Agent-driven power plant wide load cruise control method and system

Abstract

This invention discloses an agent-driven wide-load cruise control method and system for power plants, belonging to the field of intelligent control technology. Key technical points include: acquiring operating data of the power plant unit within a preset load range over a preset time period based on an agent; the operating data includes historical operating data and current operating data; obtaining the safe and stable value of the power plant unit at the current moment based on the current operating data and a preset random forest model; the preset random forest model is trained using historical operating data and boundary operating data, which is obtained from historical operating data and a variational autoencoder; and correcting PID parameters based on the deviation between the safe and stable value and the preset stable value. This invention trains the random forest using boundary operating data to improve the model's generalization ability and adaptability. Furthermore, by correcting the PID parameters based on the deviation between the safe and stable value and the preset stable value, this invention achieves adaptive adjustment of the unit control.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and more specifically to an agent-driven power plant wide-load cruise control method and system. Background Technology

[0002] In the modern power generation sector, the safe and stable operation of power plant units is of paramount importance, as it not only directly affects the reliability of power supply but also relates to energy utilization efficiency and enterprise operating costs.

[0003] While traditional PID control is widely used in power plant unit control, it has significant drawbacks. Its parameters are largely set based on experience, making it difficult to adaptively adjust and maintain optimal control performance in the face of frequent load changes and complex operating environments. Furthermore, PID parameters cannot be dynamically optimized based on the real-time operating status of the unit, making it difficult to guarantee the stability and control accuracy of the unit under different loads, leading to energy waste and increased equipment wear. To address these issues, existing technologies typically rely on the analysis of large amounts of industrial data to ensure the accuracy of PID parameters. For example, Chinese patent application CN119472553A discloses an intelligent optimization control method and platform for coal-fired power plants operating under varying loads. However, existing technologies lack sufficient mining and utilization of historical industrial data, preventing them from fully leveraging this data to achieve more precise parameter optimization. Therefore, existing technologies have shortcomings. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide an agent-driven power plant wide-load cruise control method and system. This method uses a random forest model to obtain the safe and stable value of the power plant unit under the current load condition, and automatically adjusts the PID parameters based on the deviation between the value and the preset stable value, so that the power plant unit is always in a safe and stable operating state.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides an agent-driven power plant wide-load cruise control method, comprising:

[0007] Based on the agent, the operation data of the power plant unit under a preset load range within a preset time period is obtained. The operation data includes historical operation data and current operation data.

[0008] Based on the current operating data and the preset random forest model, the safe and stable value of the power plant unit at the current moment is obtained. The preset random forest model is trained through the historical operating data and boundary operating data. The boundary operating data is obtained based on the historical operating data and the variational autoencoder. The safe and stable value is used to measure the operating status of the power plant unit under the current load.

[0009] The PID parameters are adjusted based on the deviation between the safe and stable value and the preset stable value.

[0010] As a further improvement of the present invention, the loss function in the variational autoencoder includes reconstruction loss and relative entropy loss, and the boundary running data is obtained based on the historical running data and the variational autoencoder, including:

[0011] Based on the distribution of the historical operating data, sample data located at the edge of the distribution are identified as boundary sample data;

[0012] Initialize the parameters of the variational autoencoder and define the loss function, which includes reconstruction loss and relative entropy loss;

[0013] The boundary sample data is augmented by the variational autoencoder to obtain the boundary running data.

[0014] As a further improvement of the present invention, the step of performing data augmentation processing on the boundary sample data through the variational autoencoder to obtain the boundary running data includes:

[0015] Perform an iterative operation, which includes randomly selecting multiple samples from the boundary sample data, inputting the multiple samples into the variational autoencoder to obtain reconstructed samples, calculating a reconstruction loss based on the reconstructed samples and the multiple samples, adjusting the parameters in the loss function based on the reconstruction loss, calculating the loss function based on the parameters, updating the parameters of the variational autoencoder based on the loss function, until a preset termination condition is met to obtain a converged variational autoencoder model;

[0016] Multiple latent vectors are randomly selected from the latent space of the converged variational autoencoder model;

[0017] The multiple potential vectors are input into the decoder in the converged variational autoencoder model to obtain the boundary running data.

[0018] As a further improvement of the present invention, the step of correcting the PID parameters based on the deviation between the safety stability value and the preset stability value includes:

[0019] The load value at the current moment is determined based on the current operating data;

[0020] Determine the corresponding preset stability value based on the load value;

[0021] The adjustment step size is determined based on the deviation between the safety stability value and the corresponding preset stability value;

[0022] The PID parameters are corrected based on the adjustment step size.

[0023] As a further improvement of the present invention, the step of determining the adjustment step size based on the deviation between the safety stability value and the corresponding preset stability value includes:

[0024] Obtain the deviation corresponding to the adjacent previous time step, and obtain the deviation change rate at the current time step based on the deviation corresponding to the adjacent previous time step.

[0025] Based on the deviation and the rate of change of the deviation, the overall deviation degree of the power plant unit at the current moment is calculated;

[0026] The adjustment step size is determined based on the overall deviation and the preset piecewise linear function.

[0027] As a further improvement of the present invention, the preset random forest model is trained using the historical running data and boundary running data, including:

[0028] The historical running data and the boundary running data are mixed and divided into training set and test set according to the ratio;

[0029] Based on the training set, perform random sampling with replacement to generate multiple sample sets;

[0030] For each sample set, a corresponding decision tree is trained to form the preset random forest model.

[0031] As a further improvement of the present invention, the step of training a corresponding decision tree for each sample set includes:

[0032] The features used to train the decision tree are determined based on thermodynamic principles;

[0033] Based on each sample in each sample set, a feature vector corresponding to each sample is calculated, and each element in the feature vector corresponds to a feature value;

[0034] Based on the eigenvalues, the Gini index or information gain for each eigenvalue is calculated.

[0035] Based on the Gini index or information gain of each feature, determine the features corresponding to the root node and internal nodes in each decision tree, as well as the threshold corresponding to each feature, to obtain the decision tree corresponding to each sample set.

[0036] As a further improvement of the present invention, the features include: net output energy to input energy ratio, waste heat utilization rate, entropy production rate, work capacity loss rate, and steam superheat change rate.

[0037] As a further improvement of the present invention, the step of obtaining the safety and stability value of the power plant unit at the current moment based on the current operating data and a preset random forest model includes:

[0038] Based on the current running data, the feature vector corresponding to the current running data is calculated;

[0039] The feature vectors are input into each decision tree to obtain the safe and stable value output by each decision tree;

[0040] The average of the safe and stable values ​​output by each decision tree is taken as the safe and stable value of the power plant unit at the current moment.

[0041] This invention provides an agent-driven power plant wide-load cruise control system, comprising:

[0042] The acquisition module is used to acquire the operating data of the power plant unit under a preset load range within a preset time period. The operating data includes historical operating data and current operating data.

[0043] The calculation module obtains the safety and stability value of the power plant unit at the current moment based on the current operating data and the preset random forest model. The preset random forest model is trained through the historical operating data and the boundary operating data. The boundary operating data is obtained based on the historical operating data and the variational autoencoder. The safety and stability value is used to measure the operating status of the power plant unit under the current load.

[0044] The correction module corrects the PID parameters based on the deviation between the safe and stable value and the preset stable value.

[0045] This invention trains a random forest using boundary operation data, enabling the model to better learn the operating characteristics of power plant units under different conditions, thereby improving the model's generalization ability and adaptability. Furthermore, this invention corrects PID parameters in real time based on the deviation between the safe and stable value and the preset stable value, achieving adaptive adjustment of the unit's control parameters, ensuring the unit's operational stability and control accuracy under different loads, and reducing energy waste and equipment wear. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the steps of an agent-driven power plant wide-load cruise control method according to the present invention.

[0047] Figure 2This is a flowchart of the data augmentation steps of the present invention;

[0048] Figure 3 This is a schematic diagram illustrating the training of the random forest model in this invention;

[0049] Figure 4 This is a schematic diagram illustrating the calculation of the current safe and stable value in this invention.

[0050] Figure 5 This is a schematic diagram of the overall control strategy of the power plant system in this invention. Detailed Implementation

[0051] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0052] Identical parts are indicated by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, while the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific part, respectively.

[0053] The term "and / or" in the following text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0054] Wide-load cruise control systems, based on an established unit optimization model and optimal cruise parameter settings, primarily address decision-making issues during unit operation. When load changes, they decide when to start or stop equipment, which equipment to start or stop, and determine optimal steam temperature, steam pressure, and rate of load change, thus operating the unit in a relatively safe and economical state on behalf of operators. In contrast, APS (Automatic Power Supply) focuses on unit start-up and shutdown; the procedures clearly define when to start or stop equipment, which equipment to start or stop, and the settings for various parameters during its operation. Its role is more about execution.

[0055] Wide-load cruise primarily involves in-depth optimization and improvement of the control configuration based on DCS control. Its logic design is transparent, facilitating maintenance and iteration by on-site personnel, and fulfilling all functions and performance indicators specified in the project's technical specifications, ensuring the safe and efficient operation of the unit. To further enhance the unit's automation level, while fully utilizing APS and deep peak-shaving control functions, the process systems that are still frequently operated by personnel during unit load changes are reviewed and optimized as a whole. Furthermore, the operations performed by operators under abnormal conditions are summarized, improved, and optimized, ultimately achieving less staffing, reducing the workload of operators, and lowering the risks of forced load reduction or even unplanned shutdowns caused by improper operation, thus achieving the goal of automatic cruise under unit load changes.

[0056] The ultimate goal of wide-load cruise is to reduce the workload of operators. In the actual operation of the unit, other auxiliary systems that require frequent operation by operators or parameters that require frequent adjustment can also be included in the wide-load cruise system to expand its functions.

[0057] like Figure 1 As shown in the figure, this application provides an agent-driven power plant wide-load cruise control method, including:

[0058] Based on the agent, the operation data of the power plant unit under the preset load range within the preset time period is obtained. The operation data includes historical operation data and current operation data.

[0059] Based on the current operating data and the preset random forest model, the safe and stable value of the power plant unit at the current moment is obtained. The preset random forest model is trained through historical operating data and boundary operating data. The boundary operating data is obtained based on historical operating data and variational autoencoder. The safe and stable value is used to measure the operating status of the power plant unit under the current load.

[0060] The PID parameters are adjusted based on the deviation between the safe and stable value and the preset stable value.

[0061] The preset time can be set according to actual needs, such as setting the preset time to within the past year or half a year; the operating data corresponding to each moment includes multiple parameters, such as the input energy, output energy and safe and stable value at the current moment; the PID parameters include proportional parameter (P), integral parameter (I) and derivative parameter (D). The PID controller can achieve precise regulation of the power plant unit control system by adjusting the PID parameters.

[0062] Preferably, the preset load range is 20%-100% load. For power plant units, load refers to the workload or output power undertaken by the unit during operation. For example, if a generator unit has a rated power of 1 million kilowatts but only outputs 200,000 kilowatts in actual production, then the generator unit is in a 20% load state. The load range refers to the range of load variation of the power plant unit during operation. Therefore, the operating data within the preset load range is the operating data corresponding to the power plant unit being in a 20%-100% load state.

[0063] This embodiment first generates boundary operation data using historical operation data, then trains the random forest using the boundary operation data to improve the model's generalization ability and adaptability. Based on the current operation data and the preset random forest model, it obtains the safe and stable value of the power plant unit at the current moment. Finally, it corrects the PID parameters in real time based on the deviation between the safe and stable value and the preset stable value, thus realizing precise adaptive adjustment of the unit's control parameters.

[0064] The following is combined with Figure 5 The application will be described in detail using examples from power plants. Specifically, the embodiments of this application can be used for automatic cruise control of pulverizing systems, main and reheat steam temperatures, feedwater systems, and flue gas systems in power plants, and can achieve fully automatic control of the entire unit under a load range of 20%-100%.

[0065] Unit coordinated (including main steam pressure) wide-load adaptive optimization control:

[0066] Based on the valve characteristics and load response capacity of the unit after wide-load optimization, priority is given to ensuring that the main steam pressure of the unit can be automatically tracked and optimized in real time during the unit load cruise process, so as to meet the unit's coordinated control requirements and the relative steady-state fluctuation range does not exceed ±0.2MPa.

[0067] Preferably, during low-load flexible dispatch operation of the unit, ensuring the main steam pressure of the unit remains within a reasonable range through adaptive control can be achieved by combining adaptive optimization control with adaptive recommendation for calorific value correction (dynamically adjusting the control strategy by real-time monitoring of fuel characteristics, system operating parameters, and environmental changes to achieve precise and efficient energy utilization). This ensures safe and stable control of the main steam pressure during low-load and flexible dispatch operation, guarantees stable steam temperature and pressure, and ensures sufficient extraction and stable exhaust temperature. Furthermore, the main steam pressure setting is adaptively adjusted in real time to maintain safe and economical unit operation.

[0068] It should be noted that the adaptive control method provided in this application can be freely combined with other scenarios and methods by those skilled in the art based on their own technical knowledge. As a limitation, no matter how the adaptive control method provided in this application is combined with other algorithms, it should fall within the scope of protection of this application.

[0069] Subsequently, a coordinated control approach was adopted, primarily based on indirect energy balance and secondarily on direct energy balance. Big data methods were used to reconstruct the unit's heat signal, ensuring stable unit operation while achieving rapid load response. Big data analysis and comparison with similar units were used to reconstruct the unit's sliding pressure curve, superheat function curve, water-coal ratio, and air-coal ratio. Based on the actual operating conditions of the unit, the operating curves were adaptively optimized and adjusted.

[0070] Then, the boiler main control feedforward loop can be improved, and the speed and amplitude limiting functions can be added to the boiler main control output commands. By limiting the rate of fuel input and feedwater flow, the drastic fluctuations in fuel input and feedwater flow can be prevented, which could cause local overheating of boiler metal and drastic fluctuations in steam temperature.

[0071] Furthermore, in this embodiment of the application, an optimized pressure pullback loop can be set in the turbine main control. When the main steam pressure deviation exceeds a certain value, the turbine main control assists the boiler in participating in pressure regulation. When the main steam pressure deviation continues to increase and exceeds the pressure limit action value, the turbine main control will open the regulating valve to assist the boiler in reducing the main steam pressure. This sacrifices the load control quality for a short time to stabilize the main steam pressure and ensure that the unit can operate safely and stably.

[0072] Of course, this application can also be used to adaptively adjust the calorific value correction rate of the unit. Taking the actual coal type in operation as an example, the calorific value correction method of the unit is constructed using data analysis methods, and the calorific value correction rate of the unit is adaptively adjusted to ensure rapid and accurate tracking and judgment of the heat entering the furnace when the coal quality changes significantly, so as to ensure the accuracy of the water-coal ratio and air-coal ratio, and to adaptively control the ratio of fuel, feedwater and air in steady state and dynamic process, thereby improving the dynamic adjustment quality of each major controlled parameter.

[0073] Wide-load adaptive optimization control of the pulverizing system:

[0074] The embodiments of this application can be further used in a pulverizing system. Specifically, based on the actual characteristics of the pulverizing system, adaptive control and adjustment of the primary air pressure based on the optimal coal powder concentration and the minimum primary air throttling loss can be achieved. This adaptive adjustment can apply the adaptive adjustment steps of this application, thereby making full use of online real-time coal quality data analysis, adaptively starting / stopping the pulverizing system, and reasonably controlling the operation mode of the pulverizing system to achieve safe, stable, and rapid response functions.

[0075] In this embodiment, the automatic warm-up control of the coal mill and the override control strategy for the cold and hot air dampers during the automatic start-up and shutdown of the mill can be optimized. During normal operation of the mill, based on the actual output changes of the mill, a combination of adaptive predictive control, fuzzy control, and servo control is adopted to achieve precise control of the outlet temperature, outlet air velocity, and coal powder fineness of the pulverizing system, optimize the dynamic process operating parameters of the pulverizing system throughout the entire process, and realize the automatic start-up and shutdown of the coal mill.

[0076] A linkage control strategy for primary air pressure and cold / hot air dampers of the grinding mill is constructed. It should be noted that, in this embodiment, the primary air pressure can be adaptively reduced based on the actual operating conditions of the grinding mill, specifically as follows: when the opening of all grinding mill cold / hot air dampers is too small, the primary air pressure is reduced to decrease primary air resistance throttling losses and primary air fan power consumption; when the opening of the grinding mill cold / hot air dampers is too large and loses its adjustment capability, the primary air pressure is correspondingly increased to improve the grinding mill's powder conveying capacity and outlet air temperature control capability; when the cold / hot air dampers are within a reasonable control range, if the hot air damper opening is too large and the cold air damper opening is too small, the grinding mill outlet air temperature setting is adaptively adjusted. Without increasing the risk of internal explosion in the coal mill and ensuring the dryness of the pulverized coal, the opening of the cold air damper should be increased to reduce its throttling loss. When the opening of the hot air damper is too small and the opening of the cold air damper is too large, without increasing the risk of internal explosion in the coal mill and ensuring the dryness of the pulverized coal, priority should be given to increasing the setting of the hot air volume and the output of the coal mill to reduce the hot air throttling loss and improve the economic efficiency of the coal mill operation. The control target of the primary air pressure needs to be considered not only with reference to the number of coal mills in operation, but also with real-time automatic optimization and adjustment based on the actual opening of the hot air damper as the ambient temperature changes. At the same time, intelligent intermittent correction should be performed based on the throttling level of the primary air at the coal mill outlet. On the one hand, this ensures that the throttling level of the hot air damper is within a reasonable range, achieving a balance between pulverized coal delivery capacity and energy saving. On the other hand, the intelligent intermittent correction method avoids oscillation of the primary air system during control correction.

[0077] During dynamic load changes, the primary air pressure setting can be adjusted in a timely manner according to the changes in fuel quantity. During load increases, the primary air pressure is increased; during load decreases, the primary air pressure is appropriately reduced to improve economy.

[0078] Wide-load adaptive optimization of water supply system:

[0079] This application further enables intelligent selection of feedwater pump operating mode and control strategy based on the actual operating conditions of the unit load; it achieves adaptive selection and adjustment of output across a wide load range. In the pre-control mode, when the load is below 50% (with a 2% dead zone), it automatically and seamlessly selects and shuts down one feedwater pump for standby operation; when the load increases, it automatically and seamlessly rejoins the pump, improving feedwater pump operating efficiency and reducing recirculation valve core scouring. Automatic pump shutdown and rejoining are achieved during unit start-up, shutdown, and normal operation.

[0080] In the design of the pump merging / retracting control strategy, merging is divided into three parts: increasing speed, opening the outlet valve, and output balancing. Retracting is divided into three parts: increasing output bias in the pump output balancing loop, closing the outlet valve, and reducing speed. An adaptive closed-loop control method for merging / retracting pump rates, a control method for balancing feedwater flow commands, and a dynamic adaptive proportional control strategy for the feedwater pump recirculation valve are adopted to reduce disturbances in feedwater flow during merging / retracting.

[0081] The boiler feedwater system is designed with 30% bypass feedwater piping and 100% main feedwater piping. Automatic switching of the feedwater piping is determined by feedwater flow rate, unit load, and the status of the feedwater bypass valve and feedwater pump. During feedwater bypass switching, the switching rate is adjusted by the feedwater flow rate control deviation to prevent large fluctuations in feedwater flow rate during the switching process, ultimately achieving reliable and smooth feedwater switching.

[0082] The main function of the feedwater pump recirculation valve is to ensure that the feedwater pump operates within a safe range and to assist in feedwater flow control. In terms of control strategy, the feedwater pump recirculation valve adopts a curve-shaped proportional control structure, determining the minimum flow rate of the feedwater pump based on the pump outlet pressure to ensure smooth feedwater flow control and system stability. During flexible unit operation, it adaptively adjusts the lower limit of feedwater flow protection and the lower limit of drum water level protection, while simultaneously shielding the high water level trip protection condition and implementing a high water level interconnection and drainage expansion strategy to ensure the safe and stable operation of the turbine.

[0083] It should be noted that the core concept of this application is to train the random forest using boundary running data, so that the model can better learn the operating characteristics of the power plant unit under different conditions, improve the model's generalization ability and adaptability, and the present invention corrects the PID parameters in real time according to the deviation between the safe stable value and the preset stable value, thereby realizing the adaptive adjustment of the unit control. There are no restrictions on the specific adaptive adjustment methods, such as the parameters and objects to be adjusted, and this application will not elaborate on them.

[0084] The inventive details of this application are described in detail below. This embodiment provides a step for obtaining boundary operation data based on historical operation data and a variational autoencoder, including:

[0085] Based on the distribution of historical operational data, sample data located at the edge of the distribution are identified as boundary sample data;

[0086] Boundary running data is obtained by performing data augmentation on boundary sample data using a variational autoencoder.

[0087] The loss function in the variational autoencoder includes reconstruction loss and relative entropy loss;

[0088] This embodiment considers that in the process of constructing a random forest, traditional sampling with replacement can lead to some samples being sampled multiple times while others are never sampled. Samples at the distribution boundaries, due to the randomness of sampling, cannot be fully included in the training set in multiple samplings. This results in a deviation between the constructed training set and the original historical data, failing to fully represent the distribution of the original historical data. Consequently, the random forest model's learning of these special samples is insufficient, affecting the prediction accuracy and generalization ability of the random forest model in these areas. Therefore, this embodiment uses a variational autoencoder to generate boundary data through data augmentation and mixes it with historical data to form a dataset containing more boundary sample information for subsequent training of the random forest model, thereby enhancing the model's accuracy.

[0089] Furthermore, such as Figure 2 As shown, this embodiment provides a method for performing data augmentation processing on boundary sample data using a variational autoencoder to obtain boundary running data, including:

[0090] Perform iterative operations, which include randomly selecting multiple samples from the boundary sample data, inputting the multiple samples into the variational autoencoder to obtain reconstructed samples, calculating the reconstruction loss based on the reconstructed samples and multiple samples, adjusting the parameters in the loss function based on the reconstruction loss, calculating the loss function based on the parameters, updating the parameters of the variational autoencoder based on the loss function, until the preset termination condition is met, and obtaining the converged variational autoencoder model.

[0091] Multiple latent vectors are randomly selected from the latent space of the converged variational autoencoder model;

[0092] Multiple latent vectors are input into the decoder in the converged variational autoencoder model to obtain boundary running data.

[0093] Specifically, the first step is to identify sample data located at the edge of the distribution based on the distribution of historical operational data. For example, data visualization techniques such as scatter plots or contour maps of historical operational data can be used to visually observe the data distribution and identify samples at the edge of the distribution; alternatively, density estimation can be performed on historical operational data, and data in areas with lower density can be used as boundary sample data.

[0094] Next, the encoder and decoder in the variational autoencoder are designed, and the parameters in the variational autoencoder are initialized, including the total number of iterations, the initial threshold L0 of the reconstruction loss, the threshold adjustment coefficient, etc. In this embodiment, both the encoder and decoder are designed using a multilayer perceptron architecture.

[0095] Then, an iterative operation is performed. First, multiple samples are randomly selected from the boundary sample data, and a sample vector corresponding to each sample is generated. Each sample vector is input into the encoder, and the mean and variance of the latent vector are output. Through the mean and variance, the distribution of the latent vector in the latent space can be obtained. Then, the latent vector is obtained through reparameterized sampling based on its distribution. Finally, the latent vector is decoded into reconstructed samples by the decoder. The latent vector is the feature representation of the sample vector in the latent space. The number of latent vectors and reconstructed samples are the same as the number of sample vectors.

[0096] Next, the reconstruction loss L is calculated based on the reconstructed samples and multiple sample vectors. r The reconstruction loss can be calculated based on the mean square error, specifically:

[0097]

[0098] Where n represents the number of sample vectors, x i This represents the i-th sample vector. This represents the i-th reconstructed sample.

[0099] The parameter β in the loss function is adjusted based on the reconstruction loss. Specifically, if the reconstruction loss is greater than the initial threshold L0, it indicates that the reconstruction effect of the current variational autoencoder is not ideal. Therefore, β is kept constant to avoid affecting the model's ability to learn and reconstruct data features. The reconstruction effect specifically refers to the similarity between the numerical and distributional features of the reconstructed samples and the sample vectors; the higher the similarity, the better the reconstruction effect. If the reconstruction loss is less than or equal to the initial threshold L0, it indicates that the reconstruction effect is good at this time, and therefore, the parameter β is adjusted. The focus is shifted towards relative entropy loss to achieve a dynamic balance between reconstruction loss and relative entropy loss, avoiding local optima. Simultaneously, the threshold for reconstruction loss is updated to L0 = L0 × α, where α is a threshold adjustment coefficient. By using a lower threshold to increase the requirements for reconstruction loss, the model continuously optimizes reconstruction quality during subsequent training. Preferably, L0 = 0.5 and α = 0.9.

[0100] Calculate the loss function L based on the adjusted β:

[0101] L = L r +β×L k l;

[0102]

[0103] Among them, L kl z represents the relative entropy loss. i Indicates through x i The resulting latent vector, q(z) i |x i) represents the distribution of the latent vectors, which is determined by the mean and variance of the encoder output, and q(z) represents the prior distribution of all latent vectors.

[0104] Finally, update the parameters in the variational autoencoder model, including the weight matrix and bias vector, according to the loss function. Repeat the above steps until the preset termination condition is met to obtain the converged variational autoencoder model. The preset termination condition can be that the number of iterations reaches the initially set total number of iterations or that the loss function is numerically converged.

[0105] This embodiment takes into account that in the early stages of training the variational autoencoder model, the model has a low degree of fit to the data and a large reconstruction loss. At this time, the model needs to focus on how to accurately reconstruct the input sample data in order to learn the basic features and distribution of the data. If the weight β is fixed and large at this time, it will cause the model to focus on the constraints of the latent vector distribution too early, while ignoring the fitting of the data itself, making it difficult for the model to converge to a better state. Conversely, in the later stages of training, when the reconstruction loss is small and the model has learned the features of the data well, if the weight β is still fixed and small, it will cause the model to overfit the training data and fail to make the latent vector distribution sufficiently close to the prior distribution, thereby affecting the model's generalization ability.

[0106] In summary, this embodiment adjusts the weights according to the magnitude of the reconstruction loss during model training, which can enable the model to better adapt to the characteristics of different training stages, avoid problems such as insufficient model training or overfitting caused by fixed weights, and optimize model performance.

[0107] Furthermore, such as Figure 3 As shown, this embodiment provides a step for training a preset random forest model using historical running data and boundary running data, including:

[0108] Historical operational data and boundary operational data are mixed and divided into training and test sets according to a certain ratio;

[0109] Based on the training set, perform random sampling with replacement to generate multiple sample sets, including sample set 1, sample set 2, ..., sample set N;

[0110] For each sample set, a corresponding decision tree is trained, including decision tree 1, decision tree 2, ..., decision tree N, which constitutes the preset random forest model.

[0111] Furthermore, this embodiment provides a step for training a corresponding decision tree for each sample set, including:

[0112] The features used to train the decision tree are determined based on thermodynamic principles;

[0113] For each sample in each sample set, the feature vector corresponding to each sample is calculated, and each element in the feature vector corresponds to a feature value;

[0114] Based on the eigenvalues, the Gini index or information gain for each eigenvalue is calculated.

[0115] Based on the Gini index or information gain of each feature, determine the features corresponding to the root node and internal nodes in each decision tree, as well as the threshold corresponding to each feature, to obtain the decision tree corresponding to each sample set.

[0116] Specifically, historical running data and boundary running data generated by the variational autoencoder model are first mixed to form a dataset for subsequent training of the random forest model. Then, the dataset is divided into training and test sets according to the ratio. The features and thresholds corresponding to each node in each decision tree are determined by the data in the training set, and the performance of the model is verified by the test set. Preferably, the ratio of training set to test set is 7:3.

[0117] Next, random sampling with replacement is performed on the training set to obtain multiple sample sets. Each sample set contains the same number of samples as the training set. For each sample in each sample set, a feature vector is calculated, and each element in the feature vector corresponds to a feature value.

[0118] Then, based on the Gini index or information gain of each feature, the features corresponding to the root node and internal nodes in each decision tree, as well as the threshold for each feature, are determined. Specifically, for each decision tree, when determining the features corresponding to each node, features not currently in its own decision tree are traversed. For these features, their information gain or Gini index is calculated, and the feature with the largest information gain or smallest Gini index is selected as the feature of the current node. Next, the values ​​of this feature in the sample set corresponding to the decision tree are sorted in ascending order, and each value is used as a threshold to divide the sample set into two parts. The information gain or Gini index under different partitions is calculated, and the value that maximizes the information gain or minimizes the Gini index is selected as the threshold for that feature. Specifically, if the feature is numerical, information gain is used for calculation; if it is discrete, the Gini index is used.

[0119] Finally, each decision tree is tested using a test set, and the decision trees that pass the test are retained. These decision trees that pass the test are then used to form a pre-defined decision tree model.

[0120] For example, characteristics determined according to thermodynamic principles may include: the ratio of net output energy to input energy, waste heat utilization rate, entropy production rate, work capacity loss rate, and steam superheat change rate.

[0121] Net output energy refers to the total electrical and mechanical energy output by the power plant unit under the current load, which can be calculated from the generator output power and turbine shaft power. Input energy refers to the chemical energy released by fuel combustion, which can be calculated from the fuel flow rate and the lower heating value of the fuel. This ratio reflects the energy conversion efficiency of the power plant unit. If the power plant unit is in a stable operating state, this ratio should be relatively stable. If the ratio fluctuates significantly, it indicates that there is abnormal energy loss inside the power plant unit, such as incomplete combustion or increased equipment friction, which will affect the safe and stable operation of the power plant unit.

[0122] Waste heat refers to the underutilized heat energy discharged during the operation of power plant units, such as waste heat from boiler tail flue gas and turbine exhaust steam. It can be obtained by measuring the ratio of heat recovered by waste heat recovery devices to the total waste heat. A decrease in waste heat utilization indicates increased waste heat, suggesting a malfunction in the waste heat recovery equipment, which can affect the safety of the power plant units. For example, low waste heat utilization from boiler tail flue gas leads to excessively high exhaust temperatures, exacerbating the risk of flue corrosion and threatening the long-term stable operation of the power plant units.

[0123] Entropy production rate reflects the degree of irreversibility during the operation of a power plant unit. According to the second law of thermodynamics, for a control volume, the entropy production rate can be obtained by calculating the entropy change of the working fluid at the inlet and outlet of the control volume, as well as the entropy flow caused by the heat exchanged with the surroundings. For example, suppose the entropy of the working fluid at the turbine inlet is s. in The entropy of the working fluid at the outlet is s out Let the working fluid mass flow rate be m, the heat exchanged between the turbine and the outside environment be Q, and the ambient temperature be T0. Then the entropy production rate S is:

[0124]

[0125] An increase in entropy production rate indicates an increase in irreversible losses, namely problems such as steam leakage and mechanical friction. This reduces the power plant unit's work capacity, leads to increased component wear, and affects the safe and stable operation of the power plant unit.

[0126] The power loss rate is calculated by taking the difference between the actual power output and the ideal power output as a percentage of the ideal power output. If the power loss rate increases, it indicates that the actual power output of the unit is far from the ideal state, which may be due to reasons such as equipment aging or scale buildup, which can easily affect the load-carrying capacity and stability of the power plant unit.

[0127] Steam superheat refers to the difference between steam temperature and the saturation temperature at the corresponding pressure. The rate of change of steam superheat is obtained by measuring the ratio of the change in steam superheat at different times under the current load to the time interval. An abnormal rate of change of steam superheat indicates a problem with boiler combustion or the steam system, which could threaten the safe and stable operation of the power plant unit.

[0128] This embodiment calculates the value of each feature based on the original operating data and trains the random forest model using the feature values. Compared with direct calculation using the original operating data, the new feature integrates the information contained in multiple original features, as well as the constraints and relationships provided by thermodynamic theory. It can reveal the physical processes and interactions inside the power plant system more deeply from the perspectives of energy, entropy, etc., so that the random forest model outputs more accurate safety and stability values.

[0129] Furthermore, such as Figure 4 As shown, this embodiment provides a step for obtaining the safe and stable value of a power plant unit at the current moment based on current operating data and a preset random forest model, including:

[0130] Based on the current running data, the feature vector corresponding to the current running data is calculated, and each element in the feature vector corresponds to a feature value;

[0131] The feature vectors are input into each decision tree to obtain the safe and stable value output by each decision tree, including safe and stable value 1, safe and stable value 2, ..., safe and stable value N;

[0132] The average of the safe and stable values ​​output by each decision tree is taken as the safe and stable value of the power plant unit at the current moment.

[0133] Specifically, the decision trees in the random forest constructed in this embodiment are all regression decision trees. For each decision tree, when the feature vector reaches the leaf node, the average of the safe and stable values ​​corresponding to all samples contained in the leaf node during training is used as the safe and stable value output by the decision tree.

[0134] Furthermore, this embodiment provides a step for correcting PID parameters based on the deviation between a safe stable value and a preset stable value, including:

[0135] Determine the load value at the current moment based on the current operating data;

[0136] Determine the corresponding preset stability value based on the load value;

[0137] The adjustment step size is determined based on the deviation between the safe and stable value and the corresponding preset stable value;

[0138] Adjust the PID parameters according to the step size.

[0139] Furthermore, this embodiment provides a step for determining the adjustment step size based on the deviation between the safety stability value and the corresponding preset stability value, including:

[0140] Obtain the deviation corresponding to the adjacent previous time step, and obtain the deviation change rate at the current time step based on the deviation corresponding to the adjacent previous time step.

[0141] Calculate the overall deviation degree of the power plant unit at the current moment based on the deviation and the deviation change rate;

[0142] Determine the adjustment step size according to the overall deviation degree and the preset piecewise linear function.

[0143] Specifically, when the power plant unit is under different loads, the operating characteristics of the unit will change, so the corresponding preset stable values are also different, and the preset stable values need to be determined according to the actual production needs.

[0144] Assume that the deviation at the current moment is e1, and the deviation corresponding to the adjacent previous moment is e0, and the deviation change rate The overall deviation degree P is:

[0145]

[0146] where ω is the weight coefficient, e max is the maximum deviation, Δe max is the maximum value of the deviation rate, e max and Δe max are both calculated from historical operation data.

[0147] Determine the adjustment step size ΔK according to the overall deviation degree as:

[0148]

[0149] where P1 and P2 are the segmentation points, k1, k2, and k3 represent the slopes of different intervals, and k1 < k2 < k3. The values of P1, P2, k1, k2, and k3 are all determined according to the actual production needs. Finally, correct the PID parameters at the adjacent previous moment based on the adjustment step size to obtain the corresponding PID parameters at the current moment.

[0150] In this embodiment, the step size is dynamically adjusted according to the magnitudes of the error and the error change rate. When the error or the error change rate is large, it indicates that the current safety stable value deviates far from the preset value or changes rapidly. At this time, a larger step size is required to quickly adjust the parameters; when the error and the error change rate are small, it indicates that the current safety stable value is close to the preset value and changes smoothly. At this time, a smaller step size should be adopted to improve the adjustment accuracy. In actual work, the error and the error change rate will change with the operating state. Compared with the method of fixed step size in the prior art, this embodiment can better adapt to different working conditions and environmental changes.

[0151] Furthermore, the embodiment of the present application provides an Agent-driven wide-load cruising control system for a power plant, including:

[0152] The acquisition module is used to acquire the operating data of the power plant units under the preset load range. The operating data includes historical operating data and current operating data.

[0153] The calculation module obtains the safe and stable value of the power plant unit at the current moment based on the current operating data and the preset random forest model. The preset random forest model is trained through historical operating data and boundary operating data. The boundary operating data is obtained based on historical operating data and variational autoencoder. The safe and stable value is used to measure the operating status of the power plant unit under the current load.

[0154] The correction module adjusts the PID parameters based on the deviation between the safe and stable value and the preset stable value.

[0155] Specifically, in this embodiment, data is collected by various sensors inside the power plant unit. The Agent is used to obtain the collected operating data. Based on the obtained data, the current safe and stable value and the corrected PID parameters are calculated. The calculated values ​​are then sent to the PID controller so that the PID controller can adjust according to the corrected PID parameters, making the corrected PID parameters more consistent with the operating state of the power plant unit under the current load.

[0156] This application provides an agent-driven power plant wide-load cruise control method and system, which, compared with the prior art, makes full use of industrial data (i.e., the operating data in this application embodiment). Specifically, it first generates boundary operating data through historical operating data, then trains a random forest using the boundary operating data to improve the model's generalization ability and adaptability, and obtains the safe and stable value of the power plant unit at the current moment based on the current operating data and the preset random forest model. Finally, it corrects the PID parameters in real time based on the deviation between the safe and stable value and the preset stable value, thereby realizing precise adaptive adjustment of the unit control parameters.

[0157] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0158] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0159] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0160] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An agent-driven power plant wide load cruise control method, characterized by, include: Based on the agent, the operation data of the power plant unit under a preset load range within a preset time period is obtained. The operation data includes historical operation data and current operation data. Based on the current operating data and the preset random forest model, the safe and stable value of the power plant unit at the current moment is obtained. The preset random forest model is trained through the historical operating data and boundary operating data. The boundary operating data is obtained based on the historical operating data and the variational autoencoder. The safe and stable value is used to measure the operating status of the power plant unit under the current load. The PID parameters are adjusted based on the deviation between the safety and stability value and the preset stability value. The loss function in the variational autoencoder includes reconstruction loss and relative entropy loss. The boundary running data is obtained based on the historical running data and the variational autoencoder, including: Based on the distribution of the historical operating data, sample data located at the edge of the distribution are identified as boundary sample data; The boundary sample data is augmented using the variational autoencoder to obtain the boundary running data. The step of performing data augmentation processing on the boundary sample data using the variational autoencoder to obtain the boundary running data includes: Perform an iterative operation, which includes randomly selecting multiple samples from the boundary sample data, inputting the multiple samples into the variational autoencoder to obtain reconstructed samples, calculating a reconstruction loss based on the reconstructed samples and the multiple samples, adjusting the parameters in the loss function based on the reconstruction loss, calculating the loss function based on the parameters, updating the parameters of the variational autoencoder based on the loss function, until a preset termination condition is met to obtain a converged variational autoencoder model; Multiple latent vectors are randomly selected from the latent space of the converged variational autoencoder model; The multiple potential vectors are input into the decoder in the converged variational autoencoder model to obtain the boundary running data.

2. An agent-driven power plant wide load cruise control method according to claim 1, characterized in that, The step of correcting the PID parameters based on the deviation between the safe and stable value and the preset stable value includes: The load value at the current moment is determined based on the current operating data; Determine the corresponding preset stability value based on the load value; The adjustment step size is determined based on the deviation between the safety stability value and the corresponding preset stability value; The PID parameters are corrected based on the adjustment step size.

3. An agent-driven power plant wide load cruise control method according to claim 2, characterized in that, The step of determining the adjustment step size based on the deviation between the safe and stable value and the corresponding preset stable value includes: Obtain the deviation corresponding to the adjacent previous time step, and obtain the deviation change rate at the current time step based on the deviation corresponding to the adjacent previous time step. Based on the deviation and the rate of change of the deviation, the overall deviation degree of the power plant unit at the current moment is calculated; The adjustment step size is determined based on the overall deviation and the preset piecewise linear function.

4. An agent-driven power plant wide load cruise control method according to claim 1, characterized in that, The preset random forest model is trained using the historical running data and boundary running data, including: The historical running data and the boundary running data are mixed and divided into training set and test set according to the ratio; Based on the training set, perform random sampling with replacement to generate multiple sample sets; For each sample set, a corresponding decision tree is trained to form the preset random forest model.

5. An agent-driven power plant wide load cruise control method according to claim 4, characterized in that, For each sample set, the corresponding decision tree is trained, including: The features used to train the decision tree are determined based on thermodynamic principles; Based on each sample in each sample set, a feature vector corresponding to each sample is calculated, and each element in the feature vector corresponds to a feature value; Based on the eigenvalues, the Gini index or information gain for each eigenvalue is calculated. Based on the Gini index or information gain of each feature, determine the features corresponding to the root node and internal nodes in each decision tree, as well as the threshold corresponding to each feature, to obtain the decision tree corresponding to each sample set.

6. An agent-driven power plant wide load cruise control method according to claim 5, characterized in that, The features include: net output energy to input energy ratio, waste heat utilization rate, entropy production rate, work capacity loss rate, and steam superheat change rate.

7. An agent-driven power plant wide load cruise control method according to claim 5, wherein, The step of obtaining the safety and stability value of the power plant unit at the current moment based on the current operating data and a preset random forest model includes: Based on the current running data, the feature vector corresponding to the current running data is calculated; The feature vectors are input into each decision tree to obtain the safe and stable value output by each decision tree; The average of the safe and stable values ​​output by each decision tree is taken as the safe and stable value of the power plant unit at the current moment.

8. An agent-driven power plant wide load cruise control system, characterized by, include: The acquisition module is used to acquire the operating data of the power plant unit under a preset load range within a preset time period. The operating data includes historical operating data and current operating data. The calculation module obtains the safety and stability value of the power plant unit at the current moment based on the current operating data and the preset random forest model. The preset random forest model is trained through the historical operating data and the boundary operating data. The boundary operating data is obtained based on the historical operating data and the variational autoencoder. The safety and stability value is used to measure the operating status of the power plant unit under the current load. The correction module corrects the PID parameters based on the deviation between the safety stability value and the preset stability value. The loss function in the variational autoencoder includes reconstruction loss and relative entropy loss. The boundary running data is obtained based on the historical running data and the variational autoencoder, including: Based on the distribution of the historical operating data, sample data located at the edge of the distribution are identified as boundary sample data; The boundary sample data is augmented using the variational autoencoder to obtain the boundary running data. The step of performing data augmentation processing on the boundary sample data using the variational autoencoder to obtain the boundary running data includes: Perform an iterative operation, which includes randomly selecting multiple samples from the boundary sample data, inputting the multiple samples into the variational autoencoder to obtain reconstructed samples, calculating a reconstruction loss based on the reconstructed samples and the multiple samples, adjusting the parameters in the loss function based on the reconstruction loss, calculating the loss function based on the parameters, updating the parameters of the variational autoencoder based on the loss function, until a preset termination condition is met to obtain a converged variational autoencoder model; Multiple latent vectors are randomly selected from the latent space of the converged variational autoencoder model; The multiple potential vectors are input into the decoder in the converged variational autoencoder model to obtain the boundary running data.

Images