Self-correcting and self-optimizing method for frequency modulation model of steam turbine of thermal power generating unit

Abstract

The invention provides a thermal power generating unit steam turbine frequency modulation model self-correction self-optimization-approaching method, which comprises the steps of constructing a data set, inputting the data set into a PyCT-PsLSTM prediction network model, outputting main steam pressure, valve opening and unit load and actual measurement value calculation errors, and if the errors exceed a threshold value, performing early warning and starting PID parameter optimization correction; screening non-tampered current power PID parameters from the data set, inputting the non-tampered current power PID parameters into the GPR model to obtain a predicted mean value, a standard deviation and a confidence interval, judging tampering if the parameters exceed the interval, calculating an uncertainty coefficient based on the mean value and the standard deviation, performing global search by using a particle swarm algorithm if the coefficient is greater than a threshold value, and performing local optimization if the coefficient is less than or equal to the threshold value; and obtaining an optimal solution, updating the data set, finely adjusting tampering parameters, comparing an actual value with GPR output, and judging to release or keep early warning. According to the method, the PID controller parameters and the confidence interval thereof are compared and predicted, early warning ending or warning keeping is judged, and the key parameter time sequence characteristic modeling capability is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent frequency regulation and automatic control of thermal power units, specifically to a self-correction and self-optimization method for the frequency regulation model of a steam turbine in a thermal power unit. Background Technology

[0002] In the frequency control of modern thermal power units, the dynamic characteristics of the turbine body and the parameter identification technology of various actuators are relatively mature, and current modeling and parameter acquisition methods can meet the needs of actual engineering. However, there is still considerable uncertainty in the PID parameter settings of the governor regulation system. Due to their own performance evaluation or economic optimization considerations, power plants often adjust the PID controller parameters without authorization. This human intervention may lead to a decrease in the frequency regulation system's response to grid disturbances, and even cause serious consequences such as low-frequency oscillations, system oscillations, and operational instability, posing a threat to power system safety.

[0003] Existing dispatching systems generally lack real-time monitoring and effective constraints on PID parameters, making it impossible to detect and correct abnormal operating conditions caused by parameter tampering in a timely manner. Traditional parameter tuning methods mostly rely on manual experience or offline experiments, which are difficult to adapt to the dynamic changes and complex disturbance environments of power plant operating conditions, and lack real-time early warning and adaptive correction mechanisms for parameter changes.

[0004] To address the aforementioned industry pain points, this invention collects real-time operational data such as unit frequency, main steam pressure, load, and speed, constructs and trains a data-driven model, and achieves online prediction and monitoring of the PID parameters of the governor regulation system. When the actual parameters deviate from the model prediction values ​​by more than a set threshold, the system can automatically trigger an alarm and correct the parameters, effectively preventing frequency regulation performance degradation and grid stability risks caused by human tampering, and improving the intelligence and automation level of power plant frequency regulation control. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a self-calibration and self-optimization method for frequency regulation models of steam turbines in thermal power units, aiming to solve the problems in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a self-calibration and self-optimization method for a frequency regulation model of a thermal power unit turbine, comprising: Step S1: Construct a standardized time-domain dataset; Step S2: Input the standardized time-domain dataset into the PyCT-PsLSTM prediction network model for processing, and calculate the difference between the output main steam pressure, valve opening and unit load and the actual measured values ​​at the current time. When the error of the difference calculation exceeds the preset error threshold, an early warning is triggered and the PID controller parameter prediction optimization correction mechanism is started. Step S3: In the prediction optimization and correction mechanism, the current power PID controller parameters that have not been tampered with are selected from the standardized time-domain dataset and input into the GPR model for processing. The predicted mean, predicted standard deviation and confidence interval are output. The current power PID controller parameters are compared with the confidence interval. If the current power PID controller parameters fall outside the confidence interval, it is determined that the current power PID controller parameters have been illegally tampered with, and a correction command is issued. Step S4: After issuing the correction command, first calculate the uncertainty coefficient based on the predicted mean and predicted standard deviation, compare the uncertainty coefficient with the preset coefficient threshold. If it is greater than the coefficient threshold, use the particle swarm optimization algorithm to perform a global search path for the predicted mean and predicted standard deviation. If it is less than or equal to the coefficient threshold, use the particle swarm optimization algorithm to perform local optimization exploration for the predicted mean and predicted standard deviation to obtain the current optimal solution. Then, update the standardized time domain dataset based on the current optimal solution, and fine-tune and correct the illegally tampered current power PID controller parameters based on the updated standardized time domain dataset. Step S5: By periodically comparing the actual measured values ​​with the predicted mean, predicted standard deviation, and confidence interval output by the GPR model, determine whether to lift the warning or keep the warning triggered.

[0007] Furthermore, the specific process of step S1 is as follows: The operating data of primary frequency regulation of thermal power units, including main steam pressure, is obtained based on sensors. Rotation speed Unit load Valve opening Current power PID controller parameters and target power value Key parameters; current power PID controller parameters include proportional gain. Integral gain Differential gain ; A time-domain dataset is constructed based on the operating data of primary frequency regulation of thermal power units; By normalizing, sliding window slicing, and cleaning the time-domain dataset, standardized time-series samples are obtained. The operating data of the thermal power unit under different time points and load conditions are collected. The operating data includes operating conditions covering high load, medium load, low load, frequency disturbance, and load step. Appropriate time tolerance windows are used for data points with different acquisition frequencies in the operating data of thermal power units. Processing is performed to establish a reasonable time tolerance window. Data points collected at different frequencies are considered valid data at the same time. All valid data timestamps are uniformly sorted and then distributed according to a set reasonable time tolerance window. The time stamp is iterated along the sorted timeline, and during the iteration, the timestamp is placed within the current unified and reasonable time tolerance window. Data points within the specified range are grouped together; within a reasonable time tolerance window During the grouping process, synchronously identify and remove timestamps that exhibit abnormal drift or exceed the reasonable time tolerance window. The discrete data at the threshold is used to obtain grouped data, and a linear interpolation function is applied to the grouped data to fill a reasonable time tolerance window. For missing values, Min-Max normalization is used to map the data after filling in the missing values ​​to the [0,1] interval, thus obtaining multi-source data; Wherein, the linear interpolation function and normalization represent: ; ; In the formula, This represents the estimated operating parameters of the thermal power unit at the missing positions within the time tolerance window after processing by the linear interpolation function; that is, the missing operating data values. and These represent the values ​​of two adjacent known data points in the operating data of a thermal power unit; and These represent the operating data of thermal power units. and timestamp; The original operating data of thermal power units to be normalized; This represents the minimum value of the operating data for thermal power units. This represents the maximum value of the operating data for the thermal power unit. This refers to the normalized multi-source data; The final standardized time-series samples and multi-source data form a standardized time-domain dataset containing load, main steam pressure, valve opening, and speed. ; This is the standardized time-domain dataset at time t; Let be the main steam pressure at time t; Let be the rotational speed at time t; Let be the unit load at time t; Let be the valve opening at time t; The proportional gain at time t; Let be the integral gain at time t; Let be the differential gain at time t; Let t be the target power value at time t.

[0008] Furthermore, the processing procedure of the PyCT-PsLSTM prediction network model is as follows: The PyCT-PsLSTM prediction network model includes a channel-independent unrolling module, a patching module, a feature extraction module, a channel-temporal collaborative attention module, and a long short-term memory network module. The standardized time-domain dataset is fed into the channel independence expansion module for processing, and the output of the channel independence expansion module is obtained. The output of the channel independence expansion module is input into the patching module. The patching module divides the time series in each channel of the output of the channel independence expansion module into fixed-length time slices (Patch). Each fixed-length time slice (Patch) is then mapped to a unified feature dimension through a linear embedding layer to capture the local dependency information and cross-time slice correlation in the time series of the feature dimension, thus obtaining the output of the patching module. The output of the patching module is fed into a dual-path stacked feature extraction module, which consists of two parallel sub-paths: the first parallel sub-path is a local path, composed of two cascaded one-dimensional convolution-max pooling units, which extracts short-term local features from the time series output of the patching module to obtain the output of the first sub-path; the second parallel sub-path is a global path, composed of downsampling-large convolutional kernel-upsampling units, which extracts global trend information over the time span from the output of the patching module to obtain the output of the second sub-path. The outputs of the first and second sub-paths are then fused element-wise to form a fused feature tensor. ; Channel-temporal collaborative attention module for fusion feature tensors Global average pooling is performed on both the time and spatial dimensions, followed by linear transformation and the sigmoid function to finally generate the channel attention weights. ; Then, on the fusion feature tensor Max pooling is performed on the channel dimension and spatial dimension to obtain time-weighted results. ,Will , and Element-wise multiplication is performed to obtain the attention-weighted feature output. , where ⊙ denotes element-wise multiplication; Will After being flattened, the data is input as an input vector to the Long Short-Term Memory (LSTM) network module for processing, yielding the main steam pressure, valve opening, and load as predicted values.

[0009] Furthermore, the specific process for calculating the difference between the output main steam pressure, valve opening, and unit load and the actual measured values ​​at the current moment in step S2 is as follows: When the absolute error between the predicted value and the actual measured value exceeds the set error threshold When the error threshold is not exceeded, the GPR model inference is triggered and an early warning is issued, and the PID controller parameter prediction, optimization, and correction process is initiated. It is safe to display at any time; The mean absolute error of the network, As a redundancy error threshold, the actual measured value is the main steam pressure in step S1. Unit load Valve opening .

[0010] Furthermore, the specific process of outputting the predicted mean, predicted standard deviation, and confidence interval in step S3 is as follows: When the absolute error between the predicted value and the actual measured value exceeds the set error threshold At that time, the current power PID controller parameters that have not been tampered with are selected from the standardized time domain dataset, and the real-time process variables and extended statistical features are extracted from the current power PID controller parameters that have not been tampered with as inputs for GPR model training; The real-time process variable is the valve opening at each sampling time. Main steam pressure p and unit load ; Extended statistical features include the mean moving average of unit load over the past 30-second window. ), main steam pressure standard deviation std(p) and valve opening change trend calculated by linear fitting. ); The standardized time-domain dataset is divided into training and testing datasets according to the time order. The training dataset is ,in This represents the i-th eigenvector. It is the true value of the current power PID controller parameter at the i-th corresponding time moment; The number of samples; Use radial base to capture valve opening. The nonlinear coupling between the main steam pressure p and the unit load is characterized by a linear kernel. With valve opening The steady-state correlation is used to introduce a periodic kernel to describe the unit load. The fluctuating diurnal pattern was obtained separately. ,at the same time Satisfy normalization constraints And adjust through adaptive optimization Specifically: ; ; ; ; ; In the formula, It is the covariance matrix; These are the weighting coefficients corresponding to the radial basis kernel; It is a radial basis kernel function; These are the weighting coefficients corresponding to the linear kernel; It is a linear kernel function; These are the weighting coefficients corresponding to the periodic kernel; It is a periodic kernel function; The variance of the observed noise; It is the identity matrix; The weight coefficients of the i-th kernel function; ; This is the transpose of the i-th weight vector; For feature vectors; This is the transpose of the j-th weight direction; Logarithmic probability; Logarithmic probability right The partial derivatives; Logarithmic probability right The partial derivatives; for right The partial derivatives; This is the weight vector at the (t+1)th iteration; Let be the weight vector at the t-th iteration; The gradient of the logarithmic probability with respect to the weight vector w; Let be the weight change in the t-th iteration; For learning rate, The momentum coefficient; Display the added noise variance term in the covariance matrix K. The complete covariance matrix is ​​defined as follows: ,in The training data matrix is ​​N×6 dimensional; Based on the Bayesian framework, the optimization hyperparameter set is set by estimating the maximum marginal likelihood function. The optimized hyperparameter set is used to further optimize the GPR model, resulting in a GPR model with the optimized hyperparameter set. The hyperparameter set is adjusted based on adaptive optimization. The complete covariance matrix is ​​constructed from the given information, where... The length scale of the radial base. For the periodic parameters of the periodic kernel, The smoothing coefficient of the periodic kernel. The amplitude coefficient of the periodic kernel. The standard deviation of noise. The standard deviation of the radial basis kernel; The optimized hyperparameter set's GPR model is used to calculate the predicted mean and observed noise variance of the current power PID controller parameters on the test dataset, which are expressed as follows: ; ; In the formula, To use the GPR model with the optimized hyperparameter set on the test dataset The predicted mean; k is the covariance function; For the test dataset, This represents the true values ​​of the current power PID controller parameters at the corresponding moment; The standard deviation of the predictions for the test dataset; Final output predicted mean Includes parameter values and confidence zone ; This is the predicted mean of the proportional term; This is the predicted mean of the integral term; This is the predicted mean of the differential term.

[0011] Furthermore, the specific process of obtaining the current optimal solution in step S4 is as follows: After issuing the correction command, it is first based on the predicted mean. and the predicted standard deviation Calculate the uncertainty coefficient. The coefficient threshold is 0.15, when When, execute the global search path; when At that time, local optimization exploration is carried out; when Perform a global search path; then in the global parameter space In the process, a certain number of particles are randomly generated. As corresponding candidate parameters; It is defined in the global parameter space One of the variables; The error between the predicted value and the actual measured value and the predicted mean The sum of errors with the candidate parameters constitutes the fitness function. A particle swarm optimization algorithm is used to find the optimal particle parameters until the iteration reaches the specified number of iterations or the fitness converges. The optimal position of the particle is then taken as the current optimal solution. , the formula is as follows: ; ; In the formula, For the first Fitness function for each particle; for Weighting coefficients; For the first The error between the predicted value and the actual measured value of each particle; for Weighting coefficients; For the first Each particle predicts the error between the current power PID controller parameters and the candidate parameters; For the first During the nth iteration The particle in the first The amount of change in position across each parameter dimension; Inertial weights; For the first During the nth iteration The particle in the first Speed ​​across all parameter dimensions; For individual learning factors; To generate a random number within the range [0,1], adjust... The randomness of learning; For the first The particle in the first The optimal position of a particle in each parameter dimension; For the first During the nth iteration The particle in the first Current position in each parameter dimension; As a social learning factor; To generate a random number within the range [0,1], adjust... The randomness of learning; For the entire particle swarm in the th The global optimal position in each parameter dimension; For the first During the nth iteration The particle in the first Current position in each parameter dimension; when Local optimization is explored; in the global parameter space... In this process, the particle swarm optimization algorithm is used to iteratively optimize the predicted value by considering the error between the predicted value and the actual measured value, and the predicted mean. The sum of errors with the candidate parameters forms the fitness function. The particle's velocity and position are continuously updated until the number of iterations is satisfied or the fitness converges. Finally, the particle's optimal position is taken as the current optimal solution. ; Among them, the current optimal solution Change ,express: .

[0012] Furthermore, the specific process of fine-tuning and correcting the illegally tampered current power PID controller parameters in step S4 is as follows: The set confidence bias threshold is 0.1, in the current optimal solution neighborhood Inside, the current optimal solution The feature vectors are stored together as new training samples and expanded into the standardized time-domain dataset. The hyperparameter set of illegally tampered current power PID controller parameters is fine-tuned and corrected using the maximum marginal likelihood function with the extended standardized time-domain dataset. Add a new training sample to the standardized time-domain dataset; A feature vector composed of real-time process variables and extended statistical features; Current optimal solution Change When the set confidence deviation threshold is exceeded The current optimal solution Change PID controller parameters marked as abnormal; For PID controller parameters that are out of order, the candidate parameters with the smallest error are selected from the global parameter space as the optimal control parameters. The feature vector and optimal control parameters Combine into sample groups, combine sample groups Add to standardized time-domain dataset In, that is The updated and standardized time-domain dataset is obtained; and based on the updated and standardized time-domain dataset, the hyperparameter set of the illegally tampered current power PID controller parameters is fine-tuned and corrected using the maximum marginal likelihood function to obtain the corrected current power PID controller parameters.

[0013] Furthermore, the specific process of step S5 is as follows: After executing the current power PID controller parameter correction command, the current power PID controller parameters marked as illegally tampered are continuously monitored and dynamically judged through the current power PID controller parameter optimization and correction mechanism. Specifically: Current optimal solution After deviating from the confidence interval, a continuous observation phase begins: when the absolute error between the predicted value and the actual measured value is less than a set error threshold... When the illegally modified current power PID controller parameters are adjusted, the warning is lifted. This indicates the warning is effective and the trigger warning is lifted when the absolute error between the predicted value and the actual measured value exceeds the set error threshold. If this occurs, proceed to the next cycle judgment process while maintaining the warning. Real-time operating data of primary frequency regulation of thermal power units is acquired by sensors. The GPR model is trained using the real-time operating data. Based on the trained GPR model, the corrected current power PID controller parameter operating data and real-time operating data are fitted and updated in real time. The uncertainty coefficient of the current time step is calculated based on the real-time fitting update. Uncertainty coefficient at the current time step This indicates that the judgment supports fine-tuning and correction of the current power PID controller parameters that have been illegally tampered with.

[0014] Compared with existing technologies, the present invention has the following advantages: (1) This invention proposes a feature extraction module based on the PyCT-PsLSTM prediction network model, which integrates local time structure blocks and local dependency information to improve the ability to model the time series features of key parameters such as main steam pressure, valve opening and unit load in thermal power unit operation data; it introduces the GPR model for nonlinear parameter mapping, which has the ability to output prediction variance, so that the GPR model can not only predict key operating parameters, but also provide confidence information, thereby enhancing interpretability and prediction credibility.

[0015] (2) This invention designs a dual-trigger self-correction mechanism based on the prediction error and uncertainty coefficient of the output prediction value of the PyCT-PsLSTM prediction network model. When the current power PID controller parameters deviate from the predicted value or confidence interval, it can identify and reject unreasonable parameters that may be tampered with. Based on the confidence interval of the PyCT-PsLSTM prediction network model, it explores the optimal unknown PID parameters. It uses the optimization ability of the particle swarm algorithm and the prediction fitting of the thermal power unit's operating state to find the optimal parameters and update the standardized time domain dataset and micro-batch training GPR model. This can improve the prediction accuracy of the PyCT-PsLSTM prediction network model for the thermal power unit's operating state, optimize the operating performance of the thermal power unit, reduce fuel consumption and pollutant emissions, and ensure the safe and stable operation of the unit. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] like Figure 1 As shown, the present invention provides a technical solution: a self-calibration and self-optimization method for a frequency regulation model of a thermal power unit turbine, comprising: Step S1: Construct a standardized time-domain dataset; Step S2: Input the standardized time-domain dataset into the PyCT-PsLSTM prediction network model for processing, and calculate the difference between the output main steam pressure, valve opening and unit load and the actual measured values ​​at the current time. When the error of the difference calculation exceeds the preset error threshold, an early warning is triggered and the PID controller parameter prediction optimization correction mechanism is started. Step S3: In the prediction optimization and correction mechanism, the current power PID controller parameters that have not been tampered with are selected from the standardized time-domain dataset and input into the GPR model for processing. The predicted mean, predicted standard deviation and confidence interval are output. The current power PID controller parameters are compared with the confidence interval. If the current power PID controller parameters fall outside the confidence interval, it is determined that the current power PID controller parameters have been illegally tampered with, and a correction command is issued. Step S4: After issuing the correction command, first calculate the uncertainty coefficient based on the predicted mean and predicted standard deviation, compare the uncertainty coefficient with the preset coefficient threshold. If it is greater than the coefficient threshold, use the particle swarm optimization algorithm to perform a global search path for the predicted mean and predicted standard deviation. If it is less than or equal to the coefficient threshold, use the particle swarm optimization algorithm to perform local optimization exploration for the predicted mean and predicted standard deviation to obtain the current optimal solution. Then, update the standardized time domain dataset based on the current optimal solution, and fine-tune and correct the illegally tampered current power PID controller parameters based on the updated standardized time domain dataset. Step S5: By periodically comparing the actual measured values ​​with the predicted mean, predicted standard deviation, and confidence interval output by the GPR model, determine whether to lift the warning or keep the warning triggered.

[0018] The specific process of step S1 is as follows: The operating data of primary frequency regulation of thermal power units, including main steam pressure, is obtained based on sensors. Rotation speed Unit load Valve opening Current power PID controller parameters and target power value Key parameters; current power PID controller parameters include proportional gain. Integral gain Differential gain ; A time-domain dataset is constructed based on the operating data of primary frequency regulation of thermal power units; By normalizing, sliding window slicing, and cleaning the time-domain dataset, standardized time-series samples are obtained. The operating data of the thermal power unit under different time points and load conditions are collected. The operating data includes operating conditions covering high load, medium load, low load, frequency disturbance, and load step. Appropriate time tolerance windows are used for data points with different acquisition frequencies in the operating data of thermal power units. Processing is performed to establish a reasonable time tolerance window. Data points collected at different frequencies are considered valid data at the same time. All valid data timestamps are uniformly sorted and then distributed according to a set reasonable time tolerance window. The time stamp is iterated along the sorted timeline, and during the iteration, the timestamp is placed within the current unified and reasonable time tolerance window. Data points within the specified range are grouped together; within a reasonable time tolerance window During the grouping process, synchronously identify and remove timestamps that exhibit abnormal drift or exceed the reasonable time tolerance window. The discrete data at the threshold is used to obtain grouped data, and a linear interpolation function is applied to the grouped data to fill a reasonable time tolerance window. For missing values, Min-Max normalization is used to map the data after filling in the missing values ​​to the [0,1] interval, thus obtaining multi-source data; Wherein, the linear interpolation function and normalization represent: (1); (2); In the formula, This represents the estimated operating parameters of the thermal power unit at the missing positions within the time tolerance window after processing by the linear interpolation function; that is, the missing operating data values. and These represent the values ​​of two adjacent known data points in the operating data of a thermal power unit; and These represent the operating data of thermal power units. and timestamp; The original operating data of thermal power units to be normalized; This represents the minimum value of the operating data for thermal power units. This represents the maximum value of the operating data for the thermal power unit. This refers to the normalized multi-source data; The final standardized time-series samples and multi-source data form a standardized time-domain dataset containing load, main steam pressure, valve opening, and speed. ; This is the standardized time-domain dataset at time t; Let be the main steam pressure at time t; Let be the rotational speed at time t; Let be the unit load at time t; Let be the valve opening at time t; The proportional gain at time t; Let be the integral gain at time t; Let be the differential gain at time t; Let t be the target power value at time t.

[0019] The PyCT-PsLSTM prediction network model processing procedure is as follows: The PyCT-PsLSTM prediction network model includes a channel-independent unrolling module, a patching module, a feature extraction module, a channel-temporal collaborative attention module, and a long short-term memory (LSTM) network module. The standardized time-domain dataset is fed into the channel independence expansion module for processing, and the output of the channel independence expansion module is obtained. The output of the channel independence expansion module is input into the patching module. The patching module divides the time series within each channel in the output of the channel independence expansion module into fixed-length time segments (P), for example, into N segments of length P, thereby obtaining a local time structure block of the form N×P. Each fixed-length time segment P is then mapped to a unified feature dimension through a linear embedding layer to capture the local dependency information and cross-time segment association in the time series of the feature dimension, thus obtaining the output of the patching module. The output of the patching module is fed into a dual-path stacked feature extraction module, which consists of two parallel sub-paths: the first parallel sub-path is a local path, composed of two cascaded one-dimensional convolution-max pooling units, used to extract short-term local features from the time series output of the patching module to obtain the output of the first sub-path; the second parallel sub-path is a global path, composed of downsampling-large convolutional kernel-upsampling units, used to extract global trend information over a longer time span from the output of the patching module to obtain the output of the second sub-path; the outputs of the first and second sub-paths are then fused element-wise to form a fused feature tensor. ; Channel-temporal collaborative attention module for fusion feature tensors A global average pooling GAP is applied to the time and space dimensions, followed by a linear transformation and the Sigmoid function to finally generate the channel attention weights. ; Then, on the fusion feature tensor Max pooling is performed on the channel dimension and spatial dimension to obtain time-weighted results. ,Will , and Element-wise multiplication is performed to obtain the attention-weighted feature output. , where ⊙ denotes element-wise multiplication; Will After flattening, the data is input as an input vector to the Long Short-Term Memory (LSTM) network module for processing. The main steam pressure, valve opening, and load are then used as predicted values, as follows: (3); (4); (5); (6); (7); (8); (9); In the formula, The output value of the output gate at time step t; It is the sigmoid activation function; For output gate; This is the input vector for the current time step t; This is the recursive weight matrix for the output gate; This represents the hidden state at time step t−1; This is the bias term for the output gate; To update the gate's output value at time step t; For activation functions; To update the weight matrix of the gate; To update the cyclic weight matrix of the gate; To update the gate's bias term; The output value of the input gate at time step t; It is an exponential function; This is the weight matrix of the input gate; The cyclic weight matrix of the input gate; This is the bias term for the input gate; The output value of the forget gate at time step t; This is the input weight matrix for the forget gate; Here is the recursive weight matrix for the forget gate; For the bias term of the forget gate; The pressure of the main steam at time t is the pressure at the current moment. The pressure of the main steam at the current time t−1; The valve opening at time step t; This represents the valve opening at time step t−1; The load at time step t.

[0020] The specific process for calculating the difference between the output main steam pressure, valve opening, and unit load and the actual measured values ​​at the current moment in step S2 is as follows: When the absolute error between the predicted value and the actual measured value exceeds the set error threshold When the predicted value is inconsistent with the actual measured value of the current power PID controller parameters, it triggers the GPR model inference and issues an early warning, and initiates the PID controller parameter prediction optimization and correction process; if the error threshold is not exceeded... It is safe to display at any time; The mean absolute error of the network, As a redundancy error threshold, the actual measured value is the main steam pressure in step S1. Unit load Valve opening .

[0021] The specific process of outputting the predicted mean, predicted standard deviation, and confidence interval in step S3 is as follows: When the absolute error between the predicted value and the actual measured value exceeds the set error threshold At that time, the current power PID controller parameters that have not been tampered with are selected from the standardized time domain dataset, and the real-time process variables and extended statistical features are extracted from the current power PID controller parameters that have not been tampered with as inputs for GPR model training; The real-time process variable is the valve opening at each sampling time. Main steam pressure p and unit load ; Extended statistical features include the mean moving average of unit load over the past 30-second window. ), main steam pressure standard deviation std(p) and valve opening change trend calculated by linear fitting. ); The standardized time-domain dataset was split in an 8:2 ratio according to time order, with the first 80% used as the training dataset and the remaining 20% ​​used as the test dataset. The training dataset is ,in This represents the i-th eigenvector. It is the true value of the current power PID controller parameter at the i-th corresponding time moment; The number of samples; Use radial base to capture valve opening. The nonlinear coupling between the main steam pressure p and the unit load is characterized by a linear kernel. With valve opening The steady-state correlation is used to introduce a periodic kernel to describe the unit load. The fluctuating diurnal pattern was obtained separately. ,at the same time Satisfy normalization constraints And adjust through adaptive optimization Specifically: (10); (11); (12); (13); (14); In the formula, It is the covariance matrix; These are the weighting coefficients corresponding to the radial basis kernel; It is a radial basis kernel function; These are the weighting coefficients corresponding to the linear kernel; It is a linear kernel function; These are the weighting coefficients corresponding to the periodic kernel; It is a periodic kernel function; The variance of the observed noise; It is the identity matrix; The weight coefficients of the i-th kernel function; ; This is the transpose of the i-th weight vector; For feature vectors; This is the transpose of the j-th weight direction; Logarithmic probability; Logarithmic probability right The partial derivatives; Logarithmic probability right The partial derivatives; for right The partial derivatives; This is the weight vector at the (t+1)th iteration; Let be the weight vector at the t-th iteration; The gradient of the logarithmic probability with respect to the weight vector w; Let be the weight change in the t-th iteration; For learning rate, The momentum coefficient; Display the added noise variance term in the covariance matrix K. To eliminate sensor acquisition interference; the complete covariance matrix is ​​defined as follows: ,in The training data matrix is ​​N×6 dimensional; Based on the Bayesian framework, the optimization hyperparameter set is set by estimating the maximum marginal likelihood function. The optimized hyperparameter set is used to optimize the GPR model, resulting in a GPR model with the optimized hyperparameter set. The hyperparameter set is constructed based on formulas (10)-(14) and the complete covariance matrix, where... The length scale of the radial base. For the periodic parameters of the periodic kernel, The smoothing coefficient of the periodic kernel. The amplitude coefficient of the periodic kernel. The standard deviation of noise. The standard deviation of the radial basis kernel; The optimized hyperparameter set's GPR model is used to calculate the predicted mean and observed noise variance of the PID controller parameters on the test dataset, which are expressed as follows: (15); (16); In the formula, To use the GPR model with the optimized hyperparameter set on the test dataset The predicted mean; It is the covariance function; For the test dataset, This represents the true values ​​of the current power PID controller parameters at the corresponding moment; The standard deviation of the predictions for the test dataset; Final output predicted mean Includes parameter values and 95% confidence zone ; This is the predicted mean of the proportional term; This is the predicted mean of the integral term; This is the predicted mean of the differential term.

[0022] The specific process of obtaining the current optimal solution in step S4 is as follows: After issuing the correction command, it is first based on the predicted mean. and the predicted standard deviation Calculate the uncertainty coefficient. The coefficient threshold is 0.15, when When this occurs, it indicates that the load condition is fluctuating drastically under the current operating state or that the GPR model has not yet fully learned the load condition under the current operating state, and a global search path is executed; when If the prediction is reliable, then local optimization will be explored. when Perform a global search path; then in the global parameter space In the process, a certain number of particles are randomly generated. As corresponding candidate parameters; It is defined in the global parameter space One of the variables represents the range of values ​​used to perform a global search under conditions of high uncertainty; The error between the predicted value and the actual measured value and the predicted mean The sum of errors with the candidate parameters constitutes the fitness function. A particle swarm optimization algorithm is used to find the optimal particle parameters until the iteration reaches the specified number of iterations or the fitness converges. The optimal position of the particle is then taken as the current optimal solution. , the formula is as follows: (17); (18); (19); In the formula, For the first Fitness function for each particle; for Weighting coefficients; For the first The error between the predicted value and the actual measured value of each particle; for Weighting coefficients; For the first Each particle predicts the error between the current power PID controller parameters and the candidate parameters; For the first During the nth iteration The particle in the first The amount of change in position across each parameter dimension; Inertial weights; For the first During the nth iteration The particle in the first Speed ​​across all parameter dimensions; For individual learning factors; To generate a random number within the range [0,1], adjust... The randomness of learning; For the first The particle in the first The optimal position of a particle in each parameter dimension; For the first During the nth iteration The particle in the first Current position in each parameter dimension; As a social learning factor; To generate a random number within the range [0,1], adjust... The randomness of learning; For the entire particle swarm in the th The global optimal position in each parameter dimension; For the first During the nth iteration The particle in the first Current position in each parameter dimension; when Local optimization is explored; in the global parameter space... In this process, the particle swarm optimization (PSO) algorithm is used for iterative optimization, taking the error between the predicted and actual measured values ​​and the predicted mean as the benchmark. The sum of errors with the candidate parameters forms the fitness function. The particle's velocity and position are continuously updated until the number of iterations is satisfied or the fitness converges. Finally, the particle's optimal position is taken as the current optimal solution. ; Among them, the current optimal solution Change ,express: (20).

[0023] The specific process of fine-tuning and correcting the illegally tampered current power PID controller parameters in step S4 is as follows: The set confidence bias threshold is 0.1, in the current optimal solution 10% of the neighborhood The current optimal solution The feature vectors are stored together as new training samples and expanded into the standardized time-domain dataset. The hyperparameter set of illegally tampered current power PID controller parameters is fine-tuned and corrected using the maximum marginal likelihood function with the extended standardized time-domain dataset. A new training sample has been added to the standardized time-domain dataset; A feature vector composed of real-time process variables and extended statistical features; Current optimal solution Change When the set confidence deviation threshold is exceeded The current optimal solution Change PID controller parameters marked as abnormal; For PID controller parameters that are out of order, the candidate parameters with the smallest error are selected from the global parameter space as the optimal control parameters. The feature vector and optimal control parameters Combine into sample groups, combine sample groups Add to standardized time-domain dataset In, that is The updated and standardized time-domain dataset is obtained; and based on the updated and standardized time-domain dataset, the hyperparameter set of the illegally tampered current power PID controller parameters is fine-tuned and corrected using the maximum marginal likelihood function to obtain the corrected current power PID controller parameters, thus completing the closed-loop self-evolution of the GPR model.

[0024] The specific process of step S5 is as follows: After executing the current power PID controller parameter correction command, the current power PID controller parameters marked as illegally tampered are continuously monitored and dynamically judged through the current power PID controller parameter optimization and correction mechanism. Specifically: Current optimal solution After deviating from the confidence interval, a continuous observation phase begins: when the absolute error between the predicted value and the actual measured value is less than a set error threshold... When the illegally modified current power PID controller parameters are adjusted, the warning is lifted. This indicates the warning is effective and the trigger warning is lifted when the absolute error between the predicted value and the actual measured value exceeds the set error threshold. If this occurs, proceed to the next cycle judgment process while maintaining the warning. In actual operation of thermal power units, phenomena such as slow scaling, blade erosion, and sensor drift can cause the input-output characteristics of the PyCT-PsLSTM prediction network model to deviate from the original training distribution. To address this, micro-batch augmentation training was introduced to train the GPR model, significantly improving its ability to track slow degradation factors such as scaling. Simultaneously, the corrected operating parameters of the current power PID controller were updated in real-time through fitting; specifically: Real-time operating data of primary frequency regulation of thermal power units is acquired by sensors. The GPR model is trained using the real-time operating data. Based on the trained GPR model, the corrected current power PID controller parameter operating data and real-time operating data are fitted and updated in real time. The uncertainty coefficient of the current time step is calculated based on the real-time fitting update. Uncertainty coefficient at the current time step This indicates that the judgment supports fine-tuning and correction of the current power PID controller parameters that have been illegally tampered with.

[0025] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A self-calibration and self-optimization method for a frequency regulation model of a thermal power unit turbine, characterized in that, include: Step S1: Construct a standardized time-domain dataset; Step S2: Input the standardized time-domain dataset into the PyCT-PsLSTM prediction network model for processing, and calculate the difference between the output main steam pressure, valve opening and unit load and the actual measured values ​​at the current time. When the error of the difference calculation exceeds the preset error threshold, an early warning is triggered and the PID controller parameter prediction optimization correction mechanism is started. Step S3: In the prediction optimization and correction mechanism, the current power PID controller parameters that have not been tampered with are selected from the standardized time-domain dataset and input into the GPR model for processing. The predicted mean, predicted standard deviation and confidence interval are output. The current power PID controller parameters are compared with the confidence interval. If the current power PID controller parameters fall outside the confidence interval, it is determined that the current power PID controller parameters have been illegally tampered with, and a correction command is issued. Step S4: After issuing the correction command, first calculate the uncertainty coefficient based on the predicted mean and predicted standard deviation, compare the uncertainty coefficient with the preset coefficient threshold. If it is greater than the coefficient threshold, use the particle swarm optimization algorithm to perform a global search path for the predicted mean and predicted standard deviation. If it is less than or equal to the coefficient threshold, use the particle swarm optimization algorithm to perform local optimization exploration for the predicted mean and predicted standard deviation to obtain the current optimal solution. Then, update the standardized time domain dataset based on the current optimal solution, and fine-tune and correct the illegally tampered current power PID controller parameters based on the updated standardized time domain dataset. Step S5: By periodically comparing the actual measured values ​​with the predicted mean, predicted standard deviation, and confidence interval output by the GPR model, determine whether to lift the warning or keep the warning triggered.

2. The self-calibration and self-optimization method for frequency regulation model of steam turbine in thermal power unit according to claim 1, characterized in that: The specific process of step S1 is as follows: The operating data of primary frequency regulation of thermal power units, including main steam pressure, is obtained based on sensors. Rotation speed Unit load Valve opening Current power PID controller parameters and target power value Key parameters; current power PID controller parameters include proportional gain. Integral gain Differential gain ; A time-domain dataset is constructed based on the operating data of primary frequency regulation of thermal power units; By normalizing, sliding window slicing, and cleaning the time-domain dataset, standardized time-series samples are obtained. The operating data of the thermal power unit under different time points and load conditions are collected. The operating data includes operating conditions covering high load, medium load, low load, frequency disturbance, and load step. Appropriate time tolerance windows are used for data points with different acquisition frequencies in the operating data of thermal power units. Processing is performed to establish a reasonable time tolerance window. Data points collected at different frequencies are considered valid data at the same time. All valid data timestamps are uniformly sorted and then distributed according to a set reasonable time tolerance window. The time stamp is iterated along the sorted timeline, and during the iteration, the timestamp is placed within the current unified and reasonable time tolerance window. Data points within the specified range are grouped together; within a reasonable time tolerance window During the grouping process, synchronously identify and remove timestamps that exhibit abnormal drift or exceed the reasonable time tolerance window. The discrete data at the threshold is used to obtain grouped data, and a linear interpolation function is applied to the grouped data to fill a reasonable time tolerance window. For missing values, Min-Max normalization is used to map the data after filling in the missing values ​​to the [0,1] interval, thus obtaining multi-source data; Wherein, the linear interpolation function and normalization represent: ； ； In the formula, This represents the estimated operating parameters of the thermal power unit at the missing positions within the time tolerance window after processing by the linear interpolation function; that is, the missing operating data values. and These represent the values ​​of two adjacent known data points in the operating data of a thermal power unit; and These represent the operating data of thermal power units. and timestamp; The original operating data of thermal power units to be normalized; This represents the minimum value of the operating data for thermal power units. This represents the maximum value of the operating data for the thermal power unit. This refers to the normalized multi-source data; The final standardized time-series samples and multi-source data form a standardized time-domain dataset containing load, main steam pressure, valve opening, and speed. ; This is the standardized time-domain dataset at time t; Let be the main steam pressure at time t; Let be the rotational speed at time t; Let be the unit load at time t; Let be the valve opening at time t; The proportional gain at time t; Let be the integral gain at time t; Let be the differential gain at time t; Let t be the target power value at time t.

3. The self-calibration and self-optimization method for frequency regulation model of steam turbine in thermal power unit according to claim 2, characterized in that: The PyCT-PsLSTM prediction network model processing procedure is as follows: The PyCT-PsLSTM prediction network model includes a channel-independent unrolling module, a patching module, a feature extraction module, a channel-temporal collaborative attention module, and a long short-term memory network module. The standardized time-domain dataset is fed into the channel independence expansion module for processing, and the output of the channel independence expansion module is obtained. The output of the channel independence expansion module is input into the patching module. The patching module divides the time series in each channel of the output of the channel independence expansion module into fixed-length time slices (Patch). Each fixed-length time slice (Patch) is then mapped to a unified feature dimension through a linear embedding layer to capture the local dependency information and cross-time slice correlation in the time series of the feature dimension, thus obtaining the output of the patching module. The output of the patching module is fed into a dual-path stacked feature extraction module, which consists of two parallel sub-paths: the first parallel sub-path is a local path, composed of two cascaded one-dimensional convolution-max pooling units, which extracts short-term local features from the time series output of the patching module to obtain the output of the first sub-path; the second parallel sub-path is a global path, composed of downsampling-large convolutional kernel-upsampling units, which extracts global trend information over the time span from the output of the patching module to obtain the output of the second sub-path. The outputs of the first and second sub-paths are then fused element-wise to form a fused feature tensor. ; Channel-temporal collaborative attention module for fusion feature tensors Global average pooling is performed on both the time and spatial dimensions, followed by linear transformation and the sigmoid function to finally generate the channel attention weights. ; Then, on the fusion feature tensor Max pooling is performed on the channel dimension and spatial dimension to obtain time-weighted results. ,Will , and Element-wise multiplication is performed to obtain the attention-weighted feature output. , where ⊙ denotes element-wise multiplication; Will After being flattened, the data is input as an input vector to the Long Short-Term Memory (LSTM) network module for processing, yielding the main steam pressure, valve opening, and load as predicted values.

4. The self-calibration and self-optimization method for frequency regulation model of steam turbine in thermal power unit according to claim 3, characterized in that: The specific process for calculating the difference between the output main steam pressure, valve opening, and unit load and the actual measured values ​​at the current moment in step S2 is as follows: When the absolute error between the predicted value and the actual measured value exceeds the set error threshold When this happens, the GPR model inference is triggered and an early warning is issued, and the PID controller parameter prediction, optimization and correction process is initiated. Not exceeding the error threshold It is safe to display at any time; The mean absolute error of the network, As a redundancy error threshold, the actual measured value is the main steam pressure in step S1. Unit load Valve opening .

5. The self-calibration and self-optimization method for frequency regulation model of steam turbine in thermal power unit according to claim 4, characterized in that: The specific process of outputting the predicted mean, predicted standard deviation, and confidence interval in step S3 is as follows: When the absolute error between the predicted value and the actual measured value exceeds the set error threshold At that time, the current power PID controller parameters that have not been tampered with are selected from the standardized time domain dataset, and the real-time process variables and extended statistical features are extracted from the current power PID controller parameters that have not been tampered with as inputs for GPR model training; The real-time process variable is the valve opening at each sampling time. Main steam pressure p and unit load ; Extended statistical features include the mean of unit load (mean). ), main steam pressure standard deviation std(p) and valve opening change trend calculated by linear fitting. ); The standardized time-domain dataset is divided into training and testing datasets according to the time order. The training dataset is ,in This represents the i-th eigenvector. It is the true value of the current power PID controller parameter at the i-th corresponding time moment; The number of samples; Use radial base to capture valve opening. The nonlinear coupling between the main steam pressure p and the unit load is characterized by a linear kernel. With valve opening The steady-state correlation is used to introduce a periodic kernel to describe the unit load. The fluctuating diurnal pattern was obtained separately. ,at the same time Satisfy normalization constraints And adjust through adaptive optimization Specifically: ； ； ； ； ； In the formula, It is the covariance matrix; These are the weighting coefficients corresponding to the radial basis kernel; It is a radial basis kernel function; These are the weighting coefficients corresponding to the linear kernel; It is a linear kernel function; These are the weighting coefficients corresponding to the periodic kernel; It is a periodic kernel function; The variance of the observed noise; It is the identity matrix; The weight coefficients of the i-th kernel function; ; This is the transpose of the i-th weight vector; For feature vectors; This is the transpose of the j-th weight direction; Logarithmic probability; Logarithmic probability right The partial derivatives; Logarithmic probability right The partial derivatives; for right The partial derivatives; This is the weight vector at the (t+1)th iteration; Let be the weight vector at the t-th iteration; The gradient of the logarithmic probability with respect to the weight vector w; Let be the weight change in the t-th iteration; For learning rate, The momentum coefficient; Display the added noise variance term in the covariance matrix K. The complete covariance matrix is ​​defined as follows: ,in The training data matrix is ​​N×6 dimensional; Based on the Bayesian framework, the optimization hyperparameter set is set by estimating the maximum marginal likelihood function. The optimized hyperparameter set is used to further optimize the GPR model, resulting in a GPR model with the optimized hyperparameter set. The hyperparameter set is adjusted based on adaptive optimization. The complete covariance matrix is ​​constructed from the given information, where... The length scale of the radial base. For the periodic parameters of the periodic kernel, The smoothing coefficient of the periodic kernel. The amplitude coefficient of the periodic kernel. The standard deviation of noise. The standard deviation of the radial basis kernel; The optimized hyperparameter set's GPR model is used to calculate the predicted mean and observed noise variance of the current power PID controller parameters on the test dataset, which are expressed as follows: ； ； In the formula, To use the GPR model with the optimized hyperparameter set on the test dataset The predicted mean; k is the covariance function; For the test dataset, This represents the true values ​​of the current power PID controller parameters at the corresponding moment; The standard deviation of the predictions for the test dataset; Final output predicted mean Includes parameter values and confidence zone , This is the predicted mean of the proportional term; This is the predicted mean of the integral term; This is the predicted mean of the differential term.

6. The self-calibration and self-optimization method for frequency regulation model of steam turbine in thermal power unit according to claim 5, characterized in that: The specific process of obtaining the current optimal solution in step S4 is as follows: After issuing the correction command, it is first based on the predicted mean. and the predicted standard deviation Calculate the uncertainty coefficient. The coefficient threshold is 0.15, when When, execute the global search path; when At that time, local optimization exploration is carried out; when Perform a global search path; then in the global parameter space In the process, a certain number of particles are randomly generated. As corresponding candidate parameters; It is defined in the global parameter space One of the variables; The error between the predicted value and the actual measured value and the predicted mean The sum of errors with the candidate parameters constitutes the fitness function. A particle swarm optimization algorithm is used to find the optimal particle parameters until the iteration reaches the specified number of iterations or the fitness converges. The optimal position of the particle is then taken as the current optimal solution. , the formula is as follows: ； ； In the formula, For the first Fitness function for each particle; for Weighting coefficients; For the first The error between the predicted value and the actual measured value of each particle; for Weighting coefficients; For the first Each particle predicts the error between the current power PID controller parameters and the candidate parameters; For the first During the nth iteration The particle in the first The amount of change in position across each parameter dimension; Inertial weights; For the first During the nth iteration The particle in the first Speed ​​across all parameter dimensions; For individual learning factors; To generate a random number within the range [0,1], adjust... The randomness of learning; For the first The particle in the first The optimal position of a particle in each parameter dimension; For the first During the nth iteration The particle in the first Current position in each parameter dimension; As a social learning factor; To generate a random number within the range [0,1], adjust... The randomness of learning; For the entire particle swarm in the th The global optimal position in each parameter dimension; For the first During the nth iteration The particle in the first Current position in each parameter dimension; when Local optimization is explored; in the global parameter space... In this process, the particle swarm optimization algorithm is used to iteratively optimize the predicted value by considering the error between the predicted value and the actual measured value, and the predicted mean. The sum of errors with the candidate parameters forms the fitness function. The particle's velocity and position are continuously updated until the number of iterations is satisfied or the fitness converges. Finally, the particle's optimal position is taken as the current optimal solution. ; Among them, the current optimal solution Change ,express: 。 7. The self-calibration and self-optimization method for frequency regulation model of steam turbine in thermal power unit according to claim 6, characterized in that: The specific process of fine-tuning and correcting the illegally tampered current power PID controller parameters in step S4 is as follows: The set confidence bias threshold is 0.1, in the current optimal solution neighborhood Inside, the current optimal solution The feature vectors are stored together as new training samples and expanded into the standardized time-domain dataset. The hyperparameter set of illegally tampered current power PID controller parameters is fine-tuned and corrected using the maximum marginal likelihood function with the extended standardized time-domain dataset. Add a new training sample to the standardized time-domain dataset; A feature vector composed of real-time process variables and extended statistical features; Current optimal solution Change When the set confidence deviation threshold is exceeded The current optimal solution Change PID controller parameters marked as abnormal; For PID controller parameters that are out of order, the candidate parameters with the smallest error are selected from the global parameter space as the optimal control parameters. The feature vector and optimal control parameters Combine into sample groups, combine sample groups Add to standardized time-domain dataset In, that is The updated and standardized time-domain dataset is obtained; and based on the updated and standardized time-domain dataset, the hyperparameter set of the illegally tampered current power PID controller parameters is fine-tuned and corrected using the maximum marginal likelihood function to obtain the corrected current power PID controller parameters.

8. The self-calibration and self-optimization method for frequency regulation model of steam turbine in thermal power unit according to claim 7, characterized in that: The specific process of step S5 is as follows: After executing the current power PID controller parameter correction command, the current power PID controller parameters marked as illegally tampered are continuously monitored and dynamically judged through the current power PID controller parameter optimization and correction mechanism. Specifically: Current optimal solution After deviating from the confidence interval, a continuous observation phase begins: when the absolute error between the predicted value and the actual measured value is less than a set error threshold... When the illegally modified current power PID controller parameters are adjusted, the warning is lifted. This indicates the warning is effective and the trigger warning is lifted when the absolute error between the predicted value and the actual measured value exceeds the set error threshold. If this occurs, proceed to the next cycle judgment process while maintaining the warning. Real-time operating data of primary frequency regulation of thermal power units is acquired by sensors. The GPR model is trained using the real-time operating data. Based on the trained GPR model, the corrected current power PID controller parameter operating data and real-time operating data are fitted and updated in real time. The uncertainty coefficient of the current time step is calculated based on the real-time fitting update. Uncertainty coefficient at the current time step This indicates that the judgment supports fine-tuning and correction of the current power PID controller parameters that have been illegally tampered with.

Images