Thermal power generating unit control strategy optimization method and system based on combination of DCS and AGC

Abstract

The invention relates to the technical field of distributed power generation, solves the technical problems that in the prior art, actual working conditions cannot be accurately matched, time sequence dynamic characteristics are difficult to capture, meanwhile, energy evaluation and feed-forward compensation are lack of comprehensive consideration, and control deviation and instability are likely to occur, and particularly relates to a thermal power generating unit control strategy optimization method and system based on combination of DCS and AGC. The method comprises the steps that original data of a thermal power generating unit are collected, a working condition model is constructed based on the original data, a working condition model parameter set is generated, and the original data comprise original DCS historical operation data and an AGC instruction sequence. The working condition identification accuracy is greatly improved, the prediction advancement and accuracy are effectively improved, the unit can sense the parameter change trend in advance, the control problem caused by parameter mutation is avoided, and the comprehensiveness and accuracy of energy evaluation are improved.

Description

Technical Field

[0001] This invention relates to the field of distributed generation technology, and in particular to a method and system for optimizing control strategies of thermal power units based on the combination of DCS and AGC. Background Technology

[0002] Distributed generation refers to small power generation units configured near the user side. It is an optimization method for thermal power unit control strategy that combines DCS (Distributed Control System) and AGC (Automatic Generation Control). Through DCS, local control is achieved, and AGC coordinates global output, optimizing the coordinated operation of thermal power units and distributed energy. Existing technologies mostly adopt PID control, model predictive control, or neural network algorithm control to achieve power distribution and dynamic regulation.

[0003] However, existing operating condition identification technologies are crude and cannot accurately match actual operating conditions. Prediction methods are mostly single-step or simple models, making it difficult to capture time-series dynamic characteristics. At the same time, energy assessment and feedforward compensation lack comprehensive consideration, which can easily lead to control deviations and instability. For example, in extremely cold weather, the cooling water temperature is very low, close to 0 degrees Celsius, which will reduce the turbine exhaust pressure and the steam's ability to do work in the turbine, thereby reducing the unit's load output. At this time, fixed feedforward compensation cannot detect this load change deviation caused by ambient temperature in time, resulting in a large difference between the actual load and the set load. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for optimizing the control strategy of thermal power units based on the combination of DCS and AGC. This solves the technical problems of existing technologies, such as the inability to accurately match actual operating conditions, the difficulty in capturing time-series dynamic characteristics, and the lack of comprehensive consideration of energy assessment and feedforward compensation, which easily leads to control deviations and instability.

[0005] To solve the above technical problems, the present invention provides the following technical solution: applied to the above-mentioned thermal power unit control strategy optimization method based on the combination of DCS and AGC, the steps of the method are: collecting the raw data of the thermal power unit, constructing the operating condition model based on the raw data, and generating the operating condition model parameter set. The raw data includes the original DCS historical operating data and AGC instruction sequence. The pressure setpoint is determined based on historical data thresholds, real-time status data of thermal power units is collected, and operating condition data is processed to obtain operating condition labels. Dynamic feedforward compensation is generated based on the pressure setpoint, operating condition labels, and operating condition model parameter set. The target load and emission constraints of the thermal power unit are obtained, and an optimization processing space is generated. Candidate control variable groups are selected in the optimization processing space, and weighted adaptive processing is performed to obtain fitness values. Based on the fitness values, global optimization decoding is performed to obtain the optimal instruction set. The real-time dataset of the thermal power unit is obtained. The real-time status data and the real-time dataset are subjected to energy storage fuzzy processing to generate a coefficient set and a component set. Based on the coefficient set, component set and optimal instruction set, instruction synthesis processing is performed to obtain the main control instruction set. The current unit operation data is obtained based on the master control instruction set. The model is evaluated and updated based on the current unit operation data to obtain the model update flag. The model is then retrained based on the model update flag.

[0006] Preferably, the construction of the working condition model based on the original data includes: Based on the original DCS historical operation data, an aligned variable dataset is obtained through filtering and timestamp alignment. Key performance data processing was performed on the AGC instruction sequence and alignment variable dataset to obtain historical KPI label data and historical KPI labels. Based on historical KPI tag data, the working condition model parameter set is obtained through working condition clustering.

[0007] Preferably, the operating condition data is processed to obtain operating condition labels, including: Real-time feature vectors are obtained by integrating features based on real-time status data. Based on the real-time feature vector and the parameter set of the working condition model, a standardized feature vector is generated through feature standardization processing. The standardized feature vectors and the parameter set of the working condition model are subjected to working condition matching processing to obtain working condition labels.

[0008] Preferably, a dynamic feedforward compensation amount is generated based on the pressure setpoint, operating condition label, and operating condition model parameter set, including: Multi-step dynamic prediction processing is performed based on real-time status data and operating condition model parameter set to obtain the main steam pressure prediction value and load prediction value. Based on the predicted main steam pressure, predicted load, AGC command sequence, and pressure setpoint, an estimated energy gap is generated through energy gap assessment. Dynamic feedforward processing is performed on the estimated energy gap and operating condition labels to obtain the dynamic feedforward compensation amount.

[0009] Preferably, the fitness value is obtained by weighted adaptation, including: Based on the target load and emission constraints, upper and lower limits are processed to generate an optimization space; Multiple candidate control variable groups are arbitrarily selected in the optimization processing space, and forward processing is performed on the candidate control variable groups and the parameter set of the operating condition model to obtain the predicted values ​​of key performance. The fitness value is obtained by weighting the predicted key performance values.

[0010] Preferably, global optimization decoding based on fitness values ​​includes: Population update processing is performed based on the optimization processing space to generate updated population positions; Based on the updated population position, the current global optimum is obtained through elite retention processing; Obtain all current global optimal solutions and perform optimal decoding processing to obtain the optimal instruction set.

[0011] Preferably, energy storage fuzzy processing is performed on real-time status data and real-time datasets to generate a coefficient set, including: The component set is obtained by digital filtering based on the real-time dataset; Energy storage status is assessed based on real-time status data to generate real-time energy storage margin. Dynamic fuzzy processing is performed on the real-time energy storage margin and component sets to obtain the coefficient set.

[0012] Preferably, instruction synthesis processing is performed based on the coefficient set and the optimal instruction set, including: Dynamic command generation is performed based on coefficient set and component set to obtain turbine command increment and boiler command increment; Boiler commands are synthesized based on boiler command increments, dynamic feedforward compensation, and the optimal command set to obtain boiler master control commands. The optimal instruction set and the turbine instruction increment are processed to generate a turbine main control instruction parameter set. The boiler main control instruction and the turbine main control instruction parameter set are then combined into the main control instruction set.

[0013] Preferably, the evaluation and model update are based on the current unit operating data, including: Based on the current unit operating data, feature reconstruction processing is performed to generate the current feature vector and the current operating condition label; Based on the current feature vector and the current working condition label, feature consistency is judged, and model adaptation signal is generated; The model adaptation signal is used to determine the model update and generate a model update flag.

[0014] This technical solution also provides a system for optimizing the control strategy of thermal power units based on the above-mentioned combination of DCS and AGC, the system comprising: The operating condition processing module is used to collect raw data from thermal power units, construct operating condition models based on raw data, and generate operating condition model parameter sets. The raw data includes raw DCS historical operating data and AGC instruction sequences. The feedforward compensation module is used to determine the pressure setpoint based on historical data thresholds, collect real-time status data of thermal power units, process the operating data to obtain operating condition labels, and generate dynamic feedforward compensation amount based on the pressure setpoint, operating condition labels and operating condition model parameter set. The instruction optimization module is used to obtain the target load and emission constraints of the thermal power unit, generate an optimization processing space, select candidate control variable groups in the optimization processing space, perform weighted adaptive processing to obtain fitness values, and perform global optimization decoding based on fitness values ​​to obtain the optimal instruction set. The instruction generation module is used to acquire the real-time dataset of the thermal power unit, perform energy storage fuzzy processing on the real-time status data and the real-time dataset to generate a coefficient set and a component set, and perform instruction synthesis processing based on the coefficient set, component set and optimal instruction set to obtain the main control instruction set. The model update module is used to obtain the current unit operation data based on the main control instruction set, evaluate and update the model based on the current unit operation data, obtain the model update flag, and retrain the model based on the model update flag.

[0015] By employing the above technical solution, the present invention provides a method and system for optimizing the control strategy of thermal power units based on the combination of DCS and AGC, which has at least the following beneficial effects: 1. This invention effectively solves the timing misalignment problem caused by inconsistent data sampling periods through sliding window filtering and dynamic time warping. By combining filtering and alignment, it not only preserves the local features of the original data but also eliminates time drift errors, significantly improving data quality. The use of KPI indicators that are directly linked to sliding window statistical features and power grid rules can more accurately reflect the dynamic adjustment performance of the unit. Based on the operating condition model, it realizes adaptive modeling and multi-objective optimization of operating conditions, avoiding the problem of insufficient generalization ability of a single model for complex operating conditions, and providing a more reliable data foundation and model support for the optimization of thermal power unit control strategies.

[0016] 2. This invention can quickly and accurately subdivide the unit's operating conditions into specific categories based on real-time unit status data, greatly improving the accuracy of operating condition identification. Through multi-step dynamic prediction, it can deeply mine the temporal characteristics and potential patterns in the unit's status data, effectively improving the predictive advance and accuracy. This allows the unit to perceive parameter change trends in advance, avoiding control problems caused by parameter mutations. The energy gap assessment integrates multiple data sources to comprehensively and accurately assess the unit's current energy gap status, improving the comprehensiveness and accuracy of energy assessment. The feedforward calculation process combines operating conditions, accurately selects appropriate feedforward gain, and performs amplitude limiting processing on the calculation results to ensure that the dynamic feedforward compensation can effectively make up for the energy gap while remaining within a reasonable range, improving the accuracy and stability of feedforward compensation.

[0017] 3. This invention achieves real-time and accurate prediction of key performance through an initial operating condition processing model. By combining constrained particle swarm optimization and adaptive penalty coefficient, it quickly converges to the global optimum under dynamic operating conditions, significantly shortening the processing time. It can also process data in real time, balancing load tracking accuracy and emission constraints. This solves the problems of low convergence efficiency, poor dynamic adaptability, and subjective weight allocation in traditional methods, and greatly improves load tracking accuracy and thermal power data processing efficiency.

[0018] 4. This invention accurately distinguishes between high and low frequency disturbances through frequency domain decomposition. The steam turbine focuses on high-frequency transient response, while the boiler emphasizes low-frequency trend adjustment, achieving fast and slow decoupling. Based on the energy storage margin calculation of the main steam pressure deviation, the margin is kept at one within the dead zone to avoid ineffective regulation. The fuzzy rule dynamic matching coefficient can adapt to rapid changes in operating conditions and reduce regulation conflicts. By smoothing the boiler command through first-order inertia to buffer low-frequency fluctuations and performing amplitude limiting processing on the steam turbine command, the safety of thermal power unit regulation is ensured, continuous vibration of unit equipment is avoided, and service life is effectively extended. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the thermal power unit control strategy optimization method based on the combination of DCS and AGC according to the present invention. Figure 2 This is a structural block diagram of the thermal power unit control strategy optimization system based on the combination of DCS and AGC of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0021] Example 1: Because existing technologies cannot accurately match actual operating conditions and are difficult to capture time-series dynamic characteristics, and because energy assessment and feedforward compensation lack comprehensive consideration, technical problems such as control deviations and instability are prone to occur. Please refer to [the relevant documentation / reference]. Figure 1 This embodiment provides a method and system for optimizing the control strategy of thermal power units based on the combination of DCS and AGC. It can accurately match actual operating conditions, capture the dynamic characteristics of thermal power units, comprehensively consider energy assessment and feedforward compensation, reduce control deviation and improve stability. The method includes the following steps: S1. Collect raw data from thermal power units, construct operating condition models based on raw data, and generate operating condition model parameter sets. Existing technologies mostly use simple moving averages or median filtering for data cleaning, which can easily lead to blurred signal edge features and cannot handle the time drift problem of multiple sensors. It is easy to get trapped in local optima, and a single prediction model cannot simultaneously take into account parameter prediction accuracy and multi-objective optimization requirements. For example, in practice, it cannot take into account both high efficiency and low emission requirements. To solve the above problems, the specific implementation steps are as follows: S11. Based on the original DCS historical operation data, an aligned variable dataset is obtained through filtering and timestamp alignment. The original data includes the original DCS historical operation data and AGC command sequences. The original DCS historical operation data is collected from the DCS, and the AGC command sequences are obtained from the AGC. In this step, for the input original DCS historical operation data, which includes main steam pressure, temperature, load, valve opening, coal quantity, etc., a sliding window filtering and timestamp alignment method is used. The specific steps are as follows: first, a time window of a certain length is slid across the original data sequence. For the data within the window, its average value is calculated, or other reasonable calculations can be used according to the data characteristics. For example, the median, or average value here, is used to smooth the data and eliminate data fluctuations caused by sensor noise. Then, based on the timestamp information carried by the data, the data collected by different sensors that are misaligned in time are aligned. That is, the data collected at the same time are aligned in the time dimension. Specifically, the timestamps of each sensor data are first extracted, sorted in chronological order to determine a common time reference, and then the same time interval is interpolated or truncated. Finally, the processed data is recombined in time series to obtain the timestamp-aligned data. After this processing, a time-aligned and smoothed aligned variable dataset is finally obtained, providing a consistent and reliable data foundation for subsequent models.

[0022] S12. Perform key performance data processing on the AGC instruction sequence and alignment variable dataset to obtain historical KPI label data and historical KPI labels. This step first uses a sliding window technique for key feature extraction: using a fixed time span as the window, the statistical characteristics of core parameters such as main steam pressure and unit load within the window are calculated point by point, including the arithmetic mean of parameter values, the sum of all data points divided by the number of points, the standard deviation of fluctuation being the square root of the sum of the squares of the differences between each data point and the mean divided by the number of points, and the slope of change being the difference between the first and last data points of the window divided by the time span, forming a feature vector containing features such as pressure mean, load fluctuation, and trend. Then, according to the power grid assessment rules... Calculate key performance indicators (KPIs): The regulation performance coefficient is quantified by comparing the ratio of the actual power adjustment rate to the standard rate, the integral area of ​​the actual output and the command deviation, and the weighted combination of the overshoot and adjustment time during the regulation process. The primary frequency regulation power contribution index is measured by the integral area of ​​the cumulative difference between the actual output and the theoretical required output. The final output is a complete historical dataset containing statistical characteristics and key performance indicators at each time point, as well as regulation performance coefficient and power contribution index labels arranged in time series. This provides a structured quantitative evaluation basis for subsequent operating condition clustering and model training. The KPI labels include data such as the regulation rate compliance rate, the integral area of ​​the power deviation, and the response delay time.

[0023] S13. Based on historical KPI label data, the working condition clustering process is used to obtain the working condition model parameter set. In this step, the working condition model parameter set includes working condition cluster centers, feature standardization parameters, initial dynamic prediction model, and initial working condition processing model. First, an unsupervised clustering method is used to divide the working conditions. By calculating the sum of squared Euclidean distances between each data point, the optimal number of clusters is automatically determined using the improved K-means algorithm. The historical data is divided into K typical working conditions, and the center point of each working condition is extracted as the cluster feature benchmark. At the same time, the mean and standard deviation of all feature data are recorded for subsequent standardization processing. Subsequently, dual-model training is performed. The specific process is as follows: First, a Long Short-Term Memory (LSTM) network prediction model is constructed, which is the initial dynamic prediction model. Using standardized historical data as input, the model calculates the weighted sum of feature values ​​at each time step through time step expansion and performs gating activation function operations to predict the key operating parameters of the unit at the next time step. Then, a deep neural network surrogate model is constructed, which is the initial operating condition processing model. The LSM prediction model and the deep neural network surrogate model are commonly used algorithm models, which will not be elaborated here. Using standardized historical data as input, the model undergoes nonlinear transformation through multiple hidden layers, including weight matrix multiplication and bias addition, activation function operations, which are all commonly used hidden layer processing methods, which will not be elaborated here. The output includes key indicators such as regulation performance coefficient, nitrogen oxide emission concentration, and unit efficiency. The final output includes typical operating condition cluster centers, feature standardized parameters, the initialized LSM prediction model, and the deep neural network surrogate model, providing a basic model framework and data preprocessing specifications for subsequent unit control strategy optimization based on operating condition adaptation.

[0024] This invention effectively solves the timing misalignment problem caused by inconsistent data sampling periods through sliding window filtering and dynamic time warping. By combining filtering and alignment, it not only preserves the local features of the original data but also eliminates time drift errors, significantly improving data quality. The use of KPI indicators that are directly linked to sliding window statistical features and power grid rules can more accurately reflect the dynamic adjustment performance of the unit. Based on the operating condition model, it realizes adaptive modeling and multi-objective optimization of operating conditions, avoiding the problem of insufficient generalization ability of a single model for complex operating conditions, and providing a more reliable data foundation and model support for the optimization of thermal power unit control strategies.

[0025] S2. Determine the pressure setpoint based on historical data thresholds, collect real-time status data of the thermal power unit, and process the operating data to obtain operating condition labels. Generate dynamic feedforward compensation based on the pressure setpoint, operating condition labels, and operating condition model parameter set. Existing operating condition identification is coarse and cannot accurately match actual operating conditions. Prediction methods are mostly single-step or simple models, which are difficult to capture time-series dynamic characteristics. For example, if the fuel supply is suddenly interrupted, the single-step prediction of existing technology cannot detect the rapid changes in steam parameters in time. At the same time, energy assessment and feedforward compensation lack comprehensive consideration, which is prone to control deviation and instability. For example, in severe cold, the cooling water temperature is extremely low, which will reduce the turbine exhaust pressure and change the steam's work capacity in the turbine, thereby affecting the unit's load output. At this time, fixed feedforward compensation cannot detect this load change deviation caused by ambient temperature in time, resulting in a large difference between the actual load and the set load. At the same time, key data such as main steam pressure will also fluctuate due to the influence of ambient temperature on fuel combustion efficiency, steam production, etc. Existing technology lacks comprehensive consideration of these factors, which easily leads to the problem of not being able to accurately adjust the control strategy. To solve the above problems, the specific implementation steps are as follows: S21. Based on real-time status data, feature integration processing is performed to obtain a real-time feature vector. The real-time status data includes main steam parameters, key status data of the turbine side, power grid and control system command data, and key status data of auxiliary systems, such as main steam pressure, turbine power, fan parameters, and AGC commands. In this step, a sliding window technique is first used for dynamic feature extraction: taking the current moment as the center, historical data with a fixed time span is selected as the analysis window, and the statistical characteristics of key parameters such as main steam pressure and unit load within the window are calculated. Specifically, all pressure data within the window are summed and divided by the number of data points to obtain the average pressure. Then, the square of the difference between each pressure value and the average value is calculated, and the sum is taken to obtain the standard deviation of pressure fluctuation. Finally, the pressure change slope is calculated by dividing the pressure difference at the beginning and end of the window by the time span. The same method is used to calculate the mean, standard deviation, and slope for parameters such as load. All the calculated statistical features are combined in sequence to form a real-time feature vector containing information such as average pressure, pressure fluctuation, pressure trend, and average load, providing dynamic input data for subsequent operating condition identification and control strategy adjustment.

[0026] S22. Based on the real-time feature vector and the parameter set of the operating condition model, a standardized feature vector is generated through feature standardization processing. In this step, the feature mean parameter and standard deviation parameter saved in the operating condition clustering process are first obtained, i.e., the feature standardization parameter. Each original value in the real-time feature vector is subtracted from the historical mean to obtain the median value of the deviation from the mean. Then, the median value is divided by the historical standard deviation to complete the dimensional normalization. For example, for the real-time main steam pressure feature value, the historical pressure mean is subtracted to obtain the deviation. Then, the deviation is divided by the historical pressure standard deviation to obtain the standardized pressure feature value. By performing the above subtraction and division operations on each feature, a standardized feature vector in which all features are in a similar numerical range is finally generated. This vector can more objectively reflect the relative difference between the current operating condition and the historical typical operating condition, providing input data with a unified scale for subsequent operating condition clustering matching and dynamic adjustment of control strategies.

[0027] S23. Perform operating condition matching processing on the standardized feature vector and the operating condition model parameter set to obtain the operating condition label. In this step, the difference between the real-time feature vector and the corresponding feature value of each operating condition cluster center is calculated sequentially. The difference is squared and then summed to obtain the squared distance between the real-time feature and the cluster center. The above sum of squares calculation is repeated for all operating condition clusters, and the operating condition cluster number corresponding to the smallest sum of squared distances is obtained. For example, if the sum of squared differences between the real-time feature and the center of the third type of operating condition is the smallest, then the current unit operating state is determined to belong to the third type of operating condition. Through this quantitative matching method based on Euclidean distance, the category label of the current operating condition is finally output. Alternatively, a distance threshold can be used for clustering, such as setting two clustering thresholds to classify the real-time feature into three categories, providing a classification basis for subsequent calls to the optimal control parameters or prediction models under this operating condition, and ensuring that the control strategy can dynamically adapt to different operating scenarios.

[0028] S24. Based on real-time status data and operating condition model parameter sets, multi-step dynamic prediction processing is performed to obtain the main steam pressure prediction value and load prediction value. In this step, a pre-trained LSTM dynamic prediction model is used to extrapolate future trends. The specific method is as follows: the standardized feature vector and real-time status data, such as valve opening and feedwater flow, are concatenated into the model input according to the time series. The time-series dependent features are extracted layer by layer through the recurrent units of the LSTM network. Each unit first calculates the weighted sum of the current input and the hidden state at the previous time. After nonlinear transformation by the activation function, the cell state is updated, and then the current hidden state is output. After multi-layer stacking and time step expansion, the model finally uses a fully connected layer to map the high-dimensional time series features into the main steam pressure prediction value and load prediction value at multiple future time points. For example, the model may output the main steam pressure value and load value every minute within the next five minutes, forming a continuous prediction sequence. This multi-step prediction method based on deep time series modeling can capture the long-term and short-term dependencies in the dynamic characteristics of the unit, providing a forward-looking basis for AGC command optimization and DCS control parameter adjustment, thereby improving the unit's response accuracy and stability to changes in grid load.

[0029] S25. Based on the predicted main steam pressure, predicted load, AGC command sequence, and pressure setpoint, an estimated energy gap is generated through energy gap assessment. This step first clarifies that the pressure setpoint is the target main steam pressure value set by the DCS system during unit operation based on design parameters or operational optimization objectives, reflecting the steam energy level under ideal operating conditions. Then, the difference between the real-time AGC command and the predicted load is used as the load tracking deviation term, representing the gap between the actual unit output and the grid dispatch demand. Simultaneously, the difference between the pressure setpoint and the predicted main steam pressure is used as the steam energy deviation term, reflecting... The deviation between the unit's steam generation capacity and the target value is calculated by multiplying the load tracking deviation by a weighting coefficient and then adding the result of multiplying the steam energy deviation by another weighting coefficient to obtain the comprehensive energy gap. For example, if the AGC command requires an increase in load but the predicted output is insufficient, and the predicted main steam pressure is lower than the set value, both deviations are positive. After being superimposed, the energy gap increases significantly. The final output estimated energy gap value can intuitively reflect the unit's comprehensive capacity gap in meeting the grid load demand and maintaining its own steam parameter stability, providing a quantitative basis for subsequent adjustments to control parameters such as fuel quantity and valve opening.

[0030] S26. Perform dynamic feedforward processing on the estimated energy gap and operating condition labels to obtain the dynamic feedforward compensation amount. In this step, the corresponding feedforward gain coefficient is first queried from the predefined gain table based on the operating condition label. This table is formulated by analyzing the impact of the energy gap on the unit response under different operating conditions through historical data analysis. For example, the gain coefficient is larger under high load conditions to quickly respond to changes in demand. Then, the queried gain coefficient is multiplied by the estimated energy gap to obtain the preliminary feedforward compensation value. Finally, the compensation value is limited to the maximum and minimum adjustment range allowed by the DCS system to prevent the unit operating parameters from exceeding the limit due to excessive compensation. For example, if the energy gap is positive and the operating condition is a rapid ramp type, the queried gain coefficient is 1.5. If the multiplication of the two exceeds the maximum limit value, the maximum limit value is taken as the final output. The final dynamic feedforward compensation amount will be superimposed on the original control command, enabling the unit to adjust parameters such as fuel quantity and feedwater quantity in advance, track AGC commands more accurately and maintain the stability of main steam pressure, thereby improving the unit's dynamic response capability to changes in grid load.

[0031] This invention can quickly and accurately subdivide the unit's operating conditions into specific categories based on real-time unit status data, greatly improving the accuracy of operating condition identification. Through multi-step dynamic prediction, it can deeply mine the temporal characteristics and potential patterns in the unit's status data, effectively improving the predictive advance and accuracy. This allows the unit to perceive parameter change trends in advance, avoiding control problems caused by sudden parameter changes. The energy gap assessment integrates multiple data sources to comprehensively and accurately assess the unit's current energy gap status, improving the comprehensiveness and accuracy of energy assessment. The feedforward calculation process combines operating conditions, accurately selects appropriate feedforward gain, and performs amplitude limiting processing on the calculation results to ensure that the dynamic feedforward compensation can effectively make up for the energy gap while remaining within a reasonable range, improving the accuracy and stability of feedforward compensation.

[0032] S3. Obtain the target load and emission constraints of the thermal power unit, and generate an optimization processing space. Select candidate control variable groups in the optimization processing space, and perform weighted adaptive processing to obtain fitness values. Based on the fitness values, perform global optimization decoding to obtain the optimal instruction set. Existing technologies mostly adopt fixed weights or single-objective optimization, which have problems such as slow convergence speed, easy to get trapped in local optima, lag in dynamic response, and imbalance between environmental and economic goals. For example, in the scenario of frequent grid peak shaving, traditional PID control cannot be adjusted quickly due to fixed parameters, which often leads to excessive load tracking time, thereby affecting the overall power generation management and adjustment, or emissions exceeding standards leading to air quality decline, such as excessive levels of nitrogen oxides and sulfur dioxide. To solve the above problems, the specific steps are as follows: S31. Based on the target load and emission constraints, perform upper and lower limit processing to generate an optimal processing space. In this step, the target load can be set according to the power generation target of the thermal power unit or the historical power generation target. The emission constraints can be set based on the emission product standards of the thermal power unit. In practice, refer to the national or local environmental protection standards. In terms of processing methods, firstly, based on the unit's own equipment characteristics, historical operating data and current actual operating conditions, determine the feasible upper and lower limits of the three key parameters: fuel quantity, air volume and oxygen quantity. The lower limit of fuel quantity should ensure that the unit can maintain basic combustion stability, while the upper limit should take into account factors such as boiler combustion efficiency and equipment safety capacity. The lower limit of air volume should meet the minimum air volume required for complete fuel combustion, while the upper limit should prevent problems such as furnace temperature drop and combustion instability caused by excessive air volume. The lower limit of oxygen quantity should ensure sufficient oxygen support during combustion, while the upper limit should avoid increasing flue gas heat loss and increasing nitrogen oxide generation due to excessive oxygen quantity. Meanwhile, combining the current operating status of the unit with the emission prediction model, the emission prediction model process is as follows: First, collect the fuel quantity, air volume, oxygen quantity and corresponding nitrogen oxide emissions from the unit's historical operating data to form a data sample set. Then, clean and standardize the data to eliminate the influence of dimensional differences and outliers. Next, use linear regression or simple decision tree algorithms to establish a linear or nonlinear mapping relationship between fuel quantity, air volume, oxygen quantity and nitrogen oxide emissions. In practice, linear regression algorithms are mostly used. Finally, when defining the optimization space, input the real-time or preset combination of fuel quantity, air volume and oxygen quantity into the model to quickly predict the corresponding nitrogen oxide emissions and compare them with emission constraints. Only the combination that meets the constraints is retained, thus forming an optimization processing space that meets both the target load requirements and environmental protection restrictions.

[0033] S32. In the optimization processing space, arbitrarily select multiple sets of candidate control variables, and perform forward processing on the candidate control variable sets and the parameter set of the operating condition model to obtain the predicted values ​​of key performance indicators. In this step, the candidate control variable sets are derived from multiple combinations of fuel quantity, air volume, and oxygen quantity generated by particle swarm optimization in the optimization processing space. This candidate control variable set generated by particle swarm optimization is an initial population that provides input for subsequent steps. At the same time, it combines the real-time operating condition characteristics reflected by the standardized feature vector and the benchmark information provided by the initial operating condition processing model. The initial operating condition processing model simulates the complex nonlinear relationship between fuel quantity, air volume, oxygen quantity and key performance indicators through multi-layer neuron weighted summation and activation function operation. Finally, it outputs the key performance prediction results, which include the predicted values ​​of assessment indicators, nitrogen oxide emissions, and operating efficiency. These results provide a quantitative basis for subsequent multi-objective fitness calculation and global optimal solution selection, supporting the unit to achieve economical and efficient operation while meeting load demand and environmental constraints.

[0034] S33. Based on the predicted values ​​of key performance indicators, a fitness value is obtained through weighted fitness processing. In this step, the predicted values ​​of assessment indicators and operating efficiency are positive contributors. They are weighted and summed according to preset weight coefficients, and then the penalty term for the portion of the predicted nitrogen oxide emission value exceeding the emission limit is subtracted. The penalty term is obtained by taking the positive difference between the predicted emission value and the limit and multiplying it by the penalty coefficient. The final fitness value comprehensively reflects the candidate solution's overall performance in meeting load assessment, improving operating efficiency, and controlling emission exceedances. The penalty coefficient is obtained through an adaptive penalty coefficient adjustment method. This method dynamically monitors the deviation between the predicted nitrogen oxide emission value and the limit during the optimization process. Combined with the iterative feedback of the particle swarm optimization algorithm, the iterative feedback automatically increases or decreases the penalty intensity based on the emission exceedance degree, economic performance, and convergence status of the current iteration step. The penalty coefficient is adjusted in real time, providing a quantitative evaluation basis for selecting the globally optimal control quantity combination during the particle swarm optimization process, ensuring that the unit achieves economic, environmentally friendly, and stable coordinated optimization operation under dynamic peak-shaving conditions.

[0035] S34. Population update processing is performed based on the optimization processing space to generate updated population positions. In this step, multiple initial combinations of fuel quantity, air volume, and oxygen quantity are randomly generated in the optimization decision space as the initial positions of particles. Then, according to the particle swarm optimization rules, the search inertia is maintained by combining inertial weights, and the gravitational pull of the individual historical best position and the global best position is superimposed. Uncertainty in exploration is introduced through random factors. Finally, through iterative update calculations of velocity and position, a new batch of candidate control quantity combinations that meet the constraints are generated. These updated particle swarm positions provide a dynamically optimized set of candidate solutions for subsequent fitness evaluation and global optimal solution selection, supporting the unit to achieve efficient collaborative optimization under multi-objective constraints.

[0036] S35. Based on the updated population position, the current global optimal solution is obtained through elite retention processing. In this step, for the updated population position, the performance prediction and fitness calculation process is executed sequentially for each candidate control variable combination. First, the real-time predicted values ​​of assessment indicators, emissions, and operating efficiency are obtained through forward calculation using the initial operating condition processing model. Then, the comprehensive performance of its economy, environmental protection, and load tracking capability is comprehensively evaluated through a multi-objective weighted fitness function. Subsequently, the fitness value of the current particle is compared with the individual's historical best performance and the global historical best performance. If it is better than the individual's historical best performance, the individual's optimal position is updated. If it is better than the global historical best performance, the global optimal position is updated. Finally, through the dynamic retention and transmission of the global optimal position, it is ensured that the particle swarm optimization algorithm always focuses on the candidate solution with the best comprehensive performance. This supports the unit in achieving the synergistic goals of accurate load tracking, strict emission constraints, and operating cost optimization under dynamic peak shaving conditions, providing key decision support for the efficient and stable operation of thermal power units under grid peak shaving and environmental protection requirements.

[0037] S36. Obtain all current global optimal solutions and perform optimal decoding processing to obtain the optimal instruction set. In this step, when the particle swarm optimization iteration meets the preset termination conditions, such as the maximum number of iterations or the fitness convergence threshold, the global optimal solution is decoded and parsed into specific numerical optimal fuel instructions, optimal air volume instructions, and optimal oxygen volume instructions. These instructions directly correspond to the precise control parameters of fuel quantity, air volume, and oxygen quantity during unit operation. Ultimately, this supports the unit in achieving the coordinated goals of accurate load tracking, strict emission constraints, and optimized operating costs under dynamic peak shaving conditions, ensuring that the thermal power unit achieves an economical, efficient, stable, and reliable comprehensive operating effect while meeting the grid load demand and environmental regulations.

[0038] This invention achieves real-time and accurate prediction of key performance through an initial operating condition processing model. By combining constrained particle swarm optimization and adaptive penalty coefficients, it quickly converges to the global optimum under dynamic operating conditions, significantly shortening the processing time. It can also process data in real time, balancing load tracking accuracy and emission constraints. This solves the problems of low convergence efficiency, poor dynamic adaptability, and subjective weight allocation in traditional methods, and greatly improves load tracking accuracy and thermal power data processing efficiency.

[0039] S4. Obtain the real-time dataset of the thermal power unit, perform energy storage fuzzy processing on the real-time status data and the real-time dataset to generate coefficient sets and component sets, and perform instruction synthesis processing based on the coefficient sets, component sets and optimal instruction sets to obtain the main control instruction set; Existing technologies do not perform frequency domain decomposition, which easily leads to high and low frequency disturbance coupling interference. They often force adjustment when energy storage is insufficient, causing pressure fluctuations. Dynamic decoupling depends on fixed coefficients, which is difficult to adapt to changes in operating conditions and is prone to conflicts between turbine and boiler regulation. Boiler instructions lack inertial smoothness and are prone to combustion instability due to sudden changes. Turbine instructions lack amplitude limit protection and have the risk of exceeding limits. For example, when the grid load fluctuates at high frequencies, existing technologies do not separate high and low frequencies, and the turbine is prone to response lag in practice, affecting thermal power efficiency. In addition, excessive boiler regulation can cause short-term oscillations in the unit, which seriously affects the normal operation and service life of the thermal power unit; To solve the above problems, the specific steps are as follows: S41. Based on the real-time dataset, a component set is obtained through digital filtering. In this step, the real-time dataset includes a real-time AGC command sequence and a real-time primary frequency modulation command. The component set includes low-frequency disturbance components and high-frequency disturbance components. First, the real-time AGC command sequence and the real-time primary frequency modulation command are added to obtain the total power disturbance signal. Then, a low-pass filter is used to filter the frequency of the total power disturbance signal. By setting a specific cutoff frequency, the low-pass filter allows signal components below that frequency to pass through while suppressing high-frequency components, thereby extracting the low-frequency fluctuation component in the total power disturbance. This process is achieved through the convolution operation of the digital filter. Specifically, the total power disturbance signal is linearly convolved with the impulse response of the low-pass filter to obtain the low-frequency disturbance component, while the high-pass filter... The filter performs reverse frequency filtering on the total power disturbance signal, allowing signal components above the cutoff frequency to pass through, suppressing low-frequency components, and extracting high-frequency fluctuation components. The high-pass filter's processing is also based on convolution operations, achieved through linear convolution of the total power disturbance signal with the high-pass filter's impulse response. Low-pass and high-pass filters are commonly used filtering methods, which will not be elaborated here. Finally, through the synergistic effect of the low-pass and high-pass filters, the total power disturbance signal is decomposed into low-frequency disturbance components and high-frequency disturbance components. These components can clearly characterize the frequency characteristics of unit load fluctuations. The low-frequency components reflect the long-term load change trend of the power grid, while the high-frequency components capture transient fluctuations during the unit's rapid response process, thereby improving the unit's load tracking accuracy and operational stability under complex power grid commands.

[0040] S42. Based on real-time status data, assess the energy storage status and generate real-time energy storage margin. This step first obtains the real-time main steam pressure value and the main steam pressure setpoint, calculating the absolute difference between them. This difference reflects the degree to which the current pressure deviates from the target value. Then, a pressure deviation dead zone threshold is introduced. This threshold defines the allowable pressure fluctuation range. The absolute difference is divided by the dead zone threshold to obtain the relative proportion of the pressure deviation. Finally, the real-time energy storage margin is obtained by subtracting this relative proportion from 1. When the pressure deviation is within the dead zone, the relative proportion is zero, and the energy storage margin is 1, indicating that the unit has sufficient energy storage and can stably respond to load changes. If the deviation exceeds the dead zone, the margin will be less than 1, indicating that the energy storage status is becoming tight and the control strategy needs to be adjusted to maintain operational stability. This margin indicator is directly related to the unit's dynamic peak-shaving capability. In scenarios with frequent AGC adjustments, by monitoring the energy storage margin in real time, the unit's ability to withstand load fluctuations can be accurately judged, ultimately improving the unit's comprehensive adaptability and operational reliability under complex grid commands. (Thermal power) S43. Perform dynamic fuzzy processing on the real-time energy storage margin and component sets to obtain a coefficient set. In this step, the coefficient set includes the turbine load coefficient and the boiler load coefficient. First, the real-time energy storage margin, low-frequency disturbance component, and high-frequency disturbance component are used as input variables and mapped to fuzzy sets (e.g., high, medium, low) through membership functions. Then, inference is performed based on a preset fuzzy rule base. The membership function typically uses a triangular function, trapezoidal function, Gaussian function, or bell function. For example, when the energy storage margin is sufficient and low-frequency disturbance is dominant, the rule triggers the boiler-dominated adjustment mode. The output is aggregated through a weighted average method to increase the boiler load coefficient to slowly absorb low-frequency fluctuations. When high-frequency disturbances are significant and energy storage margins are tight, the rules trigger the turbine's rapid response mode, increasing the turbine load factor to quickly track high-frequency changes. During fuzzy inference, the activation intensity of each rule is determined by the membership degree of the input variables. The final output is obtained by aggregating the results of all activated rules, resulting in the turbine load factor and boiler load factor. These factors respectively characterize the response weights of the turbine and boiler to load changes. Dynamic decoupling of the two can balance the speed of load tracking and operational stability, avoid coupling conflicts between turbine and boiler regulation, and achieve precise adaptation of turbine and boiler collaborative control. This supports the unit to achieve efficient, stable, and environmentally friendly dynamic response under complex grid commands.

[0041] S44. Dynamically generate commands based on coefficient and component sets to obtain turbine command increments and boiler command increments. In this step, the turbine command increment is first obtained by multiplying the turbine load allocation coefficient by the high-frequency disturbance component and the low-frequency disturbance component, and then adding them together. This process ensures that the turbine can respond quickly to high-frequency fluctuations while taking into account low-frequency changes, thus improving the real-time performance of load tracking. The boiler command increment is obtained by multiplying the boiler load allocation coefficient by the low-frequency disturbance component and smoothing it through a first-order inertial element. The inertial time constant controls the smoothing degree to avoid boiler-side pressure fluctuations or combustion instability caused by sudden command changes. In this allocation method, the turbine focuses on high-frequency transient regulation, while the boiler focuses on low-frequency trend adjustment. The two work together to achieve fast and slow decoupling. In the scenario of frequent AGC adjustment, this mechanism can ensure that the turbine can quickly track the high-frequency changes of the grid command, while the boiler's inertia can smooth and buffer low-frequency fluctuations, avoid regulation conflicts, and improve the stability and economy of unit operation. The final output turbine command increment and boiler command increment directly drive the adjustment of the turbine valve and boiler combustion system, supporting the unit to achieve the coordinated goals of accurate load tracking, stable pressure control and optimized operating costs in dynamic peak shaving.

[0042] S45. Based on the boiler command increment, dynamic feedforward compensation, and optimal command set, boiler commands are synthesized to obtain the boiler master control command. In this step, the dynamic feedforward compensation is first used to compensate for the boiler-side response delay, ensuring that the command responds to load changes in advance. The boiler command increment reflects the adjustment demand corresponding to the current low-frequency disturbance, and the boiler load is stably adjusted after inertial smoothing. The optimal command set provides a benchmark adjustment direction based on the global optimization result. The difference between the optimal command set and the base coal quantity forms the optimized benchmark adjustment value, which is used to correct the base coal quantity to match the optimal operating state. Finally, the boiler master control command is composed of the base coal quantity and the boiler command increment. The result is obtained by adding the four parts: dynamic feedforward compensation, incremental regulation, and optimized benchmark adjustment. This synthesis process achieves the coordinated integration of feedforward compensation, incremental regulation, and optimized benchmark. It not only ensures stable tracking of low-frequency disturbances on the boiler side, but also integrates the global optimal control objective through optimized benchmark adjustment, avoiding conflicts between local regulation and global objectives. In the scenario of frequent AGC regulation, this mechanism can balance the stability and economy of boiler regulation, support the unit to achieve the coordinated objectives of accurate load tracking, combustion efficiency optimization, and emission constraint satisfaction in dynamic peak shaving, and improve the unit's comprehensive adaptability and operational reliability under complex power grid commands.

[0043] S46. Perform instruction synthesis processing on the optimal instruction set and the turbine instruction increment to generate the turbine main control instruction parameter set. Combine the boiler main control instruction and the turbine main control instruction parameter set into the main control instruction set. In this step, the main control instruction set includes the boiler main control instruction and the turbine main control instruction parameter set. The turbine main control instruction parameter set includes the turbine main control instruction, the bottom air volume setpoint, and the bottom oxygen setpoint. First, the turbine instruction increment needs to be processed by valve position limiting, including rate limiting and position limiting: the rate limiting controls the speed of instruction change to avoid unit oscillation caused by excessive adjustment, and the position limiting constrains the instruction range to ensure that it does not exceed the equipment safety boundary. The processed increment is added to the current power to form the turbine main control command. This command responds to high-frequency disturbances while ensuring a smooth transition. The bottom-level airflow setpoint directly adopts the optimal airflow command from the optimal command set to ensure that the airflow accurately matches the combustion demand. The bottom-level oxygen setpoint directly uses the optimal oxygen command to optimize combustion efficiency and control nitrogen oxide emissions. Through this processing, the turbine main control command can quickly respond to load changes while ensuring regulation safety through a limiting mechanism. The airflow and oxygen setpoints are directly integrated into the global optimization results to achieve synergy between combustion control and environmental protection goals. The final output command directly drives the turbine valves, fans, and oxygen regulation devices, supporting the unit to achieve the composite optimization goals of accurate load tracking, improved operational stability, and compliance with emission constraints during dynamic peak shaving.

[0044] This invention accurately distinguishes between high and low frequency disturbances through frequency domain decomposition. The steam turbine focuses on high-frequency transient response, while the boiler emphasizes low-frequency trend adjustment, achieving fast and slow decoupling. Based on the energy storage margin calculation of the main steam pressure deviation, the margin is kept at one within the dead zone to avoid ineffective regulation. The fuzzy rule dynamic matching coefficient can adapt to rapid changes in operating conditions and reduce regulation conflicts. By smoothing the boiler commands through first-order inertia to buffer low-frequency fluctuations and by limiting the amplitude of the steam turbine commands, the safety of thermal power unit regulation is ensured, continuous vibration of unit equipment is avoided, and the service life is effectively extended.

[0045] S5. Obtain the current unit operation data based on the main control instruction set, evaluate and update the model based on the current unit operation data, obtain the model update flag, and retrain the model based on the model update flag; the existing technology lacks a dynamic operating condition matching and update mechanism, the feature extraction is fixed and the operating condition judgment is missing, which makes it easy to fail to adjust when the operating condition changes, thus leading to control lag; To solve the above problems, the specific implementation steps are as follows: S51. Based on the current unit operating data, feature reconstruction processing is performed to generate the current feature vector and the current operating condition label. In this step, the DCS first collects the new round of process variable data of the unit after the control command is applied, such as main steam pressure, temperature, flow rate, etc., and calculates the mean, variance, extreme values ​​and other statistics of each variable to form the original vector of actual features. Then, the original features are standardized using the output feature standardization parameters, namely the mean and standard deviation: the actual feature vector is equal to the original value of the actual feature minus the mean and divided by the standard deviation to eliminate the influence of dimensions. Finally, the standardized actual feature vector is matched with the operating condition cluster center in the operating condition model parameter set for similarity matching. By calculating the Euclidean distance or cosine similarity, the closest operating condition cluster is determined, thereby obtaining the actual operating condition label. This process realizes real-time feature extraction and accurate identification of the unit's operating status, supporting the subsequent control strategy to dynamically adjust parameters according to the current operating condition, such as adjusting the weight of fuzzy rules or optimizing the command synthesis logic, improving the load tracking accuracy and operating stability of the unit in dynamic peak shaving, and avoiding regulation failure or equipment damage caused by misjudgment of operating conditions.

[0046] S52. Based on the current feature vector and the current operating condition label, perform feature consistency judgment and generate a model adaptability signal. In this step, compare the current operating condition label with the operating condition label of the previous control cycle. If the two are inconsistent, it is determined that the operating condition has changed significantly, indicating a shift in the control environment. In the feature consistency judgment step, calculate the average distance between the current feature vector and the center of the matching operating condition cluster. Use the distance threshold to determine whether the actual operating features deviate from the original clustering pattern. If the distance continuously exceeds the threshold, it indicates that the feature distribution has shifted and the model's characterization ability has decreased. The model adaptability signal is triggered by either a significant shift in operating condition or a continuously excessive feature distance. When either condition is met, the signal is marked as unadaptable, indicating that the control strategy needs to be adjusted or the model parameters need to be updated. If neither condition is met, it is marked as adaptive, indicating that the current model is still effective. This mechanism provides a key basis for model optimization, supporting the unit to dynamically adjust the weight of fuzzy rules, optimize the instruction synthesis logic, or update the center of the operating condition cluster when the operating condition changes. This improves the adaptability of the control strategy to complex operating conditions, avoids regulation failure or equipment loss due to model failure, and optimizes the overall operating economy and stability.

[0047] S53. Perform model update judgment on the model adaptability signal and generate a model update flag. In this step, the model update flag directly inherits the adaptive signal state. If the adaptive signal is not adaptive, the update flag is synchronously marked as needing to be updated. At this time, the system archives the actual operation process data, standardized feature vectors and actual operating condition labels of this round to the historical database to form an enhanced dataset. Then, the complete process of S1 is called to retrain the operating condition model parameter set, operating condition clustering model and optimal instruction set using the newly added data to ensure that the model is always iterated based on the latest operating data. This mechanism supports the model to dynamically adapt to changes in operating conditions. For example, when the grid load mode changes abruptly or the unit characteristics drift, the control strategy can improve its accurate response capability to high-frequency fluctuations and low-frequency trends through data archiving and model retraining, avoid adjustment lag or overshoot caused by outdated models, reduce equipment losses and improve overall operational reliability.

[0048] This invention uses dynamic feature extraction and operating condition matching to adapt to changes in operating status in real time. The stability of operating conditions and consistency of features are used to accurately identify model failures. The model update mechanism is based on the latest data to retrain the model, ensuring that the control strategy always responds accurately to changes in operating conditions. Data archiving and model iteration improve load tracking accuracy, operational stability and economy, reduce equipment wear and tear, and optimize the overall operational reliability of the model.

[0049] Example 2: Because existing technologies cannot accurately match actual operating conditions and struggle to capture temporal dynamic characteristics, and because energy assessment and feedforward compensation lack comprehensive consideration, they are prone to control deviations and instability. Please refer to [link to relevant documentation]. Figure 2The diagram shown is a structural block diagram of the thermal power unit control strategy optimization system based on the combination of DCS and AGC provided in this embodiment. The system includes: operating condition processing module, feedforward compensation module, instruction optimization module, instruction generation module, and model update module. The operating condition processing module is used to collect raw data from thermal power units, construct operating condition models based on raw data, and generate operating condition model parameter sets. The raw data includes raw DCS historical operating data and AGC instruction sequences. The feedforward compensation module is used to determine the pressure setpoint based on historical data thresholds, collect real-time status data of thermal power units, process the operating data to obtain operating condition labels, and generate dynamic feedforward compensation amount based on the pressure setpoint, operating condition labels and operating condition model parameter set. The instruction optimization module is used to obtain the target load and emission constraints of the thermal power unit, generate an optimization processing space, select candidate control variable groups in the optimization processing space, perform weighted adaptive processing to obtain fitness values, and perform global optimization decoding based on fitness values ​​to obtain the optimal instruction set. The instruction generation module is used to acquire the real-time dataset of the thermal power unit, perform energy storage fuzzy processing on the real-time status data and the real-time dataset to generate a coefficient set and a component set, and perform instruction synthesis processing based on the coefficient set, component set and optimal instruction set to obtain the main control instruction set. The model update module is used to obtain the current unit operation data based on the main control instruction set, evaluate and update the model based on the current unit operation data, obtain the model update flag, and retrain the model based on the model update flag.

[0050] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code, including but not limited to disk storage, CD-ROM, optical storage, etc.

[0051] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for optimizing the control strategy of thermal power units based on the combination of DCS and AGC, characterized in that, The steps of this method are as follows: collect raw data of thermal power units, construct an operating condition model based on the raw data, and generate a parameter set for the operating condition model. The raw data includes raw DCS historical operating data and AGC instruction sequences. The pressure setpoint is determined based on historical data thresholds, real-time status data of thermal power units is collected, and operating condition data is processed to obtain operating condition labels. Dynamic feedforward compensation is generated based on the pressure setpoint, operating condition labels, and operating condition model parameter set. The target load and emission constraints of the thermal power unit are obtained, and an optimization processing space is generated. Candidate control variable groups are selected in the optimization processing space, and weighted adaptive processing is performed to obtain fitness values. Based on the fitness values, global optimization decoding is performed to obtain the optimal instruction set. The real-time dataset of the thermal power unit is obtained. The real-time status data and the real-time dataset are subjected to energy storage fuzzy processing to generate a coefficient set and a component set. Based on the coefficient set, component set and optimal instruction set, instruction synthesis processing is performed to obtain the main control instruction set. The current unit operation data is obtained based on the master control instruction set. The model is evaluated and updated based on the current unit operation data to obtain the model update flag. The model is then retrained based on the model update flag.

2. The method for optimizing the control strategy of thermal power units based on the combination of DCS and AGC according to claim 1, characterized in that, The operating condition model is constructed based on the raw data, including: Based on the original DCS historical operation data, an aligned variable dataset is obtained through filtering and timestamp alignment. Key performance data processing was performed on the AGC instruction sequence and alignment variable dataset to obtain historical KPI label data and historical KPI labels. Based on historical KPI tag data, the working condition model parameter set is obtained through working condition clustering.

3. The method for optimizing the control strategy of thermal power units based on the combination of DCS and AGC according to claim 1, characterized in that, The operating condition data is then processed to obtain operating condition labels, including: Real-time feature vectors are obtained by integrating features based on real-time status data. Based on the real-time feature vector and the parameter set of the working condition model, a standardized feature vector is generated through feature standardization processing. The standardized feature vectors and the parameter set of the working condition model are subjected to working condition matching processing to obtain working condition labels.

4. The method for optimizing the control strategy of thermal power units based on the combination of DCS and AGC according to claim 1, characterized in that, Dynamic feedforward compensation is generated based on the pressure setpoint, operating condition label, and operating condition model parameter set, including: Multi-step dynamic prediction processing is performed based on real-time status data and operating condition model parameter set to obtain the main steam pressure prediction value and load prediction value. Based on the predicted main steam pressure, predicted load, AGC command sequence, and pressure setpoint, an estimated energy gap is generated through energy gap assessment. Dynamic feedforward processing is performed on the estimated energy gap and operating condition labels to obtain the dynamic feedforward compensation amount.

5. The method for optimizing the control strategy of thermal power units based on the combination of DCS and AGC according to claim 1, characterized in that, The fitness value is obtained by weighted adaptation, including: Based on the target load and emission constraints, upper and lower limits are processed to generate an optimization space; Multiple candidate control variable groups are arbitrarily selected in the optimization processing space, and forward processing is performed on the candidate control variable groups and the parameter set of the operating condition model to obtain the predicted values ​​of key performance. The fitness value is obtained by weighting the predicted key performance values.

6. The method for optimizing the control strategy of thermal power units based on the combination of DCS and AGC according to claim 1, characterized in that, Global optimization decoding based on fitness values ​​includes: Population update processing is performed based on the optimization processing space to generate updated population positions; Based on the updated population position, the current global optimum is obtained through elite retention processing; Obtain all current global optimal solutions and perform optimal decoding processing to obtain the optimal instruction set.

7. The method for optimizing the control strategy of thermal power units based on the combination of DCS and AGC according to claim 1, characterized in that, Energy storage fuzzy processing is performed on real-time status data and real-time datasets to generate a coefficient set, including: The component set is obtained by digital filtering based on the real-time dataset; Energy storage status is assessed based on real-time status data to generate real-time energy storage margin. Dynamic fuzzy processing is performed on the real-time energy storage margin and component sets to obtain the coefficient set.

8. The method for optimizing the control strategy of thermal power units based on the combination of DCS and AGC according to claim 1, characterized in that, Instruction synthesis based on coefficient set and optimal instruction set includes: Dynamic command generation is performed based on coefficient set and component set to obtain turbine command increment and boiler command increment; Boiler commands are synthesized based on boiler command increments, dynamic feedforward compensation, and the optimal command set to obtain boiler master control commands. The optimal instruction set and the turbine instruction increment are processed to generate a turbine main control instruction parameter set. The boiler main control instruction and the turbine main control instruction parameter set are then combined into the main control instruction set.

9. The method for optimizing the control strategy of thermal power units based on the combination of DCS and AGC according to claim 1, characterized in that, Evaluation and model updates are performed based on current unit operating data, including: Based on the current unit operating data, feature reconstruction processing is performed to generate the current feature vector and the current operating condition label; Based on the current feature vector and the current working condition label, feature consistency is judged, and model adaptation signal is generated; The model adaptation signal is used to determine the model update and generate a model update flag.

10. A system applied to the thermal power unit control strategy optimization method based on the combination of DCS and AGC as described in any one of claims 1-9, characterized in that, The system includes: The operating condition processing module is used to collect raw data from thermal power units, construct operating condition models based on raw data, and generate operating condition model parameter sets. The raw data includes raw DCS historical operating data and AGC instruction sequences. The feedforward compensation module is used to determine the pressure setpoint based on historical data thresholds, collect real-time status data of thermal power units, process the operating data to obtain operating condition labels, and generate dynamic feedforward compensation amount based on the pressure setpoint, operating condition labels and operating condition model parameter set. The instruction optimization module is used to obtain the target load and emission constraints of the thermal power unit, generate an optimization processing space, select candidate control variable groups in the optimization processing space, perform weighted adaptive processing to obtain fitness values, and perform global optimization decoding based on fitness values ​​to obtain the optimal instruction set. The instruction generation module is used to acquire the real-time dataset of the thermal power unit, perform energy storage fuzzy processing on the real-time status data and the real-time dataset to generate a coefficient set and a component set, and perform instruction synthesis processing based on the coefficient set, component set and optimal instruction set to obtain the main control instruction set. The model update module is used to obtain the current unit operation data based on the main control instruction set, evaluate and update the model based on the current unit operation data, obtain the model update flag, and retrain the model based on the model update flag.

Images