A method for controlling quality of fly ash of a thermal power unit

Abstract

The application discloses a kind of thermal power unit fly ash quality control method, belong to the technical field of control optimization, including from thermal power unit system acquisition and integration condition data form standardized data set, and based on standardized data set construction prediction model, to predict the key indicators of future fly ash, then utilize the multi-objective optimization engine including expert rule and particle swarm optimization algorithm and digital twin simulation technology, generate and optimize the control instruction covering electric field operation, boiler combustion organization, hierarchical conveying path switching and flue gas conditioning, finally through instruction execution and online incremental learning algorithm update model form closed loop adaptive control.The application solves the problem that traditional control method depends on artificial experience and response lag by constructing dynamic prediction model, improves the foresight and initiative of quality control, and solves the contradiction between high-quality fly ash output rate and system operation energy consumption through multi-objective collaborative optimization.

Description

Technical Field

[0001] This invention belongs to the technical field of control optimization, and in particular relates to a method for controlling the quality of fly ash from thermal power units. Background Technology

[0002] Fly ash is a major solid waste generated by coal-fired power units and an important industrial byproduct for achieving a circular economy. With the increasing demands for raw material performance from downstream industries such as building materials, refined quality control of fly ash to achieve its high-value resource utilization has become a crucial direction for energy conservation, efficiency improvement, and green development in power plants. The core principle is to ensure quality stability and enhance economic benefits. However, in the current process of fly ash quality control in coal-fired power units, due to frequent fluctuations in the quality of coal entering the furnace, the complex and variable boiler combustion conditions, and the strong coupling and nonlinear characteristics between multiple systems such as pulverization and dust removal, traditional control methods mainly rely on the historical experience of operators and delayed adjustments based on post-event sampling and testing results. This approach cannot accurately predict future fly ash quality trends based on real-time industrial big data from unit operation. Especially when facing operating disturbances such as coal type switching and unit load adjustment, the existing control methods are difficult to carry out forward-looking coordinated regulation, which often leads to key indicators such as unburned carbon content and fineness of fly ash exceeding the high-grade standards, resulting in a significant reduction in its resource value, or even turning it into solid waste, bringing economic losses and environmental pressure to power plants. Summary of the Invention

[0003] To overcome the technical problems described in the background section, this invention provides a method for controlling the quality of fly ash in thermal power units. By constructing a dynamic prediction model, it solves the problem that traditional control methods rely on human experience and have a slow response, thereby improving the foresight and initiative of quality control. At the same time, through multi-objective collaborative optimization, it resolves the contradiction between the high-quality fly ash production rate and the system's energy consumption.

[0004] The technical solution of the present invention is: a method for quality control of fly ash from thermal power units, comprising the following steps:

[0005] S1. Collect historical operating data and real-time operating condition data of thermal power units, and form a unified raw operating condition data package after time alignment and preliminary cleaning;

[0006] S2. Clean, integrate, and standardize the original working condition data packets to generate a standardized dataset for model building;

[0007] S3. Based on standardized datasets, construct and solidify a predictive model that can reflect the dynamic characteristics of the system, and generate prediction results for key quality indicators of fly ash in the future through the predictive model.

[0008] S4. Based on the prediction results and combined with the preset processing rules, generate control instructions including electric field operation, combustion organization, delivery path and flue gas conditioning.

[0009] S5. Send control commands to the corresponding actuators in the thermal power unit and execute them to dynamically adjust the operating status of the system.

[0010] S6. Based on the new operating status data generated after the adjustment, the internal parameters of the prediction model are updated using an online incremental learning algorithm to continuously optimize the prediction accuracy of the model.

[0011] S7. Monitor key response data of the system in real time and generate a control effect evaluation report containing key performance indicators;

[0012] S8. When a key indicator in the assessment report shows an anomaly, an early warning mechanism is triggered and corrective control instructions for handling the anomaly are generated.

[0013] S9. Send the correction control command to the system's actuator for rapid intervention and closed-loop iterative adjustment;

[0014] S10. Optimal control parameters, system response data, and evaluation results are stored in a structured manner and output to the monitoring platform.

[0015] Further, step S1 includes the following steps:

[0016] S11. Collect historical operating data and real-time operating condition data from the distributed control system, fuel management information system, laboratory information management system and environmental monitoring system of thermal power units through open platform communication protocols, application programming interfaces or direct database connection;

[0017] S12. Collect core operating parameters that reflect the real-time operating status of the unit at a high sampling frequency, including boiler and combustion related data, pulverizing system related data and dust removal equipment related data;

[0018] S13. The quality analysis data of the coal fed into the furnace and the quality data of fly ash after manual sampling and testing are obtained at a low acquisition frequency and used as the target values ​​for subsequent model training.

[0019] S14. Timestamp align the high-frequency operating condition data with the low-frequency quality data, and use Lagrange interpolation to fill in missing values ​​and remove outliers.

[0020] S15. Integrate the processed data into a structured preprocessed working condition data package and pass it to step S2.

[0021] Further, step S2 includes the following steps:

[0022] S21. Obtain the preprocessed operating condition data packet generated in step S1;

[0023] S22. Normalize the preprocessed operating condition data packets to eliminate the influence of different physical dimensions and generate a normalized feature matrix; the normalization process uses the following formula:

[0024] ;

[0025] in, For the normalized data, The original data values, The minimum value in the historical dataset. The maximum value in the historical dataset;

[0026] S23. Pearson correlation coefficient analysis and L1 regularized LASSO regression model were used to combine feature screening to extract key features that are strongly correlated with fly ash quality and generate a feature optimization dataset.

[0027] S24. Perform principal component analysis on the feature optimization dataset to reduce dimensionality, extract principal components whose cumulative variance contribution rate reaches a preset threshold, and generate a standardized dataset.

[0028] S25. Verify the data quality and consistency of the standardized dataset and output a data quality verification report.

[0029] Furthermore, step S3 includes the following steps:

[0030] S31. Divide the standardized dataset output by S2 into training set, validation set and test set according to time series.

[0031] S32. Construct a long short-term memory network model for the unburned carbon content index;

[0032] S33. Construct a gradient boosting decision tree model for the characteristic particle size index of fly ash;

[0033] S34. Construct a gradient boosting decision tree model for the volcanic ash activity index.

[0034] S35. The initial prediction model is trained using the training set, and the hyperparameters of the model are optimized by grid search combined with cross-validation. Finally, the performance of the optimized model is evaluated using the test set and solidified into the final prediction model.

[0035] S36. The fly ash production rate is obtained through real-time calculation, providing a basis for decision-making regarding graded conveying and ash silo capacity management; the formula for calculating the ash production rate is:

[0036] ;

[0037] in, This represents the theoretical fly ash production rate per unit time. To measure the amount of coal fed into the furnace in real time. It refers to the ash content of the coal fed into the furnace. The unburned carbon content of fly ash is provided by the prediction model in S32.

[0038] S37. During the real-time operation phase, the operating condition data is input into the three final prediction models and combined with the calculation results of S36 to output a comprehensive prediction result data package containing four key indicators for the next 30 minutes.

[0039] Further, step S4 includes the following steps:

[0040] S41. Obtain the comprehensive prediction result of fly ash quality for the next 30 minutes generated in step S3.

[0041] S42. Generate initial control instructions based on a multi-objective optimization rule engine that includes a policy generation layer and a parameter optimization layer; the parameter optimization layer uses an objective function... To optimize the objective, a particle swarm optimization algorithm is used to search within the feasible region of the control variables. The objective function is... The expression is:

[0042] ;

[0043] in, To predict the yield of high-grade fly ash, For the predicted energy consumption per ton of ash collection, and These are the weighting coefficients;

[0044] S43. Input the initial control command into a digital twin model consisting of a boiler combustion module and an electrostatic precipitator module for simulation, and generate simulation operating condition response data;

[0045] S44. Analyze the deviation between the simulation working condition response data and the preset fly ash quality target parameters. If the deviation exceeds the allowable range, generate an instruction to optimize the correction amount.

[0046] S45. Based on the instruction, optimize the correction amount to correct the initial control instruction and output the structured final control instruction;

[0047] S46. Provides a theoretical basis for the electrostatic precipitator operation optimization instruction in the final control instruction. This theoretical basis is based on the following formula for calculating the electrostatic precipitator collection efficiency:

[0048] ;

[0049] in, For theoretical capture efficiency, It is the total area of ​​the dust collection plates. It is the flue gas flow rate. It is the dust ingress speed.

[0050] Further, step S5 includes the following steps:

[0051] S51. Send the final control command to the distributed control system and PLC controller of the thermal power unit through the industrial control network.

[0052] S52. Control the high-voltage power supply and rapping control system of the electrostatic precipitator to optimize the voltage of each electric field and the rapping cycle.

[0053] S53. Send a combustion fine-tuning command to the boiler control system to fine-tune the combustion organization parameters of the air-coal ratio and burner sway angle;

[0054] S54. Control the path switching valve of the pneumatic ash conveying system to switch the graded conveying path according to the predicted fly ash quality.

[0055] S55. Collect device status feedback data in real time during the command execution process and generate an execution log data package containing timestamps and execution results.

[0056] Further, step S6 includes the following steps:

[0057] S61. Obtain the execution log data package generated in step S5, the new operating status data after system adjustment, and the corresponding new fly ash quality test data.

[0058] S62. Using the new test data as the true label, calculate the deviation between the new test data and the model prediction results, and generate a model prediction error analysis report.

[0059] S63. Based on the model prediction error analysis report, determine whether the preset model update triggering conditions are met;

[0060] S64. When the triggering condition is met, the prediction model is retrained using new data through an online incremental learning algorithm that matches the model type; the model update strategy includes:

[0061] C1. For the Long Short-Term Memory (LSTM) network model, a periodic fine-tuning strategy is adopted, and its weight update formula is as follows:

[0062] ;

[0063] in, The updated weight matrix, This is the current weight matrix. To fine-tune the learning rate, This represents the gradient of the loss function with respect to the weights.

[0064] C2. For the gradient boosting decision tree model, the reinforcement learning strategy with residual correction is adopted. The prediction logic of the new model is as follows:

[0065] ;

[0066] in, The updated predicted value, These are the predicted values ​​from the old model. The predicted values ​​of the corrected model trained on the new data residuals;

[0067] S65. Store the updated prediction model and its version information, training date, and other metadata into the model library.

[0068] Further, step S7 includes the following steps:

[0069] S71. Collect key response data of the system after the execution of control commands using the distributed control system of thermal power units;

[0070] S72. Calculate key performance indicators, including high-grade fly ash production rate, fly ash comprehensive utilization value index, energy consumption per ton of ash collection, and environmental emission compliance rate.

[0071] S73. A control effect evaluation report is generated based on a multi-dimensional KPI threshold comparison algorithm. The report evaluates the economics of the dust removal system by calculating the energy consumption per ton of ash collection. The formula for calculating the energy consumption per ton of ash collection is:

[0072] ;

[0073] in, Energy consumption per ton of ash collection This refers to the total operating power of the electrostatic precipitator. This refers to the operating power of the pneumatic ash conveying system. It is the actual fly ash collection rate per unit time;

[0074] S74. When any key performance indicator exceeds the preset safety or economic threshold range, it shall be marked as an abnormal state in the evaluation report.

[0075] S75. Automatically generate and output assessment reports to the power plant's intelligent environmental monitoring platform according to a preset cycle.

[0076] Further, step S8 includes the following steps:

[0077] S81. When abnormal data is marked in the evaluation report of step S7, an early warning mechanism is triggered. The triggering basis for the early warning mechanism is the fly ash quality deterioration index. The calculation formula is as follows:

[0078] ;

[0079] in, The fly ash quality deterioration index. and These are the real-time values ​​for unburned carbon content and fineness, respectively. and For its target value, and These are the weighting coefficients;

[0080] S82. Use Bayesian networks for root cause diagnosis to infer the most likely source of failure and generate an abnormal cause diagnosis report.

[0081] S83. Based on the abnormal cause diagnosis report and matching the preset expert rule knowledge base, generate a draft of the correction control instruction;

[0082] S84. Review the potential impact of the revised control instruction draft through the control strategy safety simulation module, and output the final revised control instruction after confirming that there are no conflicts.

[0083] S85. Record the complete process of early warning triggering, cause diagnosis, instruction generation and execution to the system event log.

[0084] Further, step S9 includes the following steps:

[0085] S91. Send the final corrected control command generated in step S8 to the actuator module of the distributed control system of the thermal power unit.

[0086] S92. The actuator module schedules and executes correction control tasks according to preset safety and process priorities;

[0087] S93. Through rapid response closed-loop adjustment, until the monitoring data shows that the fly ash quality deterioration index has recovered to within the threshold, and generate real-time operation data of the intervention process;

[0088] S94. Verify whether the real-time operational data of the intervention process meets expectations in order to assess the effectiveness of the corrective control instructions and output a corrective intervention effect report.

[0089] S95. Perform correlation analysis between the revised intervention effect report and the abnormal cause diagnosis report to optimize the expert rule knowledge base and the revision instruction generation strategy in step S8.

[0090] The beneficial effects of this invention due to the adoption of the above-mentioned technology are as follows.

[0091] 1. In the process of controlling the quality of fly ash in thermal power units, this invention constructs a dynamic prediction model based on multi-source heterogeneous industrial big data. By deeply mining and extracting features from massive historical and real-time data, it can accurately predict key quality indicators of fly ash such as unburned carbon content, particle size distribution, and pozzolanic activity in future periods. This solves the core problem of unstable fly ash quality caused by frequent fluctuations in coal quality and changes in unit operating conditions caused by traditional control methods that rely on manual experience or static parameter settings. It significantly improves the foresight and initiative of quality control and provides a solid foundation for the high-value resource utilization of fly ash.

[0092] 2. In the process of generating and optimizing control strategies, this invention introduces a multi-objective optimization engine that combines an expert rule base and a metaheuristic algorithm. It also uses a digital twin model to perform closed-loop simulation and pre-evaluation of the initial control commands. This enables the coordinated optimization of the two core objectives of high-quality fly ash production rate and energy consumption per ton of ash collection. This solves the problem that traditional single control strategies cannot balance the contradiction between quality and efficiency, avoids the blindness and trial-and-error costs of adjusting control parameters, ensures the scientific and economical nature of control commands, and maximizes the efficiency of system operation.

[0093] 3. During the long-term operation and adaptive optimization of the system, this invention utilizes an online incremental learning algorithm to periodically fine-tune the parameters and correct the structure of the prediction model based on the latest operating data generated after system adjustment. This solves the problem of continuous decay in prediction accuracy caused by system characteristic drift due to factors such as equipment aging and environmental changes in traditional models. It ensures the long-term effectiveness and robustness of the control algorithm, enabling the entire control system to have the ability to continuously learn and evolve autonomously, thus adapting to constantly changing production conditions over a long period of time. Attached Figure Description

[0094] Figure 1 This is a flowchart illustrating the present invention.

[0095] Figure 2 This is a schematic diagram of the original working condition data acquisition and processing flow of the present invention.

[0096] Figure 3 This is a schematic diagram illustrating the construction and usage process of the predictive model reflecting the dynamic characteristics of dust removal equipment in thermal power units according to the present invention.

[0097] Figure 4 This is a flowchart illustrating the conversion of control commands based on prediction results according to the present invention.

[0098] Figure 5 This is a flowchart illustrating the key performance indicator monitoring process of the present invention. Detailed Implementation

[0099] Example 1: As Figure 1As shown, the present invention provides a method for controlling the quality of fly ash from thermal power units, comprising the following steps:

[0100] S1. Collect historical operating data and real-time operating condition data of thermal power units, and form a unified raw operating condition data package after time alignment and preliminary cleaning;

[0101] S2. Clean, integrate, and standardize the original working condition data packets to generate a standardized dataset for model building;

[0102] S3. Based on standardized datasets, construct and solidify a predictive model that can reflect the dynamic characteristics of the system, and generate prediction results for key quality indicators of fly ash in the future through the predictive model.

[0103] S4. Based on the prediction results and combined with the preset processing rules, generate control instructions including electric field operation, combustion organization, delivery path and flue gas conditioning.

[0104] S5. Send control commands to the corresponding actuators in the thermal power unit and execute them to dynamically adjust the operating status of the system.

[0105] S6. Based on the new operating status data generated after the adjustment, the internal parameters of the prediction model are updated using an online incremental learning algorithm to continuously optimize the prediction accuracy of the model.

[0106] S7. Monitor key response data of the system in real time and generate a control effect evaluation report containing key performance indicators;

[0107] S8. When a key indicator in the assessment report shows an anomaly, an early warning mechanism is triggered and corrective control instructions for handling the anomaly are generated.

[0108] S9. Send the correction control command to the system's actuator for rapid intervention and closed-loop iterative adjustment;

[0109] S10. Optimal control parameters, system response data, and evaluation results are stored in a structured manner and output to the monitoring platform.

[0110] S1 includes the following steps:

[0111] S11. Collect historical operating data and real-time operating condition data from the distributed control system, fuel management information system, laboratory information management system and environmental monitoring system of thermal power units through open platform communication protocols, application programming interfaces or direct database connection;

[0112] S12. Through the interface of the distributed control system of the thermal power unit, core operating parameters reflecting the real-time operating status of the unit are collected in high density at a sampling frequency of 1 second. These include boiler and combustion-related data such as unit load, main steam pressure, main steam temperature, total coal feed, total primary air volume, total secondary air volume, furnace outlet oxygen content, furnace characteristic temperature, and burner tilt angle of each layer; pulverizing system-related data such as operating current of each coal mill, coal mill outlet temperature, and primary air-coal mixture concentration; and dust removal equipment-related data such as electrostatic precipitator inlet flue gas temperature, inlet flue gas flow rate, primary side voltage and primary side current of each electric field.

[0113] S13. Through the interface between the fuel management information system and the laboratory information management system, collect detailed quality analysis data of the coal fed into the furnace at a batch-to-day frequency, including ash content, moisture, volatile matter, Hardgrove grindability index, and the content of major oxides, including silicon dioxide, aluminum oxide, and ferric oxide; and through the interface of the laboratory information management system, collect the quality data of fly ash after manual sampling and testing at an hourly frequency per shift, as the target value for subsequent model training, including the unburned carbon content of fly ash in each ash hopper, the characteristic particle size measured by a laser particle size analyzer, and the pozzolanic activity index measured according to the activity index test method in Appendix C of the national standard GB / T18736-2017.

[0114] S14. The high-frequency operating condition data collected in S12 and the low-frequency quality data collected in S13 are timestamped. That is, the timestamps of the low-frequency coal quality and ash quality data are associated with all the high-frequency operating condition data within their effective time period. At the same time, the integrated dataset is preliminarily cleaned, and a small number of missing values ​​are filled in using the Lagrange interpolation method, and obvious outliers caused by sensor failure are removed.

[0115] S15. The historical and real-time data, which have undergone the above processing and contain complete operating condition characteristics and corresponding fly ash quality labels, are integrated into a structured "preprocessed operating condition data package" and passed to step S2 for subsequent standardization and feature engineering.

[0116] Step S2 includes the following steps:

[0117] S21. Obtain the preprocessed operating condition data packet generated in step S1;

[0118] S22. Normalize the preprocessed data packets to generate a normalized feature matrix; the normalization process uses the following formula:

[0119] ;

[0120] in, The normalized data has a value range between [0,1], eliminating the influence of different physical dimensions. These are the raw data values, representing the instantaneous measurement value of a specific operating condition parameter. This represents the minimum value of the corresponding operating condition parameter in the historical dataset. This represents the maximum value of the corresponding operating parameters in the historical dataset. In this way, operating parameters of different units and magnitudes, such as boiler load, flue gas temperature, and electric field voltage, are processed into dimensionless values, providing a comparable data foundation for subsequent comprehensive analysis of multidimensional features and model training.

[0121] S23. Multiple feature selection algorithms are combined for screening. Pearson correlation coefficient analysis is used to remove weakly correlated features with a correlation of less than 0.1 with the target variable. An embedded feature selection is performed using a L1 regularized LASSO regression model to penalize the feature coefficients that contribute little to the target variable to zero. This achieves automated feature dimensionality reduction, extracts key features that are strongly correlated with fly ash quality from the normalized feature matrix, and generates a feature optimization dataset.

[0122] S24. Perform principal component analysis on the feature optimization dataset to reduce dimensionality, extract the principal components with a cumulative variance contribution rate of 95% as core variables, eliminate data redundancy and extract core variables to generate a standardized dataset.

[0123] S25. Verify the data quality and consistency of the standardized dataset and output a data quality verification report.

[0124] Step S3 includes the following steps:

[0125] S31. Divide the standardized dataset output in step S2 into a training set, a validation set, and a test set according to the time series in an 8:1:1 ratio for subsequent model training, tuning, and evaluation.

[0126] S32. For unburned carbon content, an indicator with strong time-series dependence, a long short-term memory network model is constructed. The input features are boiler load, air-coal ratio, furnace characteristic temperature, furnace outlet oxygen content, and volatile matter in the coal fed into the furnace. The model includes an input layer with 5 input features, two hidden layers (128 neurons per layer), a dropout layer (dropout rate p=0.2), and an output layer. The model output is the predicted unburned carbon content for the next 30 minutes.

[0127] S33. Construct a gradient boosting decision tree model for the characteristic particle size D50 of fly ash. The input features are boiler load, current of each coal mill, concentration of primary air-coal mixture, furnace characteristic temperature and Hardgrove grindability index of coal fed into the furnace. The model output is the predicted particle size D50 value for the next 30 minutes.

[0128] S34. For the volcanic ash activity index, a gradient boosting decision tree model is constructed. The input features are the furnace characteristic temperature, the oxygen content at the boiler outlet, and the content of the main oxides in the ash of the coal entering the furnace. The model output is the predicted volcanic ash activity index for the next 30 minutes.

[0129] S35. Using the training set, train the three initial prediction models constructed in S32, S33, and S34 respectively. During the training process, use the validation set and adopt a grid search combined with cross-validation to optimize the hyperparameters of the long short-term memory network model and the key hyperparameters such as the number and depth of the gradient boosting decision tree. Use the test set to perform a final performance evaluation on each optimized model and generate model accuracy evaluation data including root mean square error and mean absolute error. After confirming that the mean absolute percentage error of the model is lower than the preset accuracy requirement of 0.5%, solidify it as the final prediction model.

[0130] S36. The fly ash production rate is calculated in real time using a formula. The formula for calculating the fly ash production rate is as follows:

[0131] ;

[0132] in, This provides a theoretical fly ash production rate per unit time, serving as a basis for decision-making regarding graded conveying and ash silo capacity management. The real-time coal feed rate is obtained from the distributed control system of the thermal power unit. This refers to the ash content of the coal fed into the furnace, which is obtained from the fuel management system. The unburned carbon content in fly ash is provided by the prediction model in step S32, thereby calculating the theoretical ash production rate for the next 30 minutes.

[0133] S37. During the real-time operation phase, the operating condition data of the current moment and the previous period are input into the three final prediction models solidified in S35. Combined with the calculation in S36, a comprehensive prediction result data package containing four key indicators in the next 30 minutes, namely unburned carbon content, particle size distribution D50, volcanic ash activity index, and ash production rate, is finally output and passed to step S4 for control decision.

[0134] Step S4 includes the following steps:

[0135] S41. Obtain the fly ash quality prediction results for the next 30 minutes generated in step S3, including particle size distribution, unburned carbon content, pozzolanic activity index, and ash production rate.

[0136] S42. Generate an initial control instruction set based on a multi-objective optimization rule engine. The rule engine includes at least one strategy generation layer based on an expert rule base and a parameter optimization layer based on a metaheuristic algorithm. The strategy generation layer matches the prediction results obtained in S41 with a pre-defined quality control expert rule knowledge base to determine the adjustment direction and approximate range for each control variable, forming a preliminary strategy combination. The parameter optimization layer uses the adjustment range given by the strategy generation layer as the search space and the objective function J as the optimization objective.

[0137] ;

[0138] in, To predict the yield of high-grade fly ash, for example, fly ash with an unburned carbon content of less than 3% is considered high-grade fly ash. The predicted energy consumption per ton of ash collection; and This is a weighting coefficient that can be adjusted by operators based on market demand and energy-saving targets, and For example, setting and The values ​​are 0.7 and 0.3 respectively, indicating a greater focus on improving the output of high-quality fly ash. The particle swarm optimization algorithm searches within the feasible region of various control variables such as electric field voltage and burner swing angle to find the combination of control parameters that minimizes the objective function J, forming the initial control command. The parameters of the particle swarm optimization algorithm are set as follows: population size of 100, maximum number of iterations of 200, inertia weight of 0.8, individual learning factor c1 of 1.5, and social learning factor c2 of 1.5.

[0139] The preset quality control expert rule knowledge base includes the following rules: R1, if the unburned carbon content is predicted to exceed 4.5%, an enhanced burnout strategy is triggered, initially setting the upper secondary air opening to increase by 2-5% and the burner tilt angle to adjust upwards by 3-6 degrees; R2, if the dust particle size D50 is predicted to decrease and unburned carbon content to decrease, resulting in increased resistivity, an anti-back corona strategy is triggered, initially setting the final stage electric field voltage to decrease by 5-8kV, and an anti-wall adhesion strategy is triggered, initially setting the rapping cycle to shorten by 10-15%; R3, if the system detects changes in coal quality and predicts that the dust resistivity will exceed... If this is the case, the conditioning injection strategy is triggered, and the initial injection amount of flue gas conditioning agent SO3 is set to a base value based on the predicted resistivity calculation.

[0140] S43. Perform digital twin model simulation on the initial control commands to generate simulation operating condition response data. The digital twin model is a hybrid model, including a boiler combustion module and an electrostatic precipitator module. The boiler combustion module is a lumped parameter model based on thermodynamic equilibrium and chemical reaction kinetics. It takes as input combustion fine-tuning commands, such as air-coal ratio and burner sway angle changes, and outputs the changing trends of furnace temperature field and flue gas composition. The core of the electrostatic precipitator module is the electrostatic precipitator collection efficiency calculation formula described in step S46, where the key parameter is dust driving velocity. Instead of being a constant, it is dynamically calculated by a small neural network with two hidden layers, 16 neurons in each layer, and using the ReLU activation function, based on the predicted fly ash particle size, unburned carbon content, and input electric field voltage. In this way, by inputting the initial control command into the digital twin model, the dynamic response of the main state parameters of the system in the next 30 minutes can be quickly simulated, generating simulation operating condition response data.

[0141] S44. Analyze the deviation between the simulation working condition response data and the preset fly ash quality target parameter of less than 4.0% unburned carbon content. If the deviation exceeds the allowable range of ±5%, generate an instruction to optimize the correction amount.

[0142] S45. Based on the instruction optimization correction amount, correct the initial control instruction and output the final control instruction. The final control instruction is a structured instruction set, which includes at least the boiler combustion organization fine-tuning instruction, the electrostatic precipitator operation optimization instruction, the staged conveying path switching instruction, and the flue gas conditioning agent injection amount adjustment instruction. For example, the boiler combustion organization fine-tuning instruction adopts the key values ​​of {"burner swing angle_A layer": "+5 degrees"} and {"secondary air damper_C layer": "+3%"}, and the electrostatic precipitator operation optimization instruction adopts the key values ​​of {"electric field voltage_zone 3": "58kV"}, {"electric field voltage_zone 4": "52kV"} and {"rapping cycle_full"}. The key value for the graded conveying path switching instruction is {"Department": "15 minutes"}. The instruction is generated based on a multi-dimensional quality grading threshold matrix. For example, if the prediction result satisfies {'unburned carbon content' < 3.0%, 'activity index' > 85%}, the instruction is {"Conveying path": "No. 1 premium ash silo"}. Otherwise, if {'unburned carbon content' < 5.0%} is satisfied, the instruction is {"Conveying path": "No. 2 ordinary ash silo"}. In other cases, the instruction is {"Conveying path": "No. 3 waste silo"}. The flue gas conditioning agent injection volume adjustment instruction adopts the key value {"SO3 injection pump frequency": "35Hz"}.

[0143] S46. Calculate the theoretical collection efficiency of the electrostatic precipitator to provide a theoretical basis for optimizing the electric field voltage and current. The formula for calculating the collection efficiency of the electrostatic precipitator is:

[0144] ;

[0145] in, The theoretical dust collection efficiency of an electrostatic precipitator is a core indicator for evaluating its performance. This refers to the total area of ​​the dust collection plates, which is an inherent design parameter of the equipment. It extracts the flue gas flow rate from real-time flue gas velocity data through the electric field. The dust ingress velocity is positively correlated with various factors such as dust particle size, electric field strength, and flue gas characteristics, and is a key variable for control optimization; thus, the control system can adjust the electric field voltage to change the dust ingress velocity. Thus achieving the target capture efficiency It provides core physical model support to ensure the scientific validity and effectiveness of control commands.

[0146] Step S5 includes the following steps:

[0147] S51. The final control command is sent to the distributed control system of the thermal power unit and the PLC controller actuator module through the industrial control network.

[0148] S52. The actuator module controls the high-voltage power supply and rapping control system of the electrostatic precipitator, and optimizes the voltage of each electric field and the rapping cycle.

[0149] S53. The actuator module sends a combustion fine-tuning command to the boiler control system to fine-tune combustion organization parameters such as the air-coal ratio and burner tilt angle.

[0150] S54. The actuator module controls the path switching valve of the pneumatic ash conveying system to switch the graded conveying path according to the predicted fly ash quality.

[0151] S55. Collect device status feedback data in real time during the command execution process and generate an execution log data package containing timestamps and execution results.

[0152] Step S6 includes the following steps:

[0153] S61. Obtain the execution log data packet generated in step S5, the new operating status data after system adjustment, and the fly ash quality data obtained from the laboratory information management system corresponding to the new data timestamp;

[0154] S62. Using the new actual test data as the true label, calculate the deviation between it and the model's prediction results at the corresponding time, and generate a model prediction error analysis report that includes root mean square error and mean absolute percentage error.

[0155] S63. Based on the model prediction error analysis report, determine whether the preset model update triggering conditions are met, and select the corresponding update strategy according to the model type. The triggering conditions are set as follows: the weekly average MAPE of any model increases by more than 5% compared to the previous week or the prediction error exceeds the allowable range of 10% for three consecutive days.

[0156] S64. When the triggering condition is met, the prediction model is retrained using new data from the past 7 days through an online incremental learning algorithm that matches the model type, and an updated prediction model is generated.

[0157] The model update strategy is as follows:

[0158] C1. For the Long Short-Term Memory network model in S32, a periodic fine-tuning strategy is adopted, which means adjusting the existing model weights. As the starting point for training, a dataset containing the most recent 30 days of data, including new data, is used for retraining with a small, fine-tuned learning rate η=0.0001. The model weight update formula is as follows:

[0159] ;

[0160] in, The updated model weight matrix represents the model's state after learning new knowledge. This represents the weight matrix of the current model. This refers to fine-tuning the learning rate, a hyperparameter that controls the step size for each update and determines how quickly the model learns new data. Represents the loss function Weights The gradient represents the direction of change of the prediction error relative to the weight parameters. After 5-10 more rounds of training, the model is prevented from overfitting to new data, while retaining the knowledge learned from historical data.

[0161] C2. For the gradient boosting decision tree model in steps S33 and S34, a residual correction reinforcement learning strategy is adopted, that is, the new data is first input into the old gradient boosting decision tree model. In the middle, the predicted value is obtained. ; Calculate the true value of the new data Compared with the predicted value residuals between Then train and correct the model: using the working condition characteristics of the new data as input, and the residuals... Train a new, smaller gradient boosting decision tree model for the target label. It contains only 10-20 decision trees, and the last updated model The prediction logic is a combination of the old model and the residual correction model:

[0162] ;

[0163] in Using the input operating data, the new model, while retaining the capabilities of the old model, specifically corrects its prediction bias under the current operating conditions.

[0164] S65. Store the updated prediction model and its new version information, training date, dataset period, and other metadata in the corresponding model library for subsequent deployment and version tracking.

[0165] Step S7 includes the following steps:

[0166] S71. Collect key response data of the system after the execution of control commands using the distributed control system of thermal power units;

[0167] S72. Calculate key performance indicators, including high-grade fly ash production rate, fly ash comprehensive utilization value index, energy consumption per ton of ash collection, and environmental emission compliance rate.

[0168] S73. Based on a multi-dimensional key performance indicator threshold comparison algorithm, a control effect evaluation report is generated according to the energy consumption per ton of ash collection. The formula for calculating the energy consumption per ton of ash collection is as follows:

[0169] ;

[0170] in, Energy consumption per ton of ash collection is a core indicator for evaluating the economic efficiency of a dust removal system. The total operating power of the electrostatic precipitator includes the power consumption of the high-voltage power supply, rapping motor, etc. This indicates the operating power of the pneumatic ash conveying system. It is the actual fly ash collection rate per unit time, calculated based on the uniformity of collection volume in each ash hopper and ash discharge records. This quantifies the energy cost required to obtain a unit mass of fly ash. The energy consumption per ton of ash collection can be used to evaluate whether the control strategy can ensure the quality of fly ash while also assessing whether the overall energy saving and consumption reduction of the system operation have been achieved.

[0171] S74. When any key performance indicator exceeds the preset safety or economic threshold range, it shall be marked as an abnormal state in the evaluation report.

[0172] S75. Automatically generate and output evaluation reports to the power plant's intelligent environmental monitoring platform at a preset cycle of once every half hour.

[0173] Step S8 includes the following steps:

[0174] S81. When abnormal data is marked in the evaluation report of step S7, an early warning mechanism is triggered. The early warning trigger condition is determined by the fly ash quality deterioration index. Sure:

[0175] ;

[0176] in, This indicates the fly ash quality deterioration index; an early warning is triggered when its value exceeds a preset threshold. This indicates the real-time predicted or measured value of the unburned carbon content in fly ash. This indicates the target control value for unburned carbon content. Indicates the real-time predicted or measured value of fly ash fineness; This represents the target control value for the fineness of fly ash. and The weighting coefficient is set according to the degree of influence of unburned carbon and fineness on the comprehensive utilization value of fly ash under different working conditions. In this way, by transforming multi-dimensional quality deviations into a single early warning trigger, the early warning mechanism becomes more sensitive and accurate.

[0177] S82. Identify potential sources through root cause diagnosis algorithms, preferably using Bayesian networks for diagnosis. This involves pre-constructing conditional probability relationships between cause nodes such as coal quality mutations, coal mill malfunctions, burner blockages, and electric field flashover, and result nodes such as abnormal fly ash quality. By inputting real-time abnormal phenomena, the most likely fault source can be inferred in reverse, such as coal quality mutations, insufficient coal mill output, and specific electric field failures, and an abnormal cause diagnosis report can be generated.

[0178] S83. Based on the abnormal cause diagnosis report and matching the preset expert rule knowledge base, generate a draft of the correction control instruction;

[0179] S84. Review the potential impact of the draft control instructions through the control strategy safety simulation module. After confirming that there are no conflicts, output the final corrected control instructions. The instructions include starting the emergency ash discharge procedure, triggering electric field enhanced rapping, and switching abnormal quality fly ash to the waste silo.

[0180] S85. Record the complete process of early warning triggering, cause diagnosis, instruction generation and execution to the system event log for traceability and analysis.

[0181] Step S9 includes the following steps:

[0182] S91. Send the final corrected control command generated in step S8 to the actuator module of the distributed control system of the thermal power unit.

[0183] S92. The actuator module schedules and executes corrective control tasks according to preset safety and process priorities, such as switching waste bins > strengthening rapping > requesting load reduction.

[0184] S93. Through rapid response closed-loop adjustment, until the monitoring data shows that the fly ash quality deterioration index has recovered to within the threshold, and generate real-time operation data of the intervention process;

[0185] S94. Verify whether the real-time operational data of the intervention process meets expectations, evaluate the effectiveness of the corrective control instructions, and output a report on the corrective intervention effect.

[0186] S95. Perform correlation analysis between the revised intervention effect report and the abnormal cause diagnosis report to optimize the expert rule knowledge base and the revision instruction generation strategy in step S8.

[0187] Step S10 includes the following steps:

[0188] S101. Link and integrate the optimal control parameter set, system response data package, model evaluation report and correction intervention report using foreign keys to form a closed-loop event data archive.

[0189] S102. Write the closed-loop event data archive into the InfluxDB and PostgreSQL databases of the power plant's intelligent environmental monitoring platform in a structured format;

[0190] S103. On the user interface of the power plant's intelligent environmental monitoring platform, generate a real-time fly ash quality dashboard, control benefit analysis charts, and a system operation history traceability interface.

[0191] S104. Regularly archive and back up historical data to provide historical dataset support for subsequent long-term system performance evaluation and algorithm model iteration.

Claims

1. A method for quality control of fly ash from thermal power units, characterized in that, Includes the following steps: S1. Collect historical operating data and real-time operating condition data of thermal power units, and form a unified raw operating condition data package after time alignment and preliminary cleaning; S2. Clean, integrate, and standardize the original working condition data package to generate a standardized dataset for model building; S3. Based on the standardized dataset, construct and solidify a prediction model that reflects the dynamic characteristics of the system, and generate a comprehensive prediction result for fly ash quality through the prediction model, specifically including the following steps: S31. Divide the standardized dataset output by S2 into training set, validation set and test set according to time series. S32. Construct a long short-term memory network model for the unburned carbon content index; S33. Construct a gradient boosting decision tree model for the characteristic particle size index of fly ash; S34. Construct a gradient boosting decision tree model for the volcanic ash activity index. S35. The initial prediction model is trained using the training set, and the hyperparameters of the model are optimized by grid search combined with cross-validation. Finally, the performance of the optimized model is evaluated using the test set and solidified into the final prediction model. S36. The fly ash production rate is obtained through real-time calculation, providing a basis for decision-making regarding graded conveying and ash silo capacity management; the formula for calculating the ash production rate is: ， in, This represents the theoretical fly ash production rate per unit time. To measure the amount of coal fed into the furnace in real time, It refers to the ash content of the coal fed into the furnace. The unburned carbon content of fly ash is provided by the prediction model in S32. S37. During the real-time operation phase, the operating data is input into the three final prediction models that have been solidified, and combined with the calculation results of S36, a comprehensive prediction result data package of fly ash quality containing four key indicators in the next 30 minutes: unburned carbon content, particle size distribution, pozzolanic activity index, and ash production rate is output. S4. Based on the prediction results and in conjunction with the preset processing rules, generate control commands including electric field operation, combustion organization, delivery path and flue gas conditioning. S5. Send the control command to the corresponding actuator in the thermal power unit and execute it to dynamically adjust the operating status of the system; S6. Based on the new operating status data generated after adjustment, the internal parameters of the prediction model are updated using an online incremental learning algorithm to continuously optimize the prediction accuracy of the model; S7. Monitor key response data of the system in real time and generate a control effect evaluation report containing key performance indicators; S8. When a key indicator in the assessment report becomes abnormal, an early warning mechanism is triggered and a corrective control instruction for handling the abnormality is generated. S9. Send the correction control command to the system's actuator for rapid intervention and closed-loop iterative adjustment; S10. Optimal control parameters, system response data, and evaluation results are stored in a structured manner and output to the monitoring platform.

2. The method for controlling the quality of fly ash from thermal power units according to claim 1, characterized in that: Step S1 includes the following steps: S11. Collect historical operating data and real-time operating condition data from the distributed control system, fuel management information system, laboratory information management system and environmental monitoring system of thermal power units through open platform communication protocols, application programming interfaces or direct database connection; S12. Collect core operating parameters that reflect the real-time operating status of the unit at a high sampling frequency, including boiler and combustion related data, pulverizing system related data and dust removal equipment related data; S13. The quality analysis data of the coal fed into the furnace and the quality data of fly ash after manual sampling and testing are obtained at a low acquisition frequency and used as the target values ​​for subsequent model training. S14. Timestamp align the high-frequency operating condition data with the low-frequency quality data, and use Lagrange interpolation to fill in missing values ​​and remove outliers. S15. Integrate the processed data into a structured preprocessed working condition data package and pass it to step S2.

3. The method for controlling the quality of fly ash from thermal power units according to claim 1, characterized in that: Step S2 includes the following steps: S21. Obtain the original operating condition data packet generated in step S1; S22. Normalize the original operating condition data packet to eliminate the influence of different physical dimensions and generate a normalized feature matrix; the normalization process uses the following formula: ， in, For the normalized data, The original data values, The minimum value in the historical dataset. The maximum value in the historical dataset; S23. Pearson correlation coefficient analysis and L1 regularized LASSO regression model were used to combine feature screening to extract key features that are strongly correlated with fly ash quality and generate a feature optimization dataset. S24. Perform principal component analysis on the feature optimization dataset to reduce dimensionality, extract principal components whose cumulative variance contribution rate reaches a preset threshold, and generate a standardized dataset. S25. Verify the data quality and consistency of the standardized dataset and output a data quality verification report.

4. The method for controlling the quality of fly ash from thermal power units according to claim 1, characterized in that: Step S4 includes the following steps: S41. Obtain the comprehensive prediction result of fly ash quality for the next 30 minutes generated in step S3. S42. Generate initial control instructions based on a multi-objective optimization rule engine that includes a policy generation layer and a parameter optimization layer; the parameter optimization layer uses an objective function... To optimize the objective function, a particle swarm optimization algorithm is used to search within the feasible region of the control variables. The expression is: ， in, To predict the yield of high-grade fly ash, For the predicted energy consumption per ton of ash collection, and These are the weighting coefficients; S43. Input the initial control command into a digital twin model consisting of a boiler combustion module and an electrostatic dust removal module for simulation, and generate simulation operating condition response data; S44. Analyze the deviation between the simulation working condition response data and the preset fly ash quality target parameters. If the deviation exceeds the allowable range, generate an instruction to optimize and correct the amount. S45. Based on the instruction, optimize the correction amount to correct the initial control instruction and output a structured final control instruction; S46. Provide a theoretical basis for the electrostatic precipitator operation optimization instruction in the final control instruction, which is based on the following formula for calculating the electrostatic precipitator collection efficiency: ， in, For theoretical capture efficiency, It is the total area of ​​the dust collection plates. It is the flue gas flow rate. It is the dust ingress speed.

5. The method for controlling the quality of fly ash from thermal power units according to claim 1, characterized in that: Step S5 includes the following steps: S51. Send the final control command to the distributed control system and PLC controller of the thermal power unit through the industrial control network. S52. Control the high-voltage power supply and rapping control system of the electrostatic precipitator to optimize the voltage of each electric field and the rapping cycle. S53. Send a combustion fine-tuning command to the boiler control system to fine-tune the combustion organization parameters of the air-coal ratio and burner sway angle; S54. Control the path switching valve of the pneumatic ash conveying system to switch the graded conveying path according to the predicted fly ash quality. S55. Collect device status feedback data in real time during the command execution process and generate an execution log data package containing timestamps and execution results.

6. The method for controlling the quality of fly ash from thermal power units according to claim 5, characterized in that: Step S6 includes the following steps: S61. Obtain the execution log data package generated in step S5, the new operating status data after system adjustment, and the corresponding new fly ash quality test data. S62. Using the new test data as the true label, calculate the deviation between the new test data and the model prediction results, and generate a model prediction error analysis report. S63. Based on the model prediction error analysis report, determine whether the preset model update triggering conditions are met; S64. When the triggering condition is met, the prediction model is retrained using new data through an online incremental learning algorithm that matches the model type; the model update strategy includes: C1. For the Long Short-Term Memory (LSTM) network model, a periodic fine-tuning strategy is adopted, and its weight update formula is as follows: ， in, The updated weight matrix, This is the current weight matrix. For learning rate, This represents the gradient of the loss function with respect to the weights. C2. For the gradient boosting decision tree model, the reinforcement learning strategy with residual correction is adopted. The prediction logic of the new model is as follows: ， in, The updated predicted value, These are the predicted values ​​from the old model. The predicted values ​​of the corrected model trained on the new data residuals; S65. Store the updated prediction model, its version information, training date, and other metadata into the model library.

7. The method for controlling the quality of fly ash from thermal power units according to claim 1, characterized in that: Step S7 includes the following steps: S71. Collect key response data of the system after the execution of control commands using the distributed control system of thermal power units; S72. Calculate key performance indicators, including high-grade fly ash production rate, fly ash comprehensive utilization value index, energy consumption per ton of ash collection, and environmental emission compliance rate. S73. A control effect evaluation report is generated based on a multi-dimensional KPI threshold comparison algorithm. The report evaluates the economic efficiency of the dust removal system by calculating the energy consumption per ton of ash collection. The formula for calculating the energy consumption per ton of ash collection is as follows: ， in, Energy consumption per ton of ash collection This refers to the total operating power of the electrostatic precipitator itself. This refers to the operating power of the pneumatic ash conveying system. It is the actual fly ash collection rate per unit time; S74. When any key performance indicator exceeds the preset safety or economic threshold range, it is marked as an abnormal state in the evaluation report. S75. Automatically generate and output the assessment report to the power plant's intelligent environmental monitoring platform according to a preset cycle.

8. The method for controlling the quality of fly ash from thermal power units according to claim 1, characterized in that: Step S8 includes the following steps: S81. When abnormal data is marked in the evaluation report of step S7, an early warning mechanism is triggered. The triggering basis of the early warning mechanism is the fly ash quality deterioration index. The calculation formula is as follows: ， in, The fly ash quality deterioration index. and These are the real-time values ​​for unburned carbon content and fineness, respectively. and For its target value, and These are the weighting coefficients; S82. Use Bayesian networks for root cause diagnosis to infer the most likely source of failure and generate an abnormal cause diagnosis report. S83. Based on the abnormal cause diagnosis report and matching it with a preset expert rule knowledge base, generate a draft of the correction control instruction; S84. Review the potential impact of the revised control instruction draft through the control strategy safety simulation module, and output the final revised control instruction after confirming that there are no conflicts. S85. Record the complete process of early warning triggering, cause diagnosis, instruction generation and execution to the system event log.

9. The method for controlling the quality of fly ash from thermal power units according to claim 8, characterized in that: Step S9 includes the following steps: S91. Send the modified control command generated in step S8 to the actuator module of the distributed control system of the thermal power unit; S92. The actuator module schedules and executes correction control tasks according to preset safety and process priorities; S93. Through rapid response closed-loop adjustment, until the monitoring data shows that the fly ash quality deterioration index has recovered to within the threshold, and real-time operation data of the intervention process is generated. S94. Verify whether the real-time operation data of the intervention process meets expectations, so as to evaluate the effectiveness of the corrective control instructions and output a corrective intervention effect report. S95. The correlation analysis between the correction intervention effect report and the abnormal cause diagnosis report is used to optimize the expert rule knowledge base and correction instruction generation strategy in step S8.

Images