Ammonia desulfurization optimization control system based on machine learning algorithm

Abstract

The invention belongs to the technical field of industrial flue gas purification, and discloses an ammonia desulfurization optimization control system based on a machine learning algorithm, which comprises a feature extraction module, a working condition clustering module, a dynamic optimization module and a self-adaptive feedback module, the feature extraction module is used for collecting multi-dimensional operation parameters in a coal burning process, performing data dimension reduction through a PCA algorithm, and extracting key working condition features; the working condition clustering module identifies different operation working condition modes; the dynamic optimization module can construct a multi-modal optimization neural network based on different operation condition modes, and generates optimal control parameters in real time. Through PCA dimension reduction processing of the feature extraction module and in combination with a working condition clustering algorithm, automatic mode recognition and strategy switching under complex working conditions are achieved, fluctuation of desulfurization efficiency is reduced, and meanwhile the problems that the ammonia water adding amount depends on experience setting, raw material waste and secondary pollution are likely to be caused, and the operation and maintenance cost is large are solved.

Description

Technical Field

[0001] This application belongs to the field of industrial flue gas purification technology, specifically relating to an optimized control system for ammonia desulfurization based on machine learning algorithms. Background Technology

[0002] Industrial flue gas purification is a crucial environmental protection measure to address harmful gases and particulate matter emissions from industrial production. It primarily targets flue gas generated by industries such as coal-fired power plants, steel smelting, and cement production, which contain pollutants including sulfur dioxide, nitrogen oxides, particulate matter, and heavy metals. Currently, industrial flue gas purification technology encompasses multiple core modules, including dust removal, desulfurization, and denitrification. Among numerous desulfurization processes, ammonia-based desulfurization control technology stands out due to its unique reaction mechanism and resource utilization advantages. This technology utilizes ammonia water as an absorbent to chemically react with sulfur dioxide in the flue gas, ultimately generating ammonium sulfate fertilizer suitable for agriculture through an oxidation process. This achieves both efficient pollutant capture and transforms treatment costs into economic benefits. Compared to the traditional limestone-gypsum method, ammonia-based desulfurization has higher reactivity and a lower calcium-to-sulfur ratio requirement, making it particularly suitable for deep purification needs under high-sulfur fuel conditions. During the implementation of ammonia-based desulfurization technology, an optimized control system is required to achieve efficient operation and maintenance of the desulfurization system.

[0003] However, most current optimized control systems for ammonia-based desulfurization rely primarily on manual experience or fixed parameter models, resulting in three major problems: First, they struggle to dynamically adapt to complex operating conditions (such as load fluctuations of ±20% and coal sulfur content variations of 0.5%-3.0%), leading to desulfurization efficiency fluctuations of 15%-20%; second, the ammonia dosage is based on empirical settings, often resulting in ammonia slip rates exceeding 80%. This not only wastes raw materials but also causes secondary pollution; thirdly, the system requires real-time manual monitoring and adjustment, and the operation and maintenance cost accounts for more than 30% of the total energy consumption of the desulfurization system. At the same time, traditional intelligent control solutions require a large amount of manually labeled historical data (such as the correspondence between operating conditions and optimal parameters), which is costly and time-consuming, making it difficult to implement on a large scale in industrial scenarios. Summary of the Invention

[0004] This application provides an optimized control system for ammonia desulfurization based on machine learning algorithms, which aims to solve the problems in the existing technology, such as difficulty in dynamic adaptation under complex operating conditions, reliance on experience to set the ammonia dosage, easy waste of raw materials and secondary pollution, and high operation and maintenance costs.

[0005] An optimized control system for ammonia desulfurization based on machine learning algorithms includes a feature extraction module, an operating condition clustering module, a dynamic optimization module, and an adaptive feedback module.

[0006] The feature extraction module is used to collect multi-dimensional operating parameters during the coal combustion process and to perform data dimensionality reduction and extract key operating condition features through the PCA algorithm.

[0007] The operating condition clustering module can dynamically divide key operating condition features using self-organizing mapping networks or clustering algorithms to identify different operating condition modes.

[0008] The dynamic optimization module can construct a multimodal optimization neural network based on different operating condition modes output by the operating condition clustering module, and generate optimal control parameters in real time by combining reinforcement learning algorithms or model predictive control.

[0009] The specific details of the reinforcement learning algorithm are as follows:

[0010] 1) Parameter settings;

[0011] 2) Construct the input state of DQN based on the clustering results of the current working conditions and real-time parameters;

[0012] 3) Through - Greedy strategy for selecting actions;

[0013] 4) After the action is performed, the desulfurization efficiency is collected. ,exit Feedback data such as concentration and ammonia consumption are used to calculate the reward value R.

[0014] 5) Store the "Status-Action-Reward-Next Status" data into the experience replay pool;

[0015] 6) When the amount of data in the replay pool is ≥1000, randomly sample 32 data points to train the DQN network and output the mean square error between the target Q value and the predicted Q value.

[0016] 7) Update the target network parameters every 100 steps, repeating steps 2)-6) to achieve dynamic optimization of control parameters;

[0017] The adaptive feedback module introduces a model adaptive learning strategy, based on desulfurization efficiency and outlet... The actual concentration feedback dynamically updates the network weights.

[0018] Furthermore, the specific working content of the PCA algorithm is as follows:

[0019] 1) Parameter settings;

[0020] 2) Remove outliers from the collected multidimensional operating parameters;

[0021] 3) Standardize the data after outlier removal to eliminate the influence of units;

[0022] 4) Calculate the covariance matrix of the standardized data and solve for the eigenvalues ​​and eigenvectors;

[0023] 5) Sort the feature values ​​from largest to smallest, and select the top k feature vectors with a cumulative contribution rate ≥ 92% as key working condition features.

[0024] Furthermore, the specific working content of the organization mapping network is as follows:

[0025] 1) Parameter settings;

[0026] 2) Initialize the weights of the output layer neurons;

[0027] 3) Randomly select one sample from the feature set of PCA output and calculate its Euclidean distance to all output layer neurons;

[0028] 4) Select the neuron with the smallest distance as the "winning neuron" and determine its neighborhood range;

[0029] 5) Update the weights of the winning neuron and its neighboring neurons, where the weight update formula is as follows:

[0030]

[0031] in, This is the current weight vector. for The learning rate at time step x is the input sample;

[0032] 6) Repeat steps 3)-5) until the end of the iteration, and divide the output layer neurons into 5-8 working condition modes according to the weight similarity.

[0033] Furthermore, the specific working content of the clustering algorithm is as follows:

[0034] 1) Parameter settings;

[0035] 2) Data preprocessing: Z-score standardization is performed on the feature data after PCA dimensionality reduction to eliminate the influence of dimensional differences;

[0036] 3) Initial cluster center selection: Randomly select one sample from the sample set as the first cluster center; calculate the shortest distance from the remaining samples to the selected center. Randomly select the next center using a distance-weighted probability system, repeating this process until k initial centers are selected.

[0037] 4) Sample assignment: Calculate the Euclidean distance between each sample and the k centers, and assign the sample to the cluster to which the nearest center belongs;

[0038] 5) Update cluster centers: Calculate the mean of features of all samples within each cluster, and use it as the new cluster center;

[0039] 6) Convergence criterion: If the update magnitude of all centers is less than... If the maximum number of iterations is reached, stop iterating; otherwise, return 4).

[0040] 7) Output clustering results: Divide the samples into k working condition patterns and output the central features and sample affiliation labels for each working condition pattern.

[0041] Furthermore, the shortest distance The calculation formula is as follows:

[0042]

[0043] in, Indicates the relationship between sample x and the selected center. The Euclidean distance.

[0044] Furthermore, the reward value R is calculated as follows:

[0045]

[0046] in, For actual desulfurization efficiency, The target efficiency is 95%, and C represents the ammonia consumption. Based on the baseline consumption, To control the amount of parameter variation.

[0047] Furthermore, the specific details of the model predictive control are as follows:

[0048] 1) Parameter settings;

[0049] 2) State observation: At the current time k, the exit data is collected through the sensor. The actual output y(k) such as concentration and pH value is combined with a Kalman filter to estimate the current state of the system x(k);

[0050] 3) Future state prediction: Based on the system model and the current state x(k), predict the future state. Step output trajectory ;

[0051] 4) Rolling optimization: Under constraints, minimize the objective function and solve for the future. Optimal control sequence of steps ;

[0052] 5) Control Execution: Only execute the control variable for the current step. The subsequent control measures will not be implemented for the time being;

[0053] 6) Closed-loop feedback: waiting for one sampling period Then, collect new actual output. (k+1), return to step 2) and repeat to achieve dynamic optimization control.

[0054] Furthermore, the model update strategy is as follows: when the export... Concentration fluctuation variance > 500 ( When the desulfurization efficiency is below 92%, the model is triggered to update adaptively.

[0055] Compared with the prior art, this application has at least the following beneficial effects:

[0056] Based on further analysis and research of existing technical problems, this application achieves automatic pattern recognition and strategy switching under complex operating conditions by real-time acquisition of multi-dimensional operating parameters and PCA dimensionality reduction processing of feature extraction modules, combined with operating condition clustering algorithms (such as SOM / K-means++), thereby reducing the fluctuation of desulfurization efficiency.

[0057] Meanwhile, the dynamic optimization module, based on reinforcement learning (DQN) or model predictive control (MPC), aims to maximize desulfurization efficiency and minimize ammonia slip, thereby calculating the optimal control parameters for ammonia water dosage in real time, reducing raw material waste while avoiding secondary pollution.

[0058] This application also forms a closed-loop control chain through various modules (feature extraction → operating condition clustering → dynamic optimization → adaptive feedback), achieving millisecond-level response from data acquisition to strategy execution. The adaptive feedback mechanism replaces manual monitoring, and combined with anomaly detection algorithms, it provides early warnings of potential faults, reducing the need for manual intervention and lowering operation and maintenance costs. This breaks away from the traditional model that relies on expert experience, supporting remote diagnosis and predictive maintenance. Furthermore, the operating condition clustering module and the dynamic optimization module work together, utilizing unsupervised learning to uncover implicit operating condition-parameter relationships, reducing reliance on manually labeled data and shortening the model iteration cycle. Attached Figure Description

[0059] Figure 1 This is a block diagram of an optimized control system for ammonia desulfurization based on machine learning algorithms, provided as an embodiment of this application. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0061] like Figure 1 As shown, this application provides an optimized control system for ammonia desulfurization based on machine learning algorithms, including a feature extraction module, an operating condition clustering module, a dynamic optimization module, and an adaptive feedback module.

[0062] The feature extraction module collects multidimensional operating parameters during the coal combustion process and uses the PCA (Principal Component Analysis) algorithm to reduce the dimensionality of the data and extract key operating condition features. This is used to accurately identify and retain key principal component features characterizing the core operating conditions, providing a low-redundancy, high-information-density data foundation for subsequent modeling and optimization.

[0063] Among them, the multi-dimensional operating parameters include the desulfurization tower inlet. Concentration, flue gas flow rate, flue gas temperature, flue gas pressure, pH value, liquid-to-gas ratio, ammonia concentration, ammonia flow rate, oxidation air volume, circulating pump current, outlet Multidimensional operating parameters such as concentration.

[0064] The specific workings of the PCA algorithm are as follows:

[0065] 1) Parameter settings: Principal component contribution rate threshold ≥ 92%, data standardization method adopts Z-score (mean). =0, standard deviation σ=1), covariance matrix calculation window is 5min, k principal components are retained after dimensionality reduction (covering more than 92% of the information of the original data).

[0066] 2) Remove outliers from the collected multidimensional operating parameters (using the 3σ principle to remove data points that deviate from the mean ±3σ).

[0067] 3) Standardize the data after outlier removal to eliminate the influence of units;

[0068] 4) Calculate the covariance matrix of the standardized data and solve for the eigenvalues ​​and eigenvectors. Let the standardized data matrix be X (size n×p, n = number of samples, p = number of variables), where each column corresponds to a centered and standardized variable (i.e., each column has a mean of 0 and a standard deviation of 1). In this case, each row of X represents an observed sample, and each column corresponds to a certain operating parameter (e.g., ...). (Concentration, temperature, etc.)

[0069] Wherein, the covariance matrix The calculation formula is as follows:

[0070]

[0071] Where X is a standardized data matrix, Let X be the transpose of X, and n be the number of samples. Eigenvalues The calculation formula is as follows:

[0072]

[0073] in, This refers to the determinant operation of a matrix, where I is an n×p identity matrix. Eigenvectors... The calculation formula is as follows:

[0074]

[0075] in, To and The corresponding eigenvalues.

[0076] 5) Sort the feature values ​​from largest to smallest, and select the top k feature vectors with a cumulative contribution rate ≥ 92% as key working condition features.

[0077] The operating condition clustering module can dynamically segment key operating condition features using self-organizing map networks (SOM) or clustering algorithms (such as K-means++) to identify different operating condition modes. By dynamically quantifying the similarity and difference boundaries between different operating conditions, it provides a structured classification basis for subsequent adaptation, effectively supporting the differentiated formulation of subsequent control strategies and the optimized allocation of resources.

[0078] The specific tasks of the organization mapping network are as follows:

[0079] 1) Parameter settings: Number of input layer nodes = number of features after PCA dimensionality reduction (k), output layer is a 10×10 hexagonal grid (100 neurons in total), initial learning rate... =0.8, learning rate decay coefficient α=0.95, initial value of neighborhood radius =5, neighborhood decay coefficient β=0.9, number of iterations=5000.

[0080] 2) Initialize the weights of the output layer neurons (randomly generate vectors in the interval [-0.5, 0.5]);

[0081] 3) Randomly select one sample from the feature set of PCA output and calculate its Euclidean distance to all output layer neurons;

[0082] 4) Select the neuron with the smallest distance as the "winning neuron" and determine its neighborhood range (which decreases with the number of iterations).

[0083] 5) Update the weights of the winning neuron and its neighboring neurons, where the weight update formula is as follows:

[0084]

[0085] in, This is the current weight vector. for The learning rate at time step x is the input sample;

[0086] 6) Repeat steps 3)-5) until the end of the iteration, and divide the output layer neurons into 5-8 types of working conditions according to the weight similarity (such as "high sulfur high load working condition", "low sulfur low load working condition" and "variable load transition working condition").

[0087] The specific tasks of clustering algorithms are as follows:

[0088] 1) Parameter settings: Number of clusters k = 5-8 (determined by the elbow rule, i.e., the k value corresponding to the sum of squared errors and SSE mutation points); Euclidean distance is used as the distance metric; initial cluster centers are selected using the distance maximization principle; convergence threshold... =0.001 (stop iteration when the change in cluster centers is less than this value), maximum number of iterations = 1000 times, sample weight is uniformly 1 (all samples participate in clustering equally).

[0089] 2) Data preprocessing: The feature data after PCA dimensionality reduction are standardized by Z-score (mean μ=0, standard deviation σ=1) to eliminate the influence of dimensional differences;

[0090] 3) Initial cluster center selection: Randomly select one sample from the sample set as the first cluster center; calculate the shortest distance from the remaining samples to the selected center. Randomly select the next center using a probability weighted by the squared distance, repeating this process until k initial centers are selected, and find the shortest distance. The calculation formula is as follows;

[0091]

[0092] in, Indicates the relationship between sample x and the selected center. The Euclidean distance.

[0093] 4) Sample assignment: Calculate the Euclidean distance between each sample and the k centers, and assign the sample to the cluster to which the nearest center belongs;

[0094] 5) Update cluster centers: Calculate the mean of features of all samples within each cluster, and use it as the new cluster center;

[0095] 6) Convergence criterion: If the update magnitude (distance between old and new centers) of all centers is less than 1 / 3, then the convergence criterion is determined. If the maximum number of iterations is reached, stop iterating; otherwise, return 4).

[0096] 7) Output clustering results: Divide the samples into k operating conditions (such as "high sulfur and high load", "low sulfur and low load" and "variable load transition"), and output the central features and sample affiliation labels for each operating condition.

[0097] The dynamic optimization module can construct a multimodal optimization neural network based on the different operating conditions output by the operating condition clustering module. Combined with reinforcement learning algorithms (such as DQN) or model predictive control (MPC), it can generate optimal control parameters (including ammonia flow rate, circulating pump frequency, oxidation air volume, etc.) in real time.

[0098] The specific details of reinforcement learning algorithms (such as DQN) are as follows:

[0099] 1) Parameter settings: State space = operating condition category + current operating parameters (8-10 dimensions in total), Action space = adjustment range of 3 key control parameters (ammonia flow rate: ±5-50L / h, circulating pump frequency: ±1-5Hz, oxidation air volume: ±50-200). The reward function is as follows:

[0100]

[0101] in, For actual desulfurization efficiency, The target efficiency is 95%, and C represents the ammonia consumption. Based on the baseline consumption, To control the amount of parameter variation;

[0102] Network structure: Input layer 8-10 neurons, hidden layers 2 (64 neurons per layer, ReLU activation function), output layer = number of action spaces (e.g., 15 discrete actions), experience replay pool capacity = 10,000 records, target network update frequency = 100 steps, learning rate = 0.001, discount factor. =0.9.

[0103] 2) Construct the input state of DQN based on the clustering results of the current working conditions and real-time parameters;

[0104] 3) Through - Greedy strategy ( Initial value = 0.9, linearly decreasing to 0.1 with each step; select action (i.e., adjust the output control parameter).

[0105] 4) After the action is performed, the desulfurization efficiency is collected. ,exit Feedback data such as concentration and ammonia consumption are used to calculate the reward value R.

[0106] 5) Store the "Status-Action-Reward-Next Status" data into the experience replay pool;

[0107] 6) When the amount of data in the replay pool is ≥1000, randomly sample 32 data points to train the DQN network, and output the mean square error between the target Q value and the predicted Q value. The mean square error function is as follows;

[0108]

[0109] Where N is the number of samples in each iteration. For the target Q value, For the current network state-action pair The predicted Q value, These are the network parameters to be updated.

[0110]

[0111] in, For the immediate reward (R calculated in step 4), Represents the target network parameters. For the next state The operation with the highest Q value among all possible actions.

[0112] 7) Update the target network parameters every 100 steps, repeat steps 2)-6) to achieve dynamic optimization of control parameters.

[0113] The specific details of model predictive control are as follows:

[0114] 1) Parameter settings: Prediction time domain =10 steps (1 minute per step, predicting the system state for the next 10 minutes), control time domain =5 steps (only optimizing control variables for the next 5 steps); Sampling period =60s (optimization is performed once per minute); a linear state-space model is used. );

[0115] Constraints: Export Concentration ≤35 Ammonia slip rate ≤ 5 Ammonia flow rate ≤ 500 L / h; the diagonal elements of the objective function weight matrix Q (output error weight) are [10, 5, 3]. (Ranked by importance of concentration, pH value, and ammonia slip rate), R (weight of control variable change) = 0.1I (suppress drastic fluctuations in control variable).

[0116] 2) State observation: At the current time k, the exit data is collected through the sensor. The actual output y(k) such as concentration and pH value is combined with a Kalman filter to estimate the current state of the system x(k);

[0117] 3) Future state prediction: Based on the system model and the current state x(k), predict the future state. Step output trajectory ;

[0118] 4) Rolling optimization: Under constraints, minimize the objective function (weighted sum of output tracking errors + weighted sum of control quantity changes) to solve for the future... Optimal control sequence of steps ;

[0119] 5) Control Execution: Only execute the control variable for the current step. (For example, adjusting the opening of the ammonia water flow valve), subsequent control measures will not be implemented for the time being;

[0120] 6) Closed-loop feedback: waiting for one sampling period Then, collect new actual output. (k+1), return to step 2) and repeat to achieve dynamic optimization control.

[0121] The adaptive feedback module introduces a model adaptive learning strategy, based on the desulfurization efficiency ( ) and exports The actual concentration feedback dynamically updates the network weights, enabling continuous optimization of the control system.

[0122] Model update strategy: When the exit occurs within 5 consecutive minutes When the concentration fluctuation variance is greater than 500 (mg / Nm³)² or the desulfurization efficiency is less than 92%, the model adaptive update is triggered.

[0123] Weight adjustment logic: The incremental learning algorithm is adopted to fuse the newly collected working condition-feedback data with historical data at a 1:1 weight, and retrain the SOM clustering model and DQN optimization model. The update cycle is ≤30min.

[0124] Fault self-correction: When a sensor abnormality is detected (such as data fluctuation > 20%), it automatically switches to the backup parameter set (based on the historical best value of similar operating conditions) to ensure stable system operation.

[0125] In the aforementioned optimized control system for ammonia desulfurization based on machine learning algorithms, the system achieves automatic pattern recognition and strategy switching under complex operating conditions by real-time acquisition of multi-dimensional operating parameters and PCA dimensionality reduction processing of the feature extraction module, combined with operating condition clustering algorithms (such as SOM / K-means++), thereby reducing fluctuations in desulfurization efficiency.

[0126] Meanwhile, the dynamic optimization module, based on reinforcement learning (DQN) or model predictive control (MPC), aims to maximize desulfurization efficiency and minimize ammonia slip, thereby calculating the optimal control parameters for ammonia water dosage in real time, reducing raw material waste while avoiding secondary pollution.

[0127] Each module (feature extraction → operating condition clustering → dynamic optimization → adaptive feedback) forms a closed-loop control chain, achieving millisecond-level response from data acquisition to strategy execution. The adaptive feedback mechanism replaces manual monitoring, and combined with anomaly detection algorithms, it provides early warnings of potential faults, reducing the need for manual intervention and lowering operation and maintenance costs. This breaks away from the traditional model that relies on expert experience, supporting remote diagnostics and predictive maintenance. Furthermore, the operating condition clustering module and the dynamic optimization module work together, utilizing unsupervised learning to uncover implicit operating condition-parameter relationships, reducing reliance on manually labeled data and shortening the model iteration cycle.

[0128] Here is a specific application example:

[0129] A comparative experiment was conducted on a 100MW thermal power unit's ammonia-based desulfurization system for 30 days (total operating time 720 hours). During the experiment, the target efficiency ranged from 92% to 97%, the sulfur content of the coal was 0.8% to 2.5%, and the load was 60% to 100%. The technical effects were verified as follows:

[0130] Calculation process:

[0131] Desulfurization efficiency stability calculation: Stability rate = (duration of time the efficiency is in the target range (92%-97%) / total running time) × 100%; the stability rate of the traditional solution is 60%, the stability rate of the solution of this invention is 66.6%, and the improvement is (66.6%-60%) / 60% × 100% = 11%;

[0132] exit The reduction in concentration fluctuation variance = (traditional variance - variance of this scheme) / traditional variance × 100% = (128.5 - 75.2) / 128.5 × 100% ≈ 41.5%.

[0133] Table 1 shows the desulfurization efficiency stability rate and outlet. Table of results on the decrease in concentration fluctuation variance

[0134]

[0135] Table 1 shows that the present invention improves efficiency: the desulfurization efficiency stability rate is increased by 8-12%, and the export efficiency is improved. The variance of concentration fluctuation decreased by more than 40%;

[0136] The ammonia water consumption was calculated based on a comparative test conducted over 30 days (total operating time 720 hours) on a 100MW thermal power unit ammonia desulfurization system.

[0137] Calculation process:

[0138] The reduction in ammonia consumption = (traditional consumption - consumption under this plan) / traditional consumption × 100% = (320 - 295) / 320 × 100% = 7.8%;

[0139] The decrease in ammonia slip rate = (7.5-4.8) / 7.5×100% = 36%.

[0140] Table 2 shows the results of ammonia consumption.

[0141]

[0142] As shown in Table 2, the present invention optimizes costs, reduces ammonia consumption by 5-10%, and effectively suppresses ammonia escape.

[0143] Intelligent operation and maintenance: Reduces manual intervention by 70% and adapts to complex operating conditions such as load changes and coal type switching;

[0144] Frequency of manual intervention: Traditional methods require manual parameter adjustment every hour, averaging 24 interventions per day; the method of this invention requires ≤7 interventions per day, a reduction of (24-7) / 24×100%≈70.8%;

[0145] Adaptability to operating conditions: Under operating conditions where the sulfur content of coal abruptly increases from 1.0% to 2.5% (within 1 hour) and the load increases from 70% to 100% (within 2 hours), the solution of this invention is applicable. Concentration fluctuation ≤5 The desulfurization efficiency remained above 94%, while traditional methods saw a sharp drop in efficiency to 87% and an outlet concentration exceeding 50%. This addresses the issue of not needing to label historical data, significantly reducing the barrier to algorithm implementation and maintenance costs.

[0146] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. An optimized control system for ammonia-based desulfurization based on machine learning algorithms, characterized in that, It includes a feature extraction module, a working condition clustering module, a dynamic optimization module, and an adaptive feedback module; The feature extraction module is used to collect multi-dimensional operating parameters during the coal combustion process and to perform data dimensionality reduction and extract key operating condition features through the PCA algorithm. The operating condition clustering module can dynamically divide key operating condition features using self-organizing mapping networks or clustering algorithms to identify different operating condition modes. The dynamic optimization module can construct a multimodal optimization neural network based on different operating condition modes output by the operating condition clustering module, and generate optimal control parameters in real time by combining reinforcement learning algorithms or model predictive control. The specific details of the reinforcement learning algorithm are as follows: 1) Parameter settings; 2) Construct the input state of DQN based on the clustering results of the current working conditions and real-time parameters; 3) Through - Greedy strategy for selecting actions; 4) After the action is performed, the desulfurization efficiency is collected. ,exit Feedback data such as concentration and ammonia consumption are used to calculate the reward value R. 5) Store the "Status-Action-Reward-Next Status" data into the experience replay pool; 6) When the amount of data in the replay pool is ≥1000, randomly sample 32 data points to train the DQN network and output the mean square error between the target Q value and the predicted Q value. 7) Update the target network parameters every 100 steps, repeating steps 2)-6) to achieve dynamic optimization of control parameters; The adaptive feedback module introduces a model adaptive learning strategy, based on desulfurization efficiency and outlet... The actual concentration feedback dynamically updates the network weights.

2. The optimized control system for ammonia desulfurization based on machine learning algorithms according to claim 1, characterized in that, The specific working content of the PCA algorithm is as follows: 1) Parameter settings; 2) Remove outliers from the collected multidimensional operating parameters; 3) Standardize the data after outlier removal to eliminate the influence of units; 4) Calculate the covariance matrix of the standardized data and solve for the eigenvalues ​​and eigenvectors; 5) Sort the feature values ​​from largest to smallest, and select the top k feature vectors with a cumulative contribution rate ≥ 92% as key working condition features.

3. The optimized control system for ammonia desulfurization based on machine learning algorithms according to claim 1, characterized in that, The specific working content of the organization mapping network is as follows: 1) Parameter settings; 2) Initialize the weights of the output layer neurons; 3) Randomly select one sample from the feature set of PCA output and calculate its Euclidean distance to all output layer neurons; 4) Select the neuron with the smallest distance as the "winning neuron" and determine its neighborhood range; 5) Update the weights of the winning neuron and its neighboring neurons, where the weight update formula is as follows: ； in, This is the current weight vector. for The learning rate at time step x is the input sample; 6) Repeat steps 3)-5) until the end of the iteration, and divide the output layer neurons into 5-8 working condition modes according to the weight similarity.

4. The optimized control system for ammonia desulfurization based on machine learning algorithms according to claim 1, characterized in that, The specific working process of the clustering algorithm is as follows: 1) Parameter settings; 2) Data preprocessing: Z-score standardization is performed on the feature data after PCA dimensionality reduction to eliminate the influence of dimensional differences; 3) Initial cluster center selection: Randomly select one sample from the sample set as the first cluster center; calculate the shortest distance from the remaining samples to the selected center. Randomly select the next center using a distance-weighted probability system, repeating this process until k initial centers are selected. 4) Sample assignment: Calculate the Euclidean distance between each sample and the k centers, and assign the sample to the cluster to which the nearest center belongs; 5) Update cluster centers: Calculate the mean of features of all samples within each cluster, and use it as the new cluster center; 6) Convergence criterion: If the update magnitude of all centers is less than... If the maximum number of iterations is reached, stop iterating; otherwise, return 4). 7) Output clustering results: Divide the samples into k working condition patterns and output the central features and sample affiliation labels for each working condition pattern.

5. The optimized control system for ammonia desulfurization based on machine learning algorithms according to claim 4, characterized in that, The shortest distance The calculation formula is as follows: ； in, Indicates the relationship between sample x and the selected center. The Euclidean distance.

6. The optimized control system for ammonia desulfurization based on machine learning algorithms according to claim 1, characterized in that, The reward value R is calculated as follows: ； in, For actual desulfurization efficiency, The target efficiency is 95%, and C represents the ammonia consumption. Based on the baseline consumption, To control the amount of parameter variation.

7. The optimized control system for ammonia desulfurization based on machine learning algorithms according to claim 1, characterized in that, The specific details of the model predictive control are as follows: 1) Parameter settings; 2) State observation: At the current time k, the exit data is collected through the sensor. The actual output y(k) such as concentration and pH value is combined with a Kalman filter to estimate the current state of the system x(k); 3) Future state prediction: Based on the system model and the current state x(k), predict the future state. Step output trajectory ; 4) Rolling optimization: Under constraints, minimize the objective function and solve for the future. Optimal control sequence of steps ; 5) Control Execution: Only execute the control variable for the current step. The subsequent control measures will not be implemented for the time being; 6) Closed-loop feedback: waiting for one sampling period Then, collect new actual output. (k+1), return to step 2) and repeat to achieve dynamic optimization control.

8. The optimized control system for ammonia desulfurization based on machine learning algorithms according to claim 1, characterized in that, The model update strategy is as follows: when the exit... Concentration fluctuation variance > 500 ( When the desulfurization efficiency is below 92%, the model is triggered to update adaptively.

Images