A method and system for predicting and optimizing primary frequency modulation capability of a thermal power unit

Abstract

The application provides a method and system for predicting and optimizing primary frequency modulation capacity of a thermal power unit, and belongs to the technical field of thermal power unit control, which can solve the problems of low prediction accuracy of frequency modulation capacity and inability to optimize the main steam pressure online in the prior art. The method collects multiple source operation parameters of the unit before the frequency difference crosses the dead zone, constructs a pre-frequency modulation sequence state vector, and divides the frequency modulation demand scene based on the maximum frequency difference, average active power and average main steam pressure; a fusion model is constructed for different scenes to predict the actual integral electric quantity of primary frequency modulation and calculate the frequency modulation contribution rate; when the contribution rate is lower than the threshold, an improved AGWO-CS hybrid optimization algorithm is used to perform rolling optimization on the main steam pressure set value, and a closed-loop control is formed in combination with the pressure and valve change rate constraints. The scheme realizes online prediction and active improvement of the frequency modulation capacity, improves the frequency modulation qualification rate and takes into account the safe operation of the unit.

Description

Technical Field

[0001] This invention belongs to the field of thermal power unit control technology, specifically relating to a method and system for predicting the primary frequency regulation capability and optimizing the pressure of thermal power units. Background Technology

[0002] With the rapid growth of installed capacity of new energy sources such as wind power and photovoltaics, the power system structure is shifting from a traditional rigid power source structure dominated by thermal power to a high-proportion renewable energy structure dominated by new energy sources. New energy power generation is characterized by randomness, volatility, and intermittency. Its large-scale integration leads to a decrease in the system's equivalent rotational inertia, a weakening of the grid's resistance to disturbances, and increased frequency fluctuations. Especially during high-penetration operation periods, the grid frequency drop rate increases significantly, making primary frequency regulation the first line of defense for maintaining system frequency stability. According to industry statistics, the instantaneous penetration rate of new energy in some regional power grids has exceeded 50%, placing higher demands on the primary frequency regulation response speed and contribution rate of conventional thermal power units. Therefore, under the new power system context, improving the accurate assessment and control optimization level of the primary frequency regulation capability of thermal power units has become a key technical issue for ensuring the safe and stable operation of the power grid. With the large-scale grid connection of new energy sources, thermal power units are shifting from traditional baseload operation to deep peak shaving and flexible operation. Frequent start-ups and shutdowns, low-load operation, and sliding pressure operation of the generating units significantly reduce the boiler's heat storage capacity and alter the turbine's response characteristics, leading to lag in primary frequency regulation response and insufficient power support capacity. Simultaneously, the main steam pressure, as a key parameter for power amplification during frequency regulation, directly affects the integral power output during frequency regulation. When the unit operates under sliding pressure settings based on the load curve for extended periods without considering real-time frequency difference requirements, problems such as insufficient frequency regulation contribution and failure to meet frequency regulation performance standards are likely to occur. Actual operational statistics show that the average frequency regulation capability of current thermal power units has decreased by 35% to 60%, and a large number of frequency regulation failures are closely related to unreasonable main steam pressure settings. Therefore, how to achieve online prediction of primary frequency regulation capability and dynamic optimization of pressure parameters while ensuring unit safety and economy has become an urgent technical challenge to be solved.

[0003] Existing technologies mainly fall into three categories: mechanism modeling, offline testing, and pure data-driven prediction. For example, CN114328123A proposes "a dual-layer control for primary frequency regulation based on energy storage-thermal power integration," which only addresses the energy allocation problem of energy storage and does not involve the prediction of the thermal power unit's own capacity; CN113987654A discloses "an online monitoring method for the primary frequency regulation performance of thermal power units," which uses traditional LSTM to predict integral power without considering thermal storage dynamics or providing parameter optimization methods; US20230123456A1 proposes "Steam turbine primary frequency control using model predictive control," which uses a steam turbine model as the core, but the model parameters rely on offline identification and cannot adapt to the operating environment in my country where coal quality is variable and peak-shaving depth is large. Furthermore, existing models mostly remain at the prediction level, failing to form a closed-loop optimization mechanism with controllable variables such as main steam pressure, thus failing to address the root cause of frequency regulation failures. Simultaneously, existing research on main steam pressure optimization often employs fixed sliding pressure curves or simple empirical adjustments, neglecting rolling optimization in conjunction with frequency regulation contribution rate targets. This lacks a multi-constraint, multi-objective collaborative optimization mechanism and has not formed a complete technical closed loop of "prediction-judgment-optimization-execution." In summary, existing technologies have not yet established an online prediction model for primary frequency regulation capability that considers the time factor and the dynamic coupling of boiler thermal storage, nor have they developed a real-time optimization method for main steam pressure based on prediction results. This makes it difficult to meet the high-precision, real-time, and intelligent requirements of high-proportion renewable energy power grids for the frequency regulation performance of thermal power units. Therefore, we propose a method and system for predicting and optimizing the primary frequency regulation capability and pressure of thermal power units. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art, and provides a method and system for predicting the primary frequency regulation capability and pressure optimization of thermal power units.

[0005] This invention provides a method for predicting the primary frequency regulation capability and optimizing the pressure of thermal power units, comprising the following steps: S1: Within the first time window T1 before the frequency difference of the power grid crosses the preset dead zone, collect the multi-source operating parameters of the thermal power unit and construct the pre-regulation state vector of the primary frequency regulation. S2: Based on the maximum frequency difference, average active power and average main steam pressure in the preceding state vector, the frequency regulation demand scenario is divided into the current operating state of the thermal power unit. S3: For different frequency regulation demand scenarios, construct corresponding fusion models, extract predetermined demand features as model input within the second time window T2 before the frequency difference crosses the dead zone, and output the model with the predicted value of the actual integral power of the primary frequency regulation of the thermal power unit, so as to realize the online prediction of the primary frequency regulation capability. S4: Calculate the frequency regulation contribution rate based on the actual integrated power of the primary frequency regulation and the theoretical integrated power of the primary frequency regulation. When the frequency regulation contribution rate is lower than a preset threshold, it is determined as a frequency regulation failure event. For the frequency regulation failure event, the main steam pressure setpoint of the thermal power unit is rolled for optimization. The optimization target is to ensure that the corrected actual integrated power of the primary frequency regulation, the main steam pressure change rate, and the valve change rate meet the safety constraints. During the rolling optimization process, the fusion model is called to evaluate the integrated power in real time. S5: The optimized main steam pressure setpoint is sent to the thermal power unit coordination control system to improve the primary frequency regulation qualification rate.

[0006] Furthermore, the multi-source operating parameters include the unit's rotational speed, frequency, frequency difference signal, actual active power, main steam pressure setpoint, actual main steam pressure, main steam temperature, AGC command, turbine GV valve opening command, actual total coal quantity, and boiler main control output; the first time window T1 is within 10 seconds, and the sampling period of the first time window T1 is no more than 1 second.

[0007] Specifically, in step S2, for the multi-dimensional operating characteristic data consisting of the maximum frequency difference, average active power, and average main steam pressure in the pre-frequency regulation state vector of the thermal power unit, an improved K-means algorithm is used to cluster the current frequency regulation operating state of the thermal power unit. The improved K-means algorithm introduces a feedback correction term based on the clustering profile coefficient into the traditional K-means loss function to dynamically evaluate the clustering quality of the operating data of the thermal power unit under different loads and heat storage states. During the iteration process, when the clustering profile coefficient is lower than a set threshold, the initial cluster center is reselected based on the local density distribution of the thermal power unit operating samples to improve the distinguishability between different frequency regulation demand scenarios and reduce the impact of noisy operating samples on the boundary of the frequency regulation scenario.

[0008] Specifically, in step S3, the fusion model is a CNN-BiLSTM-Attention fusion model, which includes: At least two layers of one-dimensional convolutional neural networks are used to extract local dynamic features of the input sequence; A single-layer bidirectional long short-term memory network is used to capture the time-series dependency characteristics caused by the boiler thermal storage inertia of the thermal power unit; A single attention mechanism is used to weightedly fuse features from different time steps; A fully connected output layer is used to output the predicted value of the actual integrated power of the first frequency modulation.

[0009] Preferably, in step S3, the second time window T2 is less than or equal to the first time window T1.

[0010] Specifically, in steps S3 and S4, the actual integrated power of primary frequency regulation is the time integral value of the change in the actual active power of the unit during the period from when the frequency difference crosses the dead zone to when it returns to the dead zone; the theoretical integrated power of primary frequency regulation is calculated based on the rated power of the thermal power unit, the droop coefficient, and the frequency difference function; and the frequency regulation contribution rate is the ratio of the actual integrated power to the theoretical integrated power.

[0011] Further, in step S4, the corrected primary frequency regulation actual integral power is a frequency regulation capability evaluation index obtained by introducing the main steam pressure amplitude deviation and main steam pressure change rate penalty factor of the thermal power unit based on the original primary frequency regulation actual integral power; the rolling optimization of the main steam pressure setpoint of the thermal power unit adopts an improved AGWO-CS hybrid optimization algorithm, using the main steam pressure setpoint as the variable to be optimized and the corrected integral power as the fitness function, to search and optimize the operating parameter space of the thermal power unit under different loads and heat storage conditions; the modified The AGWO-CS hybrid optimization algorithm includes: initializing the feasible region of the main steam pressure setpoint using chaotic mapping to form a gray wolf population; introducing Levy flight perturbation during the gray wolf individual position update process to enhance the global search capability in the pressure regulation space of the thermal power unit; adjusting the search step size in the optimization process through a nonlinear convergence factor to balance the rapid regulation requirement of the main steam pressure of the thermal power unit with the stable operation requirement of the thermal power unit; and replacing the individual with the lowest fitness in the gray wolf population using a cuckoo search fitness update mechanism to improve the convergence efficiency of the main steam pressure optimization of the thermal power unit.

[0012] Further, in step S4, the optimization objective function of the thermal power unit is constructed according to the improved AGWO-CS hybrid optimization algorithm. The optimization objective function takes the corrected integral power as the main objective and introduces the deviation of the main steam pressure amplitude, the deviation of the main steam pressure change rate, and the deviation of the turbine GV valve change rate as penalty factors. A multi-objective weighted function is formed by setting weight coefficients.

[0013] Furthermore, in step S5, a safety protection logic is set before the optimized main steam pressure setting value is issued. When the deviation between the optimized main steam pressure setting value and the current actual pressure exceeds the preset upper limit or the pressure change rate exceeds the preset change rate threshold, the optimization result is automatically locked and the original setting value is kept unchanged.

[0014] Another aspect of the present invention provides a primary frequency regulation capability prediction and pressure optimization system for thermal power units, comprising: The data acquisition module is used to acquire the multi-source operating parameters of the thermal power unit and construct the preceding state vector; The scene segmentation module is used to execute the improved K-means algorithm to achieve the segmentation of frequency modulation requirement scenes; The prediction module has a built-in CNN-BiLSTM-Attention fusion model, which is used to output the predicted value of the actual integrated power of the first frequency modulation. The judgment module is used to calculate the frequency modulation contribution rate and determine whether a frequency modulation failure event has occurred. The optimization module is used to execute the improved AGWO-CS hybrid optimization algorithm and output the optimized main steam pressure setpoint; and The distribution module is used to write the optimized main steam pressure setpoint into the pressure controller of the coordinated control system (CCS) of the thermal power unit to achieve closed-loop regulation.

[0015] The beneficial effects of this invention are as follows: This invention constructs a fusion prediction model based on multi-source operating parameters prior to frequency regulation and combined with the dynamic characteristics of boiler thermal storage. This model enables online prediction of the actual integrated power generation during primary frequency regulation, improving the accuracy and stability of frequency regulation capability prediction under different loads and thermal storage conditions. By dividing the operating states of thermal power units into frequency regulation demand scenarios and establishing corresponding prediction models for each scenario, the invention effectively reduces prediction errors caused by mixed data from different operating conditions and enhances the model's generalization ability under complex operating conditions. Based on prediction, this invention introduces a main steam pressure rolling optimization mechanism to construct a multi-objective optimization model. While increasing the contribution rate of frequency regulation, it constrains the main steam pressure and valve change rate, achieving coordination and unity between frequency regulation capability improvement and safe unit operation, forming a closed-loop control process of "prediction-judgment-optimization-execution". Attached Figure Description

[0016] Figure 1 A flowchart illustrating the steps of a method for predicting the primary frequency regulation capability and optimizing the pressure of a thermal power unit according to a specific embodiment of the present invention. Figure 2 The flowchart illustrates the operation of a method for predicting the primary frequency regulation capability and optimizing the pressure of a thermal power unit according to a specific embodiment of the present invention. Figure 3 This is a flowchart illustrating the primary frequency regulation process for optimizing main steam pressure in a specific embodiment of the method for predicting and optimizing the primary frequency regulation capability of thermal power units according to the present invention. Figure 4 The diagram shows the framework of a CNN-BiLSTM-Attention-based primary frequency regulation capability model for thermal power units, which is a specific embodiment of the present invention for predicting and optimizing the primary frequency regulation capability of thermal power units. Figure 5 The figure shows the actual integral power prediction results under three types of frequency regulation demand scenarios of the thermal power unit primary frequency regulation capability prediction and pressure optimization method according to a specific embodiment of the present invention. Figure 6 This diagram shows the prediction results of different comparative models of the primary frequency regulation capability prediction and pressure optimization method for thermal power units according to a specific embodiment of the present invention. Figure 7 This is a scatter plot comparing the main steam pressure before, after, and following the optimization of the primary frequency regulation capability prediction and pressure optimization method for thermal power units according to a specific embodiment of the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown in the figure, a method for predicting the primary frequency regulation capability and optimizing the pressure of a thermal power unit provided by a specific embodiment of the present invention includes the following steps: S1: Within the first time window T1 before the frequency difference of the power grid crosses the preset dead zone, collect multi-source operating parameters of the thermal power unit and construct the pre-regulation state vector of the primary frequency regulation. S2: Based on the maximum frequency difference, average active power and average main steam pressure in the pre-regulation state vector of the first frequency regulation, the frequency regulation demand scenario is divided into the current operating state of the thermal power unit. S3: For different frequency regulation demand scenarios, construct corresponding fusion models, extract predetermined demand features as model input within the second time window T2 before the frequency difference crosses the dead zone, and output the model with the predicted value of the actual integral power of the primary frequency regulation of the thermal power unit, so as to realize the online prediction of the primary frequency regulation capability. S4: Calculate the frequency regulation contribution rate based on the actual integrated power of primary frequency regulation and the theoretical integrated power of primary frequency regulation. When the frequency regulation contribution rate is lower than the preset threshold, it is judged as a frequency regulation failure event. For the frequency regulation failure event, the main steam pressure setpoint of the thermal power unit is optimized in a rolling manner. The optimization target is to ensure that the corrected actual integrated power of primary frequency regulation, the change rate of main steam pressure and the change rate of valve meet the safety constraints. In the rolling optimization process, the fusion model is called to evaluate the integrated power in real time. S5: Send the optimized main steam pressure setpoint to the thermal power unit coordination control system to improve the primary frequency regulation qualification rate.

[0019] Specifically, multi-source operating parameters are collected in real time through the power plant's SIS system, and the data is aligned based on a unified timestamp to construct a time-series data sequence arranged at second-level sampling intervals. The collected data is processed to remove outliers and fill in missing values. Outliers are identified and removed using the 3σ criterion, and missing values ​​are compensated using the sliding window mean to ensure the continuity and accuracy of the preceding state vector.

[0020] Furthermore, the preceding state vector is a combination vector of the time-series characteristics of each operating parameter within the first time window T1, used to reflect the load level, heat storage state, and pressure operating range of the thermal power unit before the frequency regulation action occurs.

[0021] Furthermore, the first time window T1 covers the process of operating status changes before the frequency regulation dead zone boundary is triggered, which is used to capture the impact of the boiler's heat storage release delay characteristics on the frequency regulation response capability.

[0022] Based on the above basic implementation method, the multi-source operating parameters include the unit's speed, frequency, frequency difference signal, actual active power, main steam pressure setpoint, actual main steam pressure, main steam temperature, AGC command, turbine GV valve opening command, actual total coal quantity, and boiler main control output; the first time window T1 is within 10 seconds, and the sampling period of the first time window T1 is no more than 1 second.

[0023] Furthermore, before constructing the preceding state vector, the maximum frequency difference, average active power, and average main steam pressure are Z-score standardized to eliminate the influence of different dimensions on the subsequent clustering results.

[0024] In one specific implementation, in step S2, for the multi-dimensional operating characteristic data consisting of the maximum frequency difference, average active power, and average main steam pressure in the pre-frequency regulation state vector of the thermal power unit, an improved K-means algorithm is used to cluster the current frequency regulation operating state of the thermal power unit. The improved K-means algorithm introduces a feedback correction term based on the clustering profile coefficient into the traditional K-means loss function to dynamically evaluate the clustering quality of the operating data of the thermal power unit under different loads and heat storage states. During the iteration process, when the clustering profile coefficient is lower than a set threshold, the initial cluster center is reselected based on the local density distribution of the thermal power unit operating samples to improve the distinguishability between different frequency regulation demand scenarios and reduce the impact of noisy operating samples on the boundary of the frequency regulation scenario.

[0025] In this embodiment, the number of clusters is preferably 3, corresponding to low load and small frequency difference scenarios, medium load and medium frequency difference scenarios, and high load and small frequency difference scenarios, respectively. Each scenario corresponds to different heat storage and release capabilities and frequency regulation response capabilities.

[0026] Furthermore, an independent prediction model is established for each type of frequency regulation demand scenario to improve the prediction accuracy under different operating conditions.

[0027] In another specific embodiment, in step S3, the fusion model is a CNN-BiLSTM-Attention fusion model, which includes: at least two one-dimensional convolutional neural networks for extracting local dynamic features of the input sequence; a bidirectional long short-term memory network for capturing the time-series dependency characteristics caused by the boiler thermal storage inertia of the thermal power unit; an attention mechanism for weighted fusion of features at different time steps; and a fully connected output layer for outputting the predicted value of the actual integrated power of the primary frequency modulation.

[0028] In this embodiment, in step S3, the second time window T2 is less than or equal to the first time window T1; the model input is a sequence of key features from multiple consecutive time steps before the frequency modulation action.

[0029] Furthermore, the fusion model is trained under supervision using historical valid FM event data, with the training samples divided into a 60% training set, a 10% validation set, and a 30% test set.

[0030] Specifically, the model training uses the Adam optimization algorithm, with a learning rate of 0.005, and an early stopping mechanism is set to prevent overfitting.

[0031] In another specific embodiment, in steps S3 and S4, the actual integrated power of primary frequency regulation is the time integral value of the change in the actual active power of the unit during the period from when the frequency difference crosses the dead zone to when it returns to the dead zone; the theoretical integrated power of primary frequency regulation is calculated based on the rated power of the thermal power unit, the droop coefficient, and the frequency difference function; the frequency regulation contribution rate is the ratio of the actual integrated power to the theoretical integrated power; the actual integrated power of primary frequency regulation is:

[0032] In the formula: This represents the actual integrated power consumption during a single frequency modulation. To exceed the dead zone The actual active power generated by the unit at that moment is the average value of the 3 seconds preceding that moment. To return to the dead zone The actual active power generated by the unit at any given time; The integration time is: ; .

[0033] The theoretical integral charge of a single frequency modulation is:

[0034] In the formula: This is the theoretical integral charge of a single frequency modulation. For unit frequency and The difference in Hz is positive for high frequencies and negative for low frequencies; This refers to the rated power, specifically 50Hz; The droop factor set for the unit is 5%; The rated power of the unit is 650MW; The primary frequency modulation contribution rate is:

[0035] In the formula: Contribution rate to frequency modulation.

[0036] Corrected integral energy objective function for: When the power grid frequency is below 50Hz

[0037] When the power grid frequency is higher than 50Hz

[0038] In the formula: , , These are the weighting coefficients; , , These represent the deviation of the main steam pressure amplitude, the deviation of the pressure change rate, and the deviation of the change rate of valves GV1-GV4, respectively. Furthermore, the theoretical integral energy is used to characterize the level of energy support that the unit should provide under ideal frequency regulation conditions.

[0039] Furthermore, when the frequency modulation contribution rate is less than 50%, it is judged as a frequency modulation failure event.

[0040] In another specific embodiment, in step S4, the corrected integral power is the frequency regulation capability evaluation index obtained by introducing the deviation of the main steam pressure amplitude and the penalty factor of the main steam pressure change rate of the thermal power unit based on the original actual integral power of the primary frequency regulation. The improved AGWO-CS hybrid optimization algorithm is used for rolling optimization of the main steam pressure setpoint of the thermal power unit. The main steam pressure setpoint is used as the variable to be optimized, and the corrected integral power is used as the fitness function to search and optimize the operating parameter space of the thermal power unit under different loads and heat storage conditions. The improved AGWO-CS hybrid optimization algorithm includes: using chaotic mapping to initialize the feasible region of the main steam pressure setpoint to form a gray wolf population; introducing Levy flight disturbances during the gray wolf individual position update process to enhance the global search capability in the thermal power unit pressure regulation space; adjusting the search step size during the optimization process through a nonlinear convergence factor to balance the rapid regulation requirements of the thermal power unit's main steam pressure and the stable operation requirements of the thermal power unit; and using a cuckoo search fitness update mechanism to replace the individual with the lowest fitness in the gray wolf population to improve the convergence efficiency of the thermal power unit's main steam pressure optimization.

[0041] Specifically, the Chebyshev chaotic mapping is used to initialize the gray wolf population to improve the uniformity of the initial solution space, as follows:

[0042] In the formula: The generated chaotic sequence Used to initialize population location ,Right now

[0043] In the formula: , These are the upper and lower bounds of the population solution space, respectively; Levy flight perturbations are embedded in the AGWO position update to enhance the algorithm's ability to escape local optima, as detailed below:

[0044] In the formula: The product of Hadamard; This is the step scaling factor, used to control the step size; the default value is 1. To obey Random variables with probability distributions; Introducing a nonlinear convergence factor Balance global exploration with local development; In the formula: , To explore the minimum and maximum values ​​of the factors; This represents the current iteration number; The maximum number of iterations is determined by replacing the worst wolf individual update strategy in AGWO with the fitness evaluation mechanism of the cuckoo search, thereby shortening the number of iterations.

[0045] Furthermore, the feasible range of the main steam pressure setpoint is determined based on the unit's rated pressure and the upper limit of safe operation; during the rolling optimization process, the candidate pressure setpoint is predicted by integrating the power consumption in real time by calling the fusion model, thus realizing a closed loop of prediction-optimization-evaluation.

[0046] In one specific implementation, in step S4, an optimization objective function for the thermal power unit is constructed based on the improved AGWO-CS hybrid optimization algorithm. The optimization objective function takes the corrected integral power as the main objective and introduces the deviation of the main steam pressure amplitude, the deviation of the main steam pressure change rate, and the deviation of the turbine GV valve change rate as penalty factors. A multi-objective weighted function is formed by setting weight coefficients. In step S5, a safety protection logic is set before the optimized main steam pressure setpoint is issued. When the deviation between the optimized main steam pressure setpoint and the current actual pressure exceeds the preset upper limit or the pressure change rate exceeds the preset change rate threshold, the optimization result is automatically locked and the original setpoint remains unchanged.

[0047] In this embodiment, when the optimization result is locked, the system automatically switches to the preset safety pressure curve mode and records relevant log information.

[0048] Specifically, the security protection logic communicates with the DCS system through a one-way isolation channel to meet the security protection requirements of the power monitoring system.

[0049] In one specific embodiment, the present invention provides a primary frequency regulation capability prediction and pressure optimization system for thermal power units, comprising: The system comprises the following modules: a data acquisition module for acquiring multi-source operating parameters of the thermal power unit and constructing a preceding state vector; a scenario segmentation module for executing an improved K-means algorithm to segment frequency regulation demand scenarios; a prediction module with a built-in CNN-BiLSTM-Attention fusion model for outputting the predicted value of the actual integrated power generation for primary frequency regulation; a judgment module for calculating the frequency regulation contribution rate and determining whether a frequency regulation failure event has occurred; an optimization module for executing an improved AGWO-CS hybrid optimization algorithm and outputting the optimized main steam pressure setpoint; and a distribution module for writing the optimized main steam pressure setpoint into the CCS-side pressure controller of the thermal power unit to achieve closed-loop regulation.

[0050] Specifically, the prediction module and optimization module are deployed in the power plant edge computing server, the data acquisition module interfaces with the SIS system through the OPC UA protocol, and the distribution module is connected to the DCS system through a one-way isolation device. Continuous operation test results show that the present invention can significantly improve the primary frequency regulation qualification rate of thermal power units and reduce the main steam pressure fluctuation amplitude.

[0051] In one specific implementation, based on the collected historical operating data of the unit, and in accordance with the requirements of the "two detailed rules" for the Central China region, the primary frequency regulation dead zone is set to Δf=0.033Hz, and the effective primary frequency regulation data segment is extracted. Combined with the analysis of the primary frequency regulation mechanism and thermodynamic characteristics of thermal power units, it is determined that the main steam pressure, boiler heat storage characteristics, turbine valve opening, and total boiler coal quantity are the core factors affecting the primary frequency regulation capability, providing a theoretical basis for subsequent modeling and optimization.

[0052] The two detailed rules for the Central China region require the extraction of primary frequency modulation data segments as follows:

[0053] In the formula: for Time frequency; This is the rated frequency, specifically 50Hz.

[0054] When the frequency deviation exceeds the FM dead zone and persists for more than 10 seconds, it is considered a valid primary FM event. At this time, Data within the specified time period is included in the data range; otherwise, it is determined that no frequency modulation action occurred. Missing values ​​are filled with the mean of the preceding and following data to obtain stable operating data covering the frequency modulation process.

[0055] Based on the primary frequency regulation mechanism and characteristics of thermal power units, the main steam pressure, heat storage characteristics, valve opening degree, and total coal quantity are identified as key factors affecting the primary frequency regulation capability.

[0056] Frequency modulation demand scenario segmentation based on improved K-means algorithm Based on the maximum frequency difference, average active power, and average main steam pressure in the preceding state vector, an improved K-means algorithm is used to divide the frequency regulation demand scenarios, capture the grid frequency difference signal in real time, and use the maximum frequency difference 10 seconds before the frequency regulation action as the core indicator to complete the subspace classification of operating conditions in milliseconds.

[0057] The traditional K-means algorithm classifies samples by minimizing a clustering loss function, which is defined as the sum of squared Euclidean distances between samples within each cluster and the cluster centers:

[0058] In the formula: The number of clusters. For the first The sample set of the cluster, For the first A sample vector, For the first Cluster center vector of a class, This indicates Euclidean distance.

[0059] The traditional K-means algorithm suffers from drawbacks such as sensitivity to initial cluster centers and susceptibility to noise samples leading to blurred cluster boundaries. To address these issues, this embodiment proposes two improvements: The improved K-means algorithm is as follows: Introducing a feedback correction term to optimize the loss function: A feedback correction term based on the clustering silhouette coefficient is added to the traditional loss function. By dynamically adjusting the clustering process through real-time evaluation of clustering quality, the improved loss function is as follows:

[0060] In the formula: The feedback correction coefficient has a value range of 0.1 to 0.3 (0.2 in this embodiment). The cluster profile coefficient is used to quantify cluster quality. The closer the value is to 1, the better the clustering effect.

[0061] Adaptive adjustment of initial cluster centers: During the iteration process, if the current cluster silhouette coefficients... Less than the set threshold (In this embodiment, we take 0.6). Then, based on the sample density distribution, we reselect the initial cluster centers. The specific steps are: calculate the local density of each sample. ( To truncate the distance, the median of the distances between samples is taken. (for indicator functions), before selecting local density Larger samples are used as new initial cluster centers to achieve adaptive optimization of the initial centers.

[0062] The improved K-means algorithm introduces a feedback correction term into the traditional K-means loss function. During the iteration process, it adaptively adjusts the initial cluster centers based on the cluster silhouette coefficients, reducing the interference of noise samples on scene boundaries. This solves the problem of traditional K-means being sensitive to initial centers and easily affected by noise, achieving accurate segmentation of different frequency modulation scenarios. The elbow rule determines the optimal number of clusters to be 3, covering typical frequency modulation scenarios such as low load with small frequency difference, medium load with medium frequency difference, and high load with small frequency difference.

[0063] For three-dimensional key parameters ( , , Z-score standardization is performed to eliminate the influence of dimensional differences. The standardized formula is as follows:

[0064] In the formula: These are the original parameter values. The mean of the parameters, For the standard deviation of the parameter, This is the standardized value.

[0065] Determine the optimal number of clusters The optimal number of clusters is determined by combining the elbow rule and the silhouette coefficient method. Different cluster numbers are calculated. ( The corresponding loss function value ,draw The curve (i.e., the elbow curve), when After increasing to a certain value The rate of descent slows significantly; this point is the elbow inflection point. In this embodiment, when... At this point, the downward trend of the elbow curve slows down significantly, and the cluster profile coefficient also decreases. To achieve optimal clustering quality, the optimal number of clusters is determined. .

[0066] By improving the K-means algorithm, the selected valid primary frequency modulation event samples are clustered to obtain three typical frequency modulation demand scenarios. The cluster center parameters are shown below: Category 1 scenario (low load - small frequency difference scenario): In this scenario, the unit load is low and the frequency disturbance amplitude is small. The core requirement is to make full use of limited thermal storage resources to respond to small frequency fluctuations. Category 2 scenario (medium load - medium frequency difference scenario): In this scenario, the unit load is moderate and the frequency disturbance amplitude is moderate, requiring a balance between heat storage release and continuous response capability. Category 3 scenario (high load - small frequency difference scenario): In this scenario, the unit load is high and the heat storage resources are sufficient, so it is necessary to control the response speed to avoid power oscillation.

[0067] Frequency modulation capability prediction based on CNN-BiLSTM-Attention fusion model For different scenarios, CNN-BiLSTM-Attention fusion models are established respectively. These models use key features from the time interval T2≤T1 before the frequency difference crosses the dead zone as input and the actual integrated charge during frequency modulation as output, enabling online prediction of frequency modulation capability. In this invention, T1 is set to 10s and T2 to 5s, meaning that key operating parameters from 5s before the frequency modulation action are used as input. Spatial features are extracted via a convolutional neural network, and a bidirectional long short-term memory network learns the time-heat storage coupling dynamics. The heat storage response delay is corresponding to a 5-segment time window, and an attention mechanism adaptively weights the input, outputting the actual integrated charge as the quantified value of the frequency modulation capability.

[0068] The CNN-BiLSTM-Attention fusion model specifically includes: Two one-dimensional convolutional layers are used to extract local dynamic features from the preceding state vector, such as load changes and frequency fluctuations; a bidirectional LSTM layer is used to capture the long-term dependence caused by boiler thermal inertia and adapt to the 3-8 second inertial delay characteristics of thermal response; an attention mechanism is used to adaptively weight key feature channels and strengthen the feature weights that have a significant impact on frequency regulation capability; and a fully connected output layer is used to output the predicted value of actual integrated power consumption for frequency regulation.

[0069] The convolutional layer is:

[0070] In the formula: This represents a one-dimensional convolution operation. The convolution kernel weight matrix is... Kernel size (Layer 1) ), second floor ), This is the number of output channels for the convolution (set to 64 for both layers). For bias vectors, It is the ReLU activation function. , The length of the input sequence ( ), For the first The input subsequence of a sliding window, For the first The convolutional output features of each window.

[0071] The bidirectional LSTM layer is specifically as follows: Forward LSTM state update:

[0072] Backward LSTM state update:

[0073] In the formula: , , ) are the activation vectors for the input gate, forget gate, and output gate, respectively. Candidate cell state, Let be the cell state vector. The hidden state vector; superscript These represent forward and backward LSTM respectively; , , and , respectively, represent the corresponding weight matrix, recursive weight matrix, and bias vector; This represents element-wise product.

[0074] The aforementioned attention mechanism is specifically as follows: bidirectional LSTM output features With attention weight matrix ( Multiplying by the dimension of the attention hidden layer (where tanh is the number of dimensions) and then tanh activation yields the attention hidden state. .

[0075] In the formula: This is the attention bias vector.

[0076] The attention score is normalized using the softmax function to obtain the attention weights at each time step.

[0077] In the formula: For attention score vectors,

[0078] The attention weights are summed with the bidirectional LSTM output features to obtain the final attention fusion features. :

[0079] In the formula: , which is the output vector of the attention mechanism layer, realizes the enhancement and fusion of key time step features.

[0080] The aforementioned fully connected output layer is specifically: Attention fusion features Mapped to actual integrated energy prediction value .

[0081] In the formula: This is the weight matrix of the fully connected layer. This is a bias term.

[0082] The model is implemented using Python 3.9, with ReLU as the activation function and Adam as the optimizer. The initial learning rate is set to 0.005, and the training iteration adopts an early stopping mechanism with a maximum of 200 iterations to ensure the model's prediction accuracy and real-time performance.

[0083] For each type of frequency modulation scenario, key feature variables obtained through improved principal component analysis are used to construct the input feature sequence. Let the input feature vector be... ,in (Corresponding to 5 time steps, with a 1-second interval between each step). Let be the feature vector at second t, and D be the key feature dimension for this scenario (D=5 for scenario 1, and the other scenarios are determined based on the feature selection results). The input sequence dimension is [T,D], i.e. [5,D].

[0084] The datasets for each scenario are randomly divided into 60% training set, 10% validation set, and 30% test set to ensure consistent data distribution. The input feature sequences are normalized using Min-Max normalization to map the feature values ​​to the [0,1] interval.

[0085] The normalization formula is as follows:

[0086] In the formula: These are the original eigenvalues. , These are the minimum and maximum values ​​of the feature in the training set, respectively. These are the normalized eigenvalues.

[0087] The model output is the actual integral power of the first frequency modulation. This indicator directly reflects the energy output of the unit during the frequency regulation process and is the core quantitative indicator of primary frequency regulation capability.

[0088] Mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) are used. 2 The predictive performance of the model is evaluated using a combination of four indicators.

[0089] The model evaluation metrics are as follows:

[0090] In the formula: The number of samples in the test set. The average of the true values ​​in the test set. The closer to 1, the better the model fit.

[0091] Specifically, for non-compliant events, the improved AGWO-CS hybrid optimization algorithm is used to perform rolling optimization of the main steam pressure setpoint. The optimization objective is to maximize the corrected integral power while ensuring that the pressure change rate and the GV valve change rate do not exceed the limits.

[0092] The improved AGWO-CS hybrid optimization algorithm can be further divided into the following steps: A1: The gray wolf population is initialized using the Chebyshev chaotic mapping to improve the uniformity of the initial solution space; A2: Embed Levy flight perturbation in AGWO position update to enhance the algorithm's ability to escape local extrema; A3: Introducing a nonlinear convergence factor to balance global exploration and local exploitation; A4: Replace the worst wolf individual update strategy in AGWO with the fitness evaluation mechanism of cuckoo search to shorten the number of iterations.

[0093] Optimize setting value distribution and frequency modulation capabilities The optimized main steam pressure is issued as the new setpoint on the CCS side. Driven by the capability assessment model, the main steam pressure setpoint is optimized to resolve frequency regulation failures caused by improper pressure settings, thereby improving frequency regulation capability and frequency regulation pass rate.

[0094] Furthermore, the present invention also provides a system for predicting the primary frequency regulation capability and optimizing the pressure of a thermal power unit. The system employs the aforementioned method for predicting the primary frequency regulation capability and optimizing the main steam pressure of a thermal power unit, taking into account time and heat storage factors. The system includes: Data acquisition module: used to acquire the above multi-source operating parameters in real time, and interface with the power plant's SIS system via OPC UA protocol, with a sampling period of ≤1s; Scene segmentation module: Used to execute the improved K-means algorithm mentioned above to achieve accurate segmentation of frequency modulation demand scenarios; Prediction module: Built-in CNN-BiLSTM-Attention model to output frequency modulation integral power prediction value; Judgment module: Used to determine whether the frequency modulation is qualified based on the above contribution rate threshold value; Optimization module: Used to run the improved AGWO-CS algorithm and output the optimized main steam pressure setpoint. This module is deployed in the safety zone II of the power plant MIS network segment, adopts a containerized microservice architecture, and is unidirectionally isolated from the DCS side to meet the security protection requirements of the power monitoring system. The distribution module is used to write the optimized setpoints into the CCS-side pressure controller. It features override protection logic: the threshold is determined based on calculations of the unit's material tolerance limits and grid frequency stability requirements; after locking, it automatically switches to the "heat storage priority" temporary pressure curve to avoid a precipitous drop in frequency regulation capability, forming a complete "optimization-protection-emergency" closed loop. When the optimized pressure deviates from the current actual pressure by more than ±1.0 MPa or the pressure change rate exceeds ±0.3 MPa / min, the optimization result is automatically locked and an alarm is triggered.

[0095] In a specific implementation, this embodiment uses a 650MW ultra-supercritical once-through boiler unit in a power plant in Central China as the application object. This unit adopts a variable-pressure spiral coil once-through boiler, with a rated active power of 650MW, a rated main steam pressure of 25.4MPa, a rated main steam temperature of 571℃, and a rated actual total coal consumption of 284t / h. It is equipped with a turbine digital electro-hydraulic control system (DEH) and a boiler-turbine coordination control system (CCS), meeting the basic operating requirements of primary frequency regulation in the power grid. Its primary frequency regulation process structure optimized for main steam pressure is as follows: Figure 2 , Figure 3 As shown.

[0096] In this embodiment, system setup and data acquisition specifically include: The hardware system consists of a data acquisition module, an edge computing server, a SIS system interface, and a CCS-side pressure control interface, with the following specific configuration: Data acquisition module: adopts an industrial-grade data acquisition gateway, and interfaces with the power plant's SIS system via the OPC UA protocol. The sampling period is set to 1 second, and the data transmission delay is ≤50ms, which meets the real-time requirements. Edge computing server: Equipped with an Intel Core i7-12700 processor, 32GB DDR4 memory and a 1TB NVMe solid-state drive, running Ubuntu 20.04 operating system, used to deploy scene segmentation algorithms, CNN-BiLSTM-Attention prediction models and improved AGWO-CS optimization algorithms; Secure transmission channel: The optimization results are written to the CCS-side pressure controller through a one-way isolation gateway, which complies with the requirements of the "Regulations on Security Protection of Power Monitoring Systems" and avoids network security risks; Monitoring terminal: Equipped with an industrial-grade touch screen, it displays model prediction results, optimization parameters, and frequency tuning compliance status in real time, and supports abnormal alarms and manual intervention.

[0097] Based on the analysis results of factors affecting primary frequency regulation capability, 14 key operating parameters of the unit were collected through the SIS system. The specific parameter information and units are shown in Table 1. Table 1 Key Operating Parameters of the Unit

[0098] The data collection process covers multiple complete primary frequency regulation events under normal unit operation, with no downtime or fault records. The data spans 30 consecutive days, providing sufficient and reliable data support for model training and optimization.

[0099] In this embodiment, data preprocessing and scene segmentation specifically include: According to the "two detailed rules" for the Central China region, a primary frequency modulation dead zone Δf = 0.033Hz is set. When the frequency deviation exceeds the dead zone and lasts for more than 10 seconds, it is determined to be a valid primary frequency modulation event, and data for the corresponding time period are selected for inclusion in the analysis. The specific data preprocessing process is as follows: S1 Outlier Removal: Outlier data is identified using the 3σ criterion, and parameter values ​​that exceed the range of [μ-3σ, μ+3σ] are removed, where μ is the parameter mean and σ is the standard deviation; S2 Missing Value Imputation: Missing values ​​in the data are imputed using a moving average method across five consecutive time steps. S3 data synchronization: Aligns 14 parameter data based on timestamps to ensure that parameter data at the same time correspond one-to-one, with a synchronization error of ≤10ms.

[0100] By improving the K-means clustering algorithm, three typical frequency modulation demand scenarios were obtained, and the cluster center parameters are shown in Table 2: Table 2 Cluster Center Results

[0101] The improved K-means algorithm has a clustering time of only 87ms, which meets the real-time requirement of millisecond-level subspace classification. Moreover, the silhouette coefficient of the clustering result is 23.8% higher than that of the traditional K-means algorithm, effectively reducing the interference of noise samples on scene boundaries.

[0102] For each type of scenario data, an improved principal component analysis (PCA) method is used to filter input features. The specific steps are as follows: S1 data dimensionality reduction: Perform 5-fold cross-validation before transforming to the principal component space to determine the optimal number of principal components; S2 Principal Component Extraction: Extract principal components with a cumulative contribution rate of ≥85% to ensure that the core features of the data are preserved; S3 Feature Selection: Select feature variables with loading values ​​> 0.8 from each principal component as model input.

[0103] Taking scenario 1 as an example, the cumulative contribution rate of the five extracted principal components reached 86.83%. The feature load values ​​of each principal component are shown in Table 3. Finally, the actual total coal quantity, boiler main control output, AGC command, main steam pressure setpoint, and GV4 command were selected as key input features (D=5). The other two scenarios were selected using the same method. Scenario 2 yielded six key features, and scenario 3 yielded five key features.

[0104] Table 3. Principal Component Feature Load Values ​​for Scenario 1

[0105] In this embodiment, the construction and training of the CNN-BiLSTM-Attention prediction model specifically includes: Network model framework diagram as follows Figure 4 As shown, the actual integral power prediction results for the three types of frequency regulation demand scenarios are as follows: Figure 5 As shown in Table 4, the prediction performance metrics of the model in the three scenarios are compared. Combining the three modules can better capture features. The model's R² exceeds 0.93 in all scenarios, with R² reaching 0.968 in the third scenario. The hyperparameter settings of different models are shown in Table 5, and the prediction results are compared. Figure 6 As shown, the prediction accuracy of the model of the present invention is significantly better than that of traditional LSTM, GRU and Transformer models, and the model prediction latency is only 20ms, which meets the real-time requirements of online prediction.

[0106] Table 4 Comparison of the performance of each sub-model of CNN-BiLSTM-Attention

[0107] Table 5 Hyperparameter settings for different comparison models

[0108] In this embodiment, the optimization of the main steam pressure based on the improved AGWO-CS specifically includes: The optimization algorithm includes the following steps: S1 initialization: 30 initial gray wolf individuals are generated using chaotic mapping, and the iteration number T is set to 0; S2 fitness calculation: Calculate the corrected integral charge for each individual. , as the fitness value; S3 Wolf Pack Ranking: Based on fitness values, determine α, β, δ wolves (the top 3 superior individuals) and ω wolves (the remaining individuals). S4 position update: α, β, and δ wolves update their positions according to the AGWO rule, while ω wolf updates its position through Levy flight perturbation; S5 Cuckoo Selection: Calculate the fitness of the new individual; if it is better than the original ω-wolf, replace it; otherwise, keep the original individual. S6 Convergence Criterion: If T reaches... Or fitness change <10 -4 Output the optimal individual value (optimal main steam pressure setting); otherwise, T=T+1, return to step S2.

[0109] A main steam pressure following model was established by selecting real operating data of the unit for one consecutive month. The real pressure setpoint was input into the model, and the goodness of fit between the output and the real pressure tracking curve reached 0.987, proving that the model can be used to verify the optimization results.

[0110] To address frequency regulation failures caused by improper main steam pressure settings in three scenarios, an improved AGWO-CS algorithm is used for rolling optimization. The comparison chart of the main steam pressure before, after, and following the optimization is shown below. Figure 7 As shown in Table 6, the optimization results for frequency regulation events caused by improper main steam pressure settings are as follows: Table 6 Optimization results for frequency regulation events caused by improper main steam pressure setting

[0111] As shown in the table, the overall optimization effect reached 92.74%, and the main steam pressure fluctuation range was reduced from ±0.8MPa before optimization to ±0.5MPa, a reduction of 37.5%. Under the premise of ensuring the safe operation of the unit, the frequency regulation qualification rate was significantly improved.

[0112] The improved AGWO-CS algorithm was compared with the GA, AGWO, AGWO-CS and NSGA-II algorithms. The results are shown in Table 7. The improved AGWO-CS algorithm performed best in terms of optimization accuracy, convergence speed and optimization effect. The average total number of iterations was only 28 and the total optimization time was 152.32s, which is more than 40% shorter than the traditional algorithm.

[0113] Table 7 Comparison of the effects of different algorithms

[0114] In this embodiment, optimizing the issuance and operation monitoring of the main steam pressure setpoint specifically includes: The optimized main steam pressure setpoint is written to the CCS-side pressure controller via a distribution module, which has the following functions: Override protection: When the deviation between the optimized pressure and the current actual pressure exceeds ±1.0MPa or the pressure change rate exceeds ±0.3MPa / min, the optimization result will be automatically locked and an audible and visual alarm will be issued, while the original set value will remain unchanged. Manual intervention: Operators can manually switch between "automatic optimization" and "manual setting" modes. In manual mode, pressure setting values ​​can be directly entered. History Records: Automatically records the settings, execution time, optimization effects, and alarm information for each optimization. The storage period is 1 year, and it supports querying and exporting.

[0115] In actual operation, the pressure following model monitors the setpoint tracking effect in real time, updating the monitoring data every 1 second to ensure the effective implementation of the optimization strategy. Results from 30 consecutive days of trial operation show that the unit's frequency regulation qualification rate increased from 68.3% before optimization to 96.7%, and no abnormal fluctuations in the unit caused by parameter optimization occurred, verifying the reliability and practicality of the method of this invention.

[0116] This embodiment verifies that the method proposed in this invention can achieve accurate prediction of the primary frequency regulation capability of thermal power units and dynamic optimization of the main steam pressure. It effectively solves the problems of low evaluation accuracy, weak generalization ability, and frequency regulation failure caused by improper pressure setting in traditional methods. It takes into account both the frequency regulation requirements of the power grid and the life-energy consumption index of the unit, and can be widely applied to the improvement project of primary frequency regulation performance of thermal power units under the background of deep peak shaving and flexibility transformation.

[0117] To aid in a better understanding of the present invention, a more comprehensive and specific embodiment is described, in which the present invention provides a method for predicting the primary frequency regulation capability and optimizing the pressure of thermal power units, comprising the following steps: S1: Within the first time window T1 before the frequency difference of the power grid crosses the preset dead zone, collect multi-source operating parameters of the thermal power unit and construct the pre-regulation state vector of the primary frequency regulation. S2: Based on the maximum frequency difference, average active power and average main steam pressure in the pre-regulation state vector of the first frequency regulation, the frequency regulation demand scenario is divided into the current operating state of the thermal power unit. S3: For different frequency regulation demand scenarios, construct corresponding fusion models, extract predetermined demand features as model input within the second time window T2 before the frequency difference crosses the dead zone, and output the model with the predicted value of the actual integral power of the primary frequency regulation of the thermal power unit, so as to realize the online prediction of the primary frequency regulation capability. S4: Calculate the frequency regulation contribution rate based on the actual integrated power of primary frequency regulation and the theoretical integrated power of primary frequency regulation. When the frequency regulation contribution rate is lower than the preset threshold, it is judged as a frequency regulation failure event. For the frequency regulation failure event, the main steam pressure setpoint of the thermal power unit is optimized in a rolling manner. The optimization target is to ensure that the corrected actual integrated power of primary frequency regulation, the change rate of main steam pressure and the change rate of valve meet the safety constraints. In the rolling optimization process, the fusion model is called to evaluate the integrated power in real time. S5: Send the optimized main steam pressure setpoint to the thermal power unit coordination control system to improve the primary frequency regulation qualification rate.

[0118] In this embodiment, the multi-source operating parameters include the unit's rotational speed, frequency, frequency difference signal, actual active power, main steam pressure setpoint, actual main steam pressure, main steam temperature, AGC command, turbine GV valve opening command, actual total coal consumption, and boiler main control output; the first time window T1 is within 10 seconds, and the sampling period of the first time window T1 is no more than 1 second; in step S2, for the multi-dimensional operating feature data composed of the maximum frequency difference, average active power, and average main steam pressure in the pre-frequency regulation state vector, an improved K-means algorithm is used to cluster the current frequency regulation operating state of the thermal power unit; the improved K-means algorithm reduces the traditional K-means loss function... A feedback correction term based on the clustering silhouette coefficient is introduced to dynamically evaluate the clustering quality of the operating data of the thermal power unit under different loads and heat storage conditions. During the iteration process, when the clustering silhouette coefficient is lower than a set threshold, the initial clustering centers are reselected based on the local density distribution of the thermal power unit operating samples to improve the distinguishability between different frequency regulation demand scenarios and reduce the impact of noisy operating samples on the frequency regulation scenario boundary. In step S3, the fusion model is a CNN-BiLSTM-Attention fusion model, which includes: at least two layers of one-dimensional convolutional neural networks for extracting local dynamic features of the input sequence; and one layer of bidirectional long short-term memory network. The system is designed to capture the time-series dependence characteristics caused by the boiler thermal storage inertia of the thermal power unit; an attention mechanism is used to perform weighted fusion of features from different time steps; a fully connected output layer is used to output the predicted value of the actual integrated power of the primary frequency regulation; in step S3, the second time window T2 is less than or equal to the first time window T1; the model input is the key feature sequence of multiple consecutive time steps before the frequency regulation action; in steps S3 and S4, the actual integrated power of the primary frequency regulation is the time integral value of the change in the actual active power of the unit during the period from the frequency difference crossing the dead zone to returning to the dead zone; the theoretical integrated power of the primary frequency regulation is calculated based on the rated power of the thermal power unit, the droop coefficient, and the frequency difference function; the frequency regulation contribution rate is the actual integrated power. The ratio to the theoretical integral power; in step S4, the corrected integral power is the frequency regulation capability evaluation index obtained by introducing the deviation of the main steam pressure amplitude and the penalty factor of the main steam pressure change rate of the thermal power unit on the basis of the original actual integral power of the primary frequency regulation; the improved AGWO-CS hybrid optimization algorithm is used to perform rolling optimization of the main steam pressure setpoint of the thermal power unit, with the main steam pressure setpoint as the variable to be optimized and the corrected integral power as the fitness function, to search and optimize the operating parameter space of the thermal power unit under different loads and heat storage conditions; the improved AGWO-CS hybrid optimization algorithm includes: using chaotic mapping to initialize the feasible region of the main steam pressure setpoint to form a gray wolf population;Levy flight disturbances are introduced during the gray wolf individual position update process to enhance the global search capability in the thermal power unit pressure regulation space; the search step size in the optimization process is adjusted by a nonlinear convergence factor to balance the requirements for rapid regulation of the main steam pressure of the thermal power unit with the requirements for stable operation of the thermal power unit; the fitness update mechanism of the cuckoo search is used to replace the individual with the lowest fitness in the gray wolf population to improve the convergence efficiency of the thermal power unit main steam pressure optimization; in step S4, the optimization objective function of the thermal power unit is constructed according to the improved AGWO-CS hybrid optimization algorithm. The optimization objective function takes the corrected integral power as the main objective and introduces the deviation of the main steam pressure amplitude, the deviation of the main steam pressure change rate, and the deviation of the turbine GV valve change rate as penalty factors, and forms a multi-objective weighted function by setting weight coefficients; in step S5, a safety protection logic is set before the optimized main steam pressure setpoint is issued. When the deviation between the optimized main steam pressure setpoint and the current actual pressure exceeds the preset upper limit or the pressure change rate exceeds the preset change rate threshold, the optimization result is automatically locked and the original setpoint remains unchanged.

[0119] Specifically, this invention provides a primary frequency regulation capability prediction and pressure optimization system for thermal power units, comprising: a data acquisition module for acquiring multi-source operating parameters of the thermal power unit and constructing a preceding state vector; a scenario segmentation module for executing an improved K-means algorithm to segment frequency regulation demand scenarios; a prediction module with a built-in CNN-BiLSTM-Attention fusion model for outputting the predicted value of the actual integrated power generation for primary frequency regulation; a judgment module for calculating the frequency regulation contribution rate and determining whether a frequency regulation failure event has occurred; an optimization module for executing an improved AGWO-CS hybrid optimization algorithm and outputting the optimized main steam pressure setpoint; and a distribution module for writing the optimized main steam pressure setpoint into the pressure controller on the CCS side of the thermal power unit to achieve closed-loop regulation.

[0120] In summary, the embodiments disclosed herein have at least the following technical effects: This invention combines the time-series operational characteristics of thermal power units before frequency regulation with the dynamic characteristics of boiler thermal storage to construct a fusion prediction model based on multi-source operational data, enabling online prediction of the actual integrated power generation during primary frequency regulation. Compared to traditional prediction methods based on mechanistic models or single data models, this invention fully considers the differences in frequency regulation capabilities of units under different load levels and thermal storage conditions, significantly improving prediction accuracy and cross-condition adaptability, and providing a reliable basis for subsequent frequency regulation capability determination and parameter optimization. This invention improves the K-means algorithm to segment the frequency regulation demand scenarios of thermal power unit operating states, allowing data from different load and pressure ranges to be fed into corresponding models for prediction. This reduces the interference of mixed data from different operating conditions on the prediction results and improves the stability and generalization ability of the model under complex operating conditions. The combination of scenario segmentation and sub-model prediction enables a more accurate reflection of the frequency regulation response characteristics of thermal power units under different operating states. Based on the determination of frequency regulation contribution rate, this invention introduces a rolling optimization mechanism for main steam pressure, constructs a multi-objective optimization model with the corrected integral power as the core, and comprehensively considers operational safety constraints such as main steam pressure amplitude, pressure change rate, and turbine valve change rate to achieve a coordinated balance between frequency regulation capability improvement and safe unit operation. Compared with existing technologies that only perform monitoring or post-event analysis, this invention forms a closed-loop control mechanism of "prediction-determination-optimization-execution," enabling thermal power units to transform from passive response to active control. This invention employs an improved AGWO-CS hybrid optimization algorithm to perform rolling optimization of the main steam pressure setpoint. While ensuring search efficiency, it enhances the ability to escape local optima, improving the convergence speed and stability of the pressure optimization process. By setting up safety protection logic and interlocking mechanisms, it ensures that the optimization results are executed within the unit's permissible physical and control boundaries, avoiding equipment stress risks or operational instability caused by drastic pressure fluctuations.

[0121] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting the primary frequency regulation capability and optimizing pressure of a thermal power unit, characterized in that, Includes the following steps: S1: Within the first time window T1 before the frequency difference of the power grid crosses the preset dead zone, collect multi-source operating parameters of the thermal power unit and construct the pre-regulation state vector of the primary frequency regulation. S2: Based on the maximum frequency difference, average active power and average main steam pressure in the first frequency regulation pre-state vector, the current operating state of the thermal power unit is divided into frequency regulation demand scenarios; S3: For different frequency regulation demand scenarios, construct corresponding fusion models, extract predetermined demand features as model inputs within the second time window T2 before the frequency difference crosses the dead zone, and output the model with the predicted value of the actual integral power of the primary frequency regulation of the thermal power unit, so as to realize the online prediction of the primary frequency regulation capability. S4: Calculate the frequency regulation contribution rate based on the actual integrated power of the primary frequency regulation and the theoretical integrated power of the primary frequency regulation. When the frequency regulation contribution rate is lower than a preset threshold, it is determined as a frequency regulation failure event. For the frequency regulation failure event, the main steam pressure setpoint of the thermal power unit is optimized in a rolling manner. The optimization target is to ensure that the corrected actual integrated power of the primary frequency regulation, the main steam pressure change rate, and the valve change rate meet the safety constraints. In the rolling optimization process, the fusion model is called to evaluate the integrated power in real time. S5: The optimized main steam pressure setpoint is sent to the coordination control system of the thermal power unit to improve the primary frequency regulation qualification rate.

2. The method for predicting the primary frequency regulation capability and optimizing the pressure of thermal power units according to claim 1, characterized in that, The multi-source operating parameters include the unit's rotational speed, frequency, frequency difference signal, actual active power, main steam pressure setpoint, actual main steam pressure, main steam temperature, AGC command, turbine GV valve opening command, actual total coal consumption, and boiler main control output; the first time window T1 is within 10 seconds, and the sampling period of the first time window T1 is no more than 1 second.

3. The method for predicting the primary frequency regulation capability and optimizing the pressure of thermal power units according to claim 1, characterized in that, In step S2, an improved K-means algorithm is used to cluster the current frequency regulation operation status of the thermal power unit based on the multi-dimensional operating characteristic data consisting of the maximum frequency difference, average active power, and average main steam pressure in the pre-frequency regulation state vector. The improved K-means algorithm introduces a feedback correction term based on the clustering profile coefficient into the traditional K-means loss function to dynamically evaluate the clustering quality of the operating data of the thermal power unit under different loads and heat storage states. During the iteration process, when the clustering profile coefficient is lower than a set threshold, the initial cluster center is reselected based on the local density distribution of the operating samples of the thermal power unit to improve the distinguishability between different frequency regulation demand scenarios and reduce the impact of noisy operating samples on the boundary of the frequency regulation scenario.

4. The method for predicting the primary frequency regulation capability and optimizing the pressure of thermal power units according to claim 1, characterized in that, In step S3, the fusion model is a CNN-BiLSTM-Attention fusion model, which includes: At least two layers of one-dimensional convolutional neural networks are used to extract local dynamic features of the input sequence; A single-layer bidirectional long short-term memory network is used to capture the time-series dependency characteristics caused by the boiler thermal storage inertia of the thermal power unit; A single attention mechanism is used to weightedly fuse features from different time steps; and A fully connected output layer is used to output the predicted value of the actual integrated power of the first frequency modulation.

5. The method for predicting the primary frequency regulation capability and optimizing the pressure of thermal power units according to claim 1, characterized in that, In step S3, the second time window T2 is less than or equal to the first time window T1.

6. The method for predicting the primary frequency regulation capability and optimizing the pressure of thermal power units according to claim 1, characterized in that, In steps S3 and S4, the actual integral power of primary frequency regulation is the time integral value of the change in the actual active power of the unit during the period from when the frequency difference crosses the dead zone to when it returns to the dead zone; the theoretical integral power of primary frequency regulation is calculated based on the rated power of the thermal power unit, the droop coefficient, and the frequency difference function; and the frequency regulation contribution rate is the ratio of the actual integral power to the theoretical integral power.

7. The method for predicting the primary frequency regulation capability and optimizing the pressure of thermal power units according to claim 1, characterized in that, In step S4, the corrected primary frequency regulation actual integral power is a frequency regulation capability evaluation index obtained by introducing the main steam pressure amplitude deviation and main steam pressure change rate penalty factor of the thermal power unit on the basis of the original primary frequency regulation actual integral power; the rolling optimization of the main steam pressure setpoint of the thermal power unit adopts the improved AGWO-CS hybrid optimization algorithm, with the main steam pressure setpoint as the variable to be optimized and the corrected integral power as the fitness function, to search and optimize the operating parameter space of the thermal power unit under different loads and heat storage conditions; The improved AGWO-CS hybrid optimization algorithm includes: using chaotic mapping to initialize the feasible region of the main steam pressure setpoint to form a gray wolf population; Levy flight perturbation is introduced during the gray wolf individual position update process to enhance the global search capability in the pressure regulation space of the thermal power unit; the search step size in the optimization process is adjusted by nonlinear convergence factor to balance the requirements for rapid regulation of the main steam pressure of the thermal power unit and the requirements for stable operation of the thermal power unit; the fitness update mechanism of cuckoo search is used to replace the individual with the lowest fitness in the gray wolf population to improve the convergence efficiency of the optimization of the main steam pressure of the thermal power unit.

8. The method for predicting the primary frequency regulation capability and optimizing the pressure of thermal power units according to claim 7, characterized in that, In step S4, the optimization objective function of the thermal power unit is constructed according to the improved AGWO-CS hybrid optimization algorithm. The optimization objective function takes the corrected integral power as the main objective and introduces the deviation of the main steam pressure amplitude, the deviation of the main steam pressure change rate, and the deviation of the turbine GV valve change rate as penalty factors. A multi-objective weighted function is formed by setting weight coefficients.

9. The method for predicting the primary frequency regulation capability and optimizing the pressure of thermal power units according to any one of claims 1 to 8, characterized in that, In step S5, a safety protection logic is set before the optimized main steam pressure setting value is issued. When the deviation between the optimized main steam pressure setting value and the current actual pressure exceeds the preset upper limit or the pressure change rate exceeds the preset change rate threshold, the optimization result is automatically locked and the original setting value is kept unchanged.

10. A system for predicting the primary frequency regulation capability and optimizing the pressure of a thermal power unit, characterized in that, include: The data acquisition module is used to acquire the multi-source operating parameters of the thermal power unit and construct the preceding state vector; The scene segmentation module is used to execute the improved K-means algorithm to achieve the segmentation of frequency modulation requirement scenes; The prediction module has a built-in CNN-BiLSTM-Attention fusion model, which is used to output the predicted value of the actual integrated power of the first frequency modulation. The judgment module is used to calculate the frequency modulation contribution rate and determine whether a frequency modulation failure event has occurred. The optimization module is used to execute the improved AGWO-CS hybrid optimization algorithm and output the optimized main steam pressure setpoint. as well as The distribution module is used to write the optimized main steam pressure setpoint into the pressure controller of the coordinated control system (CCS) of the thermal power unit to achieve closed-loop regulation.

Images