the clustering algorithm, the Q-table of the reinforcement learning model, and the prediction model using the newly added steady-state operating data when the incremental update triggering conditions are met, and to retain historical model versions for fault rollback during the update process. The incremental update triggering conditions include: the time interval since the last incremental update reaches a preset update cycle, and the number of samples of newly added steady-state operating data accumulated since the last incremental update reaches a preset minimum data volume threshold. In some embodiments, the working condition clustering and identification module 120 is specifically used for: Extract a multi-dimensional operating condition feature space composed of the multiple historical operating parameters from the steady-state operating data segment; The multidimensional working condition feature space is discretized into multiple regular grids, each grid being a micro-working condition unit. Historical operating data samples within each grid are statistically analyzed, and a temporary label is pre-assigned to each grid. Based on the topological adjacency relationship between grids, and using preset sample size thresholds, liquid-to-gas ratio similarity, liquid-to-gas ratio zone consistency, and pump group structure consistency conditions, adjacent grids that meet the conditions in historical operation data samples are clustered and merged to form multiple connected operating condition regions. For each connected operating condition region, the pre-assigned temporary label corresponding to the grid with the largest number of samples in that region is used as the final operating condition label for that region.

[0076] In some embodiments, the reinforcement learning recommendation module 140 is specifically used for: The decision state of the pre-trained reinforcement learning model is constructed using the current operating condition label and the current pump group operating status information. Based on the current input decision state, the Q value of each optional pump combination action is queried from the Q table corresponding to the current operating condition label; wherein, the Q value represents the expected cumulative reward obtained by performing the corresponding pump combination action under the current input decision state, and the expected cumulative reward is calculated by a multi-factor weighted reward function, wherein the reward factors of the multi-factor weighted reward function include a power factor for encouraging energy saving, a load balancing factor for encouraging load balancing, a switching factor for limiting the number of pump start-ups and shutdowns, and an emission factor for giving segmented rewards or penalties based on emission concentration ranges; An ε-greedy strategy is adopted to select at least one pump combination action as the candidate pump combination action based on the Q value.

[0077] In some embodiments, the rapid response decision module 150 is specifically used for: The effect risk level, trend intensity level, and structural mismatch level are assessed based on the emission concentration, operating parameters, and current pump unit operating status information in the current operating data, respectively. The effect risk level represents the urgency of achieving emission standards, the trend intensity level represents the rate of deterioration of operating conditions, and the structural mismatch level represents the necessity of adjusting the current pump unit operation. The action bias is determined based on the effect risk level and the trend intensity level. The action bias is one of maintaining the current pump combination, switching the pump combination, or increasing the number of operating pumps. If the action bias is to switch pump combinations, the switching priority of the pumps in the action bias is corrected according to the structural mismatch level; Based on the action bias, action optimization is performed in the union of the similar pump combination set and the historical preferred pump combination set to determine at least one pump combination action as the candidate pump combination action; wherein, the similar pump combination set refers to the set of all pump combination actions in the preset action space whose Hamming distance from the current pump combination action does not exceed a preset distance threshold, and the historical preferred pump combination set refers to the set of several pump combination actions that have been used the most times or have the highest operating effect evaluation in the historical operating data corresponding to the current operating condition label.

[0078] In some embodiments, the engineering constraint and decision module 170 is specifically used for: Based on maximizing the cumulative reward, the future trend of pH value corresponding to each candidate pump combination action is compared with the preset slurry pH value process range. If the future trend of the pH value corresponding to the candidate pump combination action being checked remains stable within the preset slurry pH value process range, then the check is passed, and the candidate pump combination action is determined as the recommended pump combination action. If all candidate pump combination actions fail the safety check, then the current pump combination action is determined to be the recommended pump combination action.

[0079] In some embodiments, the pH prediction module 160 is specifically used for: Based on the pump combination configuration corresponding to each candidate pump combination action, the current operating data, and the slurry pH value at multiple time steps, a prediction model based on the XGBoost algorithm is used to predict the slurry pH value change trend within a preset time period after each candidate pump combination action is executed.

[0080] In some embodiments, the engineering constraint and decision module 170 is further configured to: If the difference in the number of operating pumps between the recommended pump combination action and the current pump combination action exceeds a preset number threshold, a step-by-step approximation strategy is activated. At least one transition pump combination configuration is determined based on the recommended pump combination action, and the slurry circulation pump group is controlled to execute each of the transition pump combination configurations in sequence until the recommended pump combination action is finally executed. In this case, the change in the number of operating pumps in each pump group adjustment does not exceed the preset number threshold.

[0081] The data-driven wet desulfurization pump group optimization system provided by this invention employs gridded clustering and adjacent merging clustering algorithms to accurately identify and classify the operating conditions of the desulfurization system. This ensures that the pump group optimization strategy is highly adapted to the current operating conditions, effectively avoiding the problems of unstable desulfurization efficiency and energy waste caused by operating condition mismatch. It makes pump combination adjustments more targeted and scientific. Furthermore, it combines a reinforcement learning model to output candidate pump combinations in real time and uses a prediction model to predict the future trend of slurry pH value after the execution of each candidate pump combination action. Based on the predicted future pH value change trends, it also performs a forward-looking safety check on the corresponding candidate pump combination actions. Simultaneously, it constructs a non-intrusive model incremental update mechanism, enabling… The optimization strategy can continuously evolve with equipment operation, overcoming the poor adaptability of traditional fixed strategies and effectively responding to new operating conditions. It forms a closed-loop intelligent optimization system of state perception, decision optimization, effect prediction, and model update. The operating condition clustering and identification module provides the foundation for pump group optimization, the pH prediction module provides effect feedback for pump group optimization, and the incremental training module continuously optimizes the model based on the operating effect. The various modules form an efficient linkage mechanism, which can make precise adjustments to the pump group according to real-time changes in operating conditions. Then, based on the adjusted pH value changes, energy consumption data, and SO2 emission data, the control strategy is optimized in a timely manner, realizing continuous iterative upgrades and ensuring the stability and efficiency of the desulfurization system operation.

[0082] Specific limitations regarding the data-driven wet desulfurization pump set optimization system can be found in the limitations of the data-driven wet desulfurization pump set optimization method described above, and will not be repeated here. Each module in the aforementioned data-driven wet desulfurization pump set optimization system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0083] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0084] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Those skilled in the art can make various equivalent changes and improvements based on the above embodiments, and all equivalent variations or modifications made within the scope of the claims should fall within the protection scope of the present invention.

Claims

1. A data-driven optimization method for wet desulfurization pump sets, characterized in that, include: The historical and current operating data of the desulfurization system are acquired, and the historical operating data is preprocessed and a steady-state operating data segment is extracted; wherein, the operating data includes multiple operating parameters and pump group operating status information used to reflect changes in operating conditions; Based on the steady-state operation data segment, a clustering algorithm is used to divide the operating condition feature space composed of multiple historical operating parameters into multiple operating condition regions, and a unique operating condition label is assigned to each operating condition region. The current operating data is matched with the operating condition area to identify the current operating condition label; Based on the current operating condition label and the current pump group operating status information, a pre-trained reinforcement learning model is used to determine at least one candidate pump combination action from the preset action space according to the principle of maximizing cumulative reward. Each action in the action space corresponds to a pump combination configuration. Predict the future trend of slurry pH value after each of the candidate pump combinations performs its operation; The corresponding candidate pump combination actions are safety-verified based on the future change trend of each pH value, and a recommended pump combination action is determined from the at least one candidate pump combination action based on the verification result, and its execution is controlled.

2. The data-driven wet desulfurization pump set optimization method as described in claim 1, characterized in that, Based on the steady-state operating data segment, a clustering algorithm is used to divide the operating condition feature space composed of multiple historical operating parameters into multiple operating condition regions, and a unique operating condition label is assigned to each operating condition region. Specifically, this includes: Extract a multi-dimensional operating condition feature space composed of the multiple historical operating parameters from the steady-state operating data segment; The multidimensional working condition feature space is discretized into multiple regular grids, each grid being a micro-working condition unit. Historical operating data samples within each grid are statistically analyzed, and a temporary label is pre-assigned to each grid. Based on the topological adjacency relationship between grids, and using preset sample size thresholds, liquid-to-gas ratio similarity, liquid-to-gas ratio zone consistency, and pump group structure consistency conditions, adjacent grids that meet the conditions in historical operation data samples are clustered and merged to form multiple connected operating condition regions. For each connected operating condition region, the pre-assigned temporary label corresponding to the grid with the largest number of samples in that region is used as the final operating condition label for that region.

3. The data-driven wet desulfurization pump set optimization method as described in claim 1, characterized in that, After identifying the current operating condition label, the process also includes: Determine the current operating mode of the desulfurization system based on current operating data; If the current operation mode is steady state, the step of determining at least one candidate pump combination action from the preset action space based on the current operating condition label and the current pump group operating status information, using a pre-trained reinforcement learning model and the principle of maximizing cumulative reward is executed. If the current operating mode is rapidly changing, a preset risk assessment and action decision-making rule is adopted to determine at least one candidate pump combination action based on the urgency of emission compliance and the trend of operating condition changes.

4. The data-driven wet desulfurization pump set optimization method as described in claim 1 or 3, characterized in that, Based on the current operating condition label and the current pump group operating status information, and using a pre-trained reinforcement learning model, at least one candidate pump combination action is determined from a preset action space according to the principle of maximizing cumulative reward. Specifically, this includes: The decision state of the pre-trained reinforcement learning model is constructed using the current operating condition label and the current pump group operating status information. Based on the current input decision state, the Q value of each optional pump combination action is queried from the Q table corresponding to the current operating condition label; wherein, the Q value represents the expected cumulative reward obtained by performing the corresponding pump combination action under the current input decision state, and the expected cumulative reward is calculated by a multi-factor weighted reward function, wherein the reward factors of the multi-factor weighted reward function include a power factor for encouraging energy saving, a load balancing factor for encouraging load balancing, a switching factor for limiting the number of pump start-ups and shutdowns, and an emission factor for giving segmented rewards or penalties based on emission concentration ranges; An ε-greedy strategy is adopted to select at least one pump combination action as the candidate pump combination action based on the Q value.

5. The data-driven wet desulfurization pump set optimization method as described in claim 3, characterized in that, The operational data also includes emission concentrations. The method employs preset risk assessment and action decision rules to determine at least one candidate pump combination action based on the urgency of emission compliance and the trend of operating condition changes. Specifically, this includes: The effect risk level, trend intensity level, and structural mismatch level are assessed based on the emission concentration, operating parameters, and current pump unit operating status information in the current operating data, respectively. The effect risk level represents the urgency of achieving emission standards, the trend intensity level represents the rate of deterioration of operating conditions, and the structural mismatch level represents the necessity of adjusting the current pump unit operation. The action bias is determined based on the effect risk level and the trend intensity level. The action bias is one of maintaining the current pump combination, switching the pump combination, or increasing the number of operating pumps. If the action is biased towards switching pump combinations, the switching priority is adjusted according to the structural mismatch level; Based on the action bias, action optimization is performed in the union of the similar pump combination set and the historical preferred pump combination set to determine at least one pump combination action as the candidate pump combination action; wherein, the similar pump combination set refers to the set of all pump combination actions in the preset action space whose Hamming distance from the current pump combination action does not exceed a preset distance threshold, and the historical preferred pump combination set refers to the set of several pump combination actions that have been used the most times or have the highest operating effect evaluation in the historical operating data corresponding to the current operating condition label.

6. The data-driven wet desulfurization pump set optimization method as described in claim 1, characterized in that, The step of performing a safety check on the corresponding candidate pump combination actions based on the future change trend of each of the pH values, and determining a recommended pump combination action from the at least one candidate pump combination action based on the check result, specifically includes: Based on maximizing the cumulative reward, the future trend of pH value corresponding to each candidate pump combination action is compared with the preset slurry pH value process range. If the future trend of the pH value corresponding to the candidate pump combination action being checked remains stable within the preset slurry pH value process range, then the check is passed, and the candidate pump combination action is determined as the recommended pump combination action. If all candidate pump combination actions fail the safety check, then the current pump combination action is determined to be the recommended pump combination action.

7. The data-driven wet desulfurization pump set optimization method as described in claim 1, characterized in that, The prediction of the future trend of slurry pH value after the execution of each of the candidate pump combination actions specifically includes: Based on the pump combination configuration corresponding to each candidate pump combination action, the current operating data, and the slurry pH value at multiple time steps, a prediction model based on the XGBoost algorithm is used to predict the slurry pH value change trend within a preset time period after each candidate pump combination action is executed.

8. The data-driven wet desulfurization pump set optimization method as described in claim 7, characterized in that, The data-driven wet desulfurization pump set optimization method further includes: The system checks whether the preset incremental update triggering conditions are met. The incremental update triggering conditions include: the time interval since the last incremental update reaches the preset update cycle, and the number of samples of newly added steady-state operating data accumulated since the last incremental update reaches the preset minimum data volume threshold. When the incremental update triggering condition is met, the newly added steady-state operating data is used to perform a non-intrusive hot update on the working condition region divided by the clustering algorithm, the Q-table of the reinforcement learning model, and the prediction model, and historical model versions are retained during the update process for fault rollback.

9. The data-driven wet desulfurization pump set optimization method as described in claim 7, characterized in that, After determining the recommended pump combination action from the at least one candidate pump combination action based on the verification result, the method further includes: If the difference in the number of operating pumps between the recommended pump combination action and the current pump combination action exceeds a preset number threshold, a step-by-step approximation strategy is activated. At least one transition pump combination configuration is determined based on the recommended pump combination action, and the slurry circulation pump group is controlled to execute each of the transition pump combination configurations in sequence until the recommended pump combination action is finally executed. In this case, the change in the number of operating pumps in each pump group adjustment does not exceed the preset number threshold.

10. A data-driven wet desulfurization pump set optimization system, characterized in that, include: The data acquisition and preprocessing module is used to acquire historical and current operating data of the desulfurization system, preprocess the historical operating data, and extract steady-state operating data segments; wherein, the operating data includes multiple operating parameters and pump group operating status information to reflect changes in operating conditions; The working condition clustering and identification module is used to divide the working condition feature space composed of multiple historical operating parameters based on the steady-state operating data segment using a clustering algorithm, forming multiple working condition regions, and assigning a unique working condition label to each working condition region. The identification and judgment module is used to match the current operating data with the operating condition area, identify the current operating condition label, and determine whether the current operating mode of the desulfurization system is a steady-state operating mode or a rapidly changing operating mode based on the current operating data. The reinforcement learning recommendation module has a built-in pre-trained reinforcement learning model, which is used to determine at least one candidate pump combination action from a preset action space based on the current working condition label and the current pump group operating status information in steady-state operation mode, according to the principle of maximizing cumulative reward. Each action in the action space corresponds to a pump combination configuration. The rapid response decision module is used to determine at least one candidate pump combination action based on the urgency of emission compliance and the trend of changing operating conditions by adopting preset risk assessment and action decision rules in rapidly changing operating modes. The pH prediction module is used to predict the future trend of slurry pH value after the execution of each candidate pump combination action. The engineering constraint and decision module is used to perform safety verification on the corresponding candidate pump combination actions based on the future change trend of each pH value, and to determine the recommended pump combination action from the at least one candidate pump combination action based on the verification result. The incremental training module is used to perform non-intrusive hot updates on the working condition regions divided by the clustering algorithm, the Q-table of the reinforcement learning model, and the prediction model using newly added steady-state operating data, and retains historical model versions for fault rollback during the update process.

Images