Quality monitoring method for thermotechnical control loop of thermal power plant

Abstract

The invention provides a thermal control loop quality monitoring method for a thermal power plant, and aims to improve the thermal control loop performance of the thermal power plant. According to the method, firstly, relevant operation parameters of a PID control loop are collected online in real time and normalized, then performance is evaluated by using a minimum variance or a generalized minimum variance, and actual, ideal and historical optimal indexes are compared. And aiming at the condition that the performance is influenced by unreasonable PID parameters, data-driven modeling and an advanced control technology are utilized to optimize the parameters. And setting a threshold range, and respectively triggering alarm and carrying out trend prediction and early warning on the performance index deviation and long-term performance degradation problems. A control strategy is adjusted according to evaluation and detection information, the effect is continuously monitored, key factors are found in combination with multi-dimensional data fusion and big data analysis, and a rule is excavated. Performance reports are generated regularly, key indexes are displayed, iterative optimization is carried out based on results of all links, it is guaranteed that the loop is in the optimal state, and the reliability and efficiency of thermal power plant operation are improved.

Description

Technical Field

[0001] The invention relates to the field of control loop quality monitoring, and in particular to a thermal control loop quality monitoring method for a thermal power plant. Background Art

[0002] In the thermal control loop of a thermal power plant, its operating quality is crucial to the safe, stable and efficient operation of the power plant. With the expansion of the scale of thermal power plants and the increasing complexity of the operating environment, thermal control loops face many challenges. Traditional monitoring methods often find it difficult to comprehensively and accurately evaluate the performance of the control loop. For example, simple performance indicator calculations may not take into account the combined effects of multiple factors such as process, equipment, and load, resulting in inaccurate judgments on the state of the control loop. Moreover, when faced with a large amount of historical data, there is a lack of effective means to mine the hidden patterns and laws, making it difficult to warn of possible failures or performance degradation in advance. In addition, when the performance of the control loop is abnormal, the control strategy cannot be adjusted in a timely and automatic manner to adapt to changes in operating conditions, and complex manual analysis and intervention are required, which is not only inefficient, but may also cause delays or aggravation of the problem due to human errors.

[0003] Therefore, there is an urgent need for a more advanced, comprehensive and intelligent thermal control loop quality monitoring method to improve the overall performance and reliability of thermal control systems in thermal power plants, ensure the stable operation of power plants and adapt to the ever-changing technical requirements. Summary of the invention

[0004] The main purpose of the present invention is to provide a method for monitoring the quality of a thermal control loop in a thermal power plant, so as to solve the problem that when the performance of the control loop is abnormal, the control strategy cannot be adjusted in time and automatically to adapt to the change in operating conditions, and complex manual analysis and intervention are required, which is not only inefficient but may also cause delays or aggravation of the problem due to human errors.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for monitoring the quality of a thermal control loop of a thermal power plant, the method comprising: S1. Collect all operating parameters related to the PID control loop online in real time, including input signal, output signal, set value, actual measurement value, and normalize the collected data; S2. Analyze the performance of the existing PID control loop using the minimum variance or generalized minimum variance performance evaluation method; Calculate the actual performance of the current PID control loop and compare it with the ideal or historical best performance; S3. For control loops whose control performance is affected by unreasonable PID parameter settings, these parameters are re-optimized through data-driven modeling methods; Use adaptive control and predictive control advanced control technology to adjust the proportional, integral and differential coefficients of the PID controller according to real-time data to improve control performance; S4. Set a reasonable threshold range. When the performance index of the PID control loop deviates from the normal range, the alarm mechanism is triggered; For performance degradation issues during long-term changes, a trend prediction model is established to provide early warning of possible failures or performance degradation; S5. Automatically or manually adjust the control strategy based on the performance evaluation results and abnormal detection information to ensure that the control loop is always in the optimal working state; Continuously monitor the effects of adjustments and record them for subsequent analysis and further optimization; S6. Combine the influence of process, equipment and load factors to conduct multi-dimensional data fusion analysis to find out the key factors affecting the performance of the control loop; Use big data analysis technology to discover patterns and rules hidden in large amounts of historical data, providing reference for future control optimization; S7. Generate control loop performance reports regularly, including but not limited to performance indicator change trends, comparisons before and after optimization, and abnormal event records; Key performance indicators are displayed through a graphical interface, allowing operators to intuitively understand the status and health of the control loop; S8. Based on the results of the above steps, continuously iterate and optimize algorithms and control strategies to cope with increasingly complex operating conditions and technological advances.

[0006] In the preferred embodiment, the specific steps of step S2 are: S2.1. The data in step S1 is organized into a data file, and historical data and current data are extracted from the given data file and converted into a matrix form suitable for calculation; historical data matrix for Dimension, current moment data vector for dimension; Normalization: Normalize all data to ensure that the numerical ranges of different variables are consistent and eliminate the impact of dimensional differences on subsequent calculations; For the historical data matrix Elements in , the normalized elements The calculation formula is: ; This formula can unify the value range of data into the interval [0,1], so that data between different variables have the same "status" in subsequent operations, ensuring the rationality of the calculation results, especially in subsequent comparative operations such as distance, to avoid deviations caused by differences in the magnitude of the original data; For the current moment data vector Elements in , the normalized elements The calculation formula is the same: ; S2.2. Construct prediction model: According to the generalized minimum variance control theory, a prediction model of the controlled variable is established, taking into account the past control input and the actual value of the controlled variable The linear relationship between the two models; the least squares method is used to estimate the model parameters and ; Calculate performance indicators: Use the prediction model to calculate the generalized minimum variance performance indicator at the current moment , the formula is: ; in, represents the mathematical expectation operation, is the actual value of the controlled variable at the current moment, is the predicted value; this indicator can quantitatively measure the performance of the current PID control loop; S2.3. Calculate the similar interval corresponding to the historical best performance index: Select the historical best moment: Review historical data and select performance indicators The smallest front moments, forming the best historical data set ; Calculate the dynamic similarity distance: For each data vector at the best historical moment , calculate the data vector normalized with the current moment The dynamic similarity distance , the formula is: ; in, It is The dynamic weight of a variable is calculated as: ; It is an exponential parameter that adjusts the similarity distance calculation characteristics and takes a value greater than 0; S2.4. Comparison and evaluation: Determine the similarity interval: Based on the calculated dynamic similarity distance , find the similar interval corresponding to the nearest historical best moment ; Performance comparison: Calculate the average generalized minimum variance performance index of historical data in a similar interval , and with the current moment The calculation formula for comparison is: ; in, Similar interval The number of historical data points contained in like , indicating that the current control loop performance is good; like , it indicates that the control strategy needs to be further optimized.

[0007] In the preferred embodiment, the specific steps of step S3 are: S3.1. Extract the required data from the given data file, including the historical control input sequence , the actual value sequence of historical controlled variables , historical PID parameter sequence, and current real-time control input and the actual value of the controlled variable ; Normalize all collected data to ensure fair "contribution" of data of different magnitudes in subsequent calculations; S3.2, build dynamic fuzzy neural network DFNN model: determine the network structure of DFNN, including input layer, fuzzification layer, fuzzy rule layer, defuzzification layer and output layer; the input layer node selects the normalized control input value at the current moment, the actual value of the controlled variable and the relevant data of the previous moments; The Gaussian membership function is used to fuzzify the input, converting the clear input data into a membership representation with fuzzy semantics, which is more in line with the uncertainty in the actual control process; According to the fuzzified membership values, fuzzy rules are combined and operated, and the activation strength of each fuzzy rule is calculated by product reasoning. The output of the fuzzy rule layer is defuzzified to obtain a clear output value, i.e., the predicted PID parameter value, which is calculated using the weighted average method; S3.3, Combine multimodal fusion mechanism for model optimization: Establish a prediction model of the controlled variable based on historical data, and obtain the PID parameter values ​​recommended by the predictive control by solving the optimization problem according to the set control objectives ( , , ); The idea of ​​model reference adaptive control is adopted to build a reference model. By comparing the difference between the actual controlled variable and the reference model output, the PID parameters are dynamically adjusted using the adaptive law. , , ), so that the control loop is as close as possible to the performance of the reference model; The PID parameters predicted by the DFNN model, the PID parameters recommended by the predictive control, and the PID parameters adjusted by the adaptive control are weighted and fused to obtain the final PID parameters ( , , ); Fusion weight , , Dynamically adjust according to different control stages and the degree of trust in each mode; S3.4, real-time update and feedback adjustment: apply the final fused PID parameters to the actual PID controller to obtain the real-time control input and actual value of the controlled variable at the next moment; The newly acquired data is added to the historical data, and data preprocessing, model training and parameter optimization are re-performed to achieve adaptive PID parameter optimization and continuous improvement of control performance.

[0008] In the preferred embodiment, the specific steps of step S4 are: S4.1. Extract the performance index data sequence related to the PID control loop from the given data file And the corresponding timestamp information , to clarify the time sequence of each data point; Variational mode decomposition: transform the performance index data series Decompose into IMF components , decomposition is achieved by solving variational problems; VMD can adaptively decompose complex data into multiple relatively stable modal components with different frequency characteristics, which helps to mine the variation information of different scales hidden in the data; Characteristic statistics calculation: Calculate the characteristic statistics of each IMF component, such as the mean , Standard Deviation , skewness and Kurtosis ,These statistics can describe the distribution characteristics of each modal component from different perspectives,,providing rich characteristic information for the subsequent construction of performance evaluation and,trend prediction models; S4.2. Set threshold range and alarm mechanism: Determine a reasonable threshold range based on the characteristic statistics of each IMF component corresponding to the performance indicator data under historical normal operation; , assuming its normal threshold range is , determined by the empirical formula: ; in, It is the first in the historical normal data The average of the means of the IMF components, is the corresponding standard deviation, is the number of historical normal data points selected, The corresponding confidence level is The standard normal distribution quantile of ; Real-time monitoring of the characteristic statistics of each IMF component after the decomposition of the current performance indicator data. Once a characteristic statistic of a certain IMF component exceeds its corresponding threshold range, an alarm is triggered and the alarm time is recorded. and abnormal characteristic information, and send alarm prompts to operators to facilitate rapid location and troubleshooting of problems; S4.3. Build a trend prediction model: The extracted characteristic statistics of each IMF component and the corresponding timestamp information form an input sample set , the output is the predicted value of the performance index corresponding to the future moment ; The long short-term memory network (LSTM) is used to build a trend prediction model. The core calculation formula inside the LSTM unit includes the input gate, forget gate, output gate and cell state update. The LSTM network can handle the long-term dependencies in time series data well, and selectively remember or forget historical information through a complex gating mechanism to improve prediction accuracy. Use historical data to train the constructed LSTM model and optimize the model parameters by minimizing the mean square error (MSE) loss function between the predicted value and the true value. The MSE calculation formula is: ; By minimizing this function, the model is prompted to continuously adjust parameters to improve prediction accuracy; S4.4. Early warning and fault prevention: Use the trained LSTM trend prediction model to predict the performance indicators at future moments and obtain the prediction result sequence ; Analyze the changing trend of performance indicators based on the prediction results and set warning thresholds. If it is predicted that the performance indicators at a certain moment in the future will exceed the warning threshold range, issue a warning message in advance to remind relevant personnel to take corresponding fault prevention or control performance optimization measures.

[0009] Adjust PID parameters and check related equipment to avoid or reduce the impact of possible faults on the entire system.

[0010] In the preferred embodiment, the specific steps of step S5 are: S5.1. Extract performance evaluation result data from the given data file , anomaly detection information data And the current values ​​of control strategy related parameters ; Standardization: Performance evaluation result data For standardization, a standardization formula based on data distribution adaptation is adopted: ; in, It is The mean of the performance evaluation indicators, is the corresponding variance, is a very small positive number. is an adaptive weight; for anomaly detection information and control strategy parameters , perform simple preprocessing operations such as normalization or discretization according to the actual situation; S5.2. Constructing a hybrid intelligent optimization model: Particle swarm optimization algorithm PSO and simulated annealing algorithm SA fusion: Defining particle position vector and the particle velocity vector Through information sharing and position and speed update among particles, the control strategy parameter space is continuously searched to find the parameter combination that makes the performance better; the acceptance criterion of simulated annealing algorithm is introduced to improve the problem that PSO is prone to fall into local optimum; Define comprehensive performance evaluation index function , whose calculation formula is derived based on the performance evaluation result data and anomaly detection information, can comprehensively and quantitatively measure the overall performance under different control strategy parameter combinations, and provide a clear optimization goal for the hybrid intelligent optimization model; S5.3, Control strategy parameter adjustment: Using the constructed hybrid intelligent optimization model, based on the current performance evaluation results and anomaly detection information, start iterative optimization of control strategy parameters; initialize the position, speed, individual optimal, global optimal and other related parameters of the particle swarm, set the initial temperature and other parameters of the simulated annealing algorithm; in each iteration, update the position of the particle, calculate the comprehensive performance evaluation index function value corresponding to each particle , update the individual optimal and global optimal positions and the corresponding function values; continue to iterate until the preset number of iterations is reached Or if the convergence condition is met, the control strategy parameter value corresponding to the final global optimal position is used as the adjusted control strategy parameter ; S5.4. Implement adjustments and continuously monitor the effects: Applied to the actual control loop, it starts to continuously monitor the operation of the control loop, collects performance evaluation related data and abnormal detection information in real time, performs data preprocessing, and obtains new standardized performance evaluation result data and anomaly detection information vector ; Effect evaluation: Calculate the adjusted comprehensive performance evaluation index function value , use the same calculation formula as before, but substitute the adjusted data for calculation to quantitatively evaluate the control effect after adjustment; record the performance evaluation results, anomaly detection information, and comprehensive performance evaluation index function values ​​corresponding to each monitoring moment for further analysis; S5.5. Further optimization feedback based on monitoring results: Regularly analyze the recorded monitoring data to check whether the control loop still has room for performance improvement or abnormal conditions have not been effectively improved; If any optimizable problems are found, return to step S5.2 and reuse the hybrid intelligent optimization model, taking the latest monitoring data as input, continue to adjust the control strategy parameters, repeat steps 5.3 and 5.4 to achieve continuous optimization of the control strategy and ensure that the control loop is always in the optimal working state.

[0011] In the preferred embodiment, the specific steps of step S6 are: S6.1. Extract process-related data from a given data file , equipment related data , load related data And control loop performance indicator data ; Remove obviously erroneous or missing data points, use a filling method based on data distribution similarity for missing data, and use the covariance relationship between process variables to reasonably fill in missing values ​​to ensure data integrity and consistency; S6.2. Integrate process, equipment and load data to build a fusion data matrix ; Multivariate Kernel Density Estimation MVKDE: Fusion Data Matrix Perform multivariate kernel density estimation to obtain the probability density distribution characteristics of the data and extract the eigenvector MVKDE can describe the distribution of data on multi-dimensional data. The extracted feature vectors can more comprehensively reflect the overall characteristics and local change information of the data, and capture the complex patterns and potential relationships in the data. S6.3. Key factor mining based on deep belief network DBN: DBN construction and training: Construct a deep belief network composed of multiple layers of restricted Boltzmann machines (RBM) and extract the fusion data feature vector The data is input into DBN as the input layer; the RBM is trained by the contrastive divergence (CD) algorithm so that the network can learn the potential patterns and rules in the input feature vector; Fine-tuning and importance score calculation: Based on the trained DBN, the control loop performance index data is introduced through the back propagation algorithm Fine-tune the network to further optimize the network's learning of the relationship between key factors and performance indicators; calculate the importance score of each input feature , sort the features according to the scores and find out the key factors affecting the control loop performance; S6.4. Big data analysis and optimization reference mining: Data segmentation: Divide all historical data into multiple data segments in chronological order ; Statistical feature calculation: For each data segment, calculate the average value of the control loop performance index and standard deviation , and the statistical characteristics of each process, equipment, and load variable; Sequential pattern mining: Build a model based on sequential pattern mining to find the pattern rules between the statistical characteristics of variables and the changes in control loop performance indicators between different data segments, and mine frequent pattern sets. ; These patterns can discover the influence of different variable combinations on control loop performance from the time series changes of a large amount of historical data, and provide a reference for the optimization of future control loops; S6.5. Result integration and visualization: Analysis report: Integrate the key factors affecting the control loop performance and the optimization reference mode obtained from big data analysis to form a detailed analysis report; Visual presentation: Heat map, bar chart, and line chart visualization techniques are used to intuitively present the importance scores of key factors, the relationship between variables and performance indicators under different modes, and other results, making it easier for engineers and technicians to understand and further analyze them, so as to better utilize these results in the actual control optimization process.

[0012] In the preferred embodiment, the specific steps of steps S7-S8 are: A1. Extract control loop related performance index data PI, optimization operation record data OR and abnormal event data AE from the given data file; perform weighted moving average processing on the performance index data, adopt an exponential decreasing weight formula to reduce noise interference in the data, make the change trend of the performance index clearer, and emphasize the importance of recent data; A2. Build an autoencoder neural network, which includes an input layer, a hidden layer, and an output layer. By minimizing the reconstruction error, the autoencoder learns the characteristic pattern of normal data. Calculate the anomaly score of each data point, quantify the degree of anomaly, and determine the time and severity of the abnormal event by comparing it with the set anomaly threshold; A3. Use the time series decomposition method to decompose the performance indicator data into trend components, seasonal components and residual components, and intuitively display the long-term change trend of the performance indicators over time; for each optimization operation, calculate the degree of improvement of the performance indicators after optimization, and quantify the improvement effect of the optimization operation on the control loop performance; Combine abnormal event data, abnormality scores and corresponding time information to form a detailed abnormal event report, including the time when the abnormality occurred, abnormality scores, possible causes, etc. Organize the performance indicator change trend, comparison results before and after optimization, abnormal event records and other information into a complete control loop performance report to facilitate engineers and technicians to understand and further analyze; A4. Select several key performance indicators for visualization, draw a line graph of changes over time, mark the time point when abnormal events occur in the graph, and help operators intuitively see the relationship between abnormal events and changes in performance indicators; draw a box plot of performance indicators to show their distribution characteristics, and help operators understand the overall distribution and discreteness of performance indicators, so as to judge the stability and reliability of the control loop; A5. Iterative optimization algorithm and control strategy: Based on the information in the control loop performance report, analyze the problems and deficiencies of the current algorithm and control strategy, and optimize them in a targeted manner; adjust the structure and parameters of the autoencoder to improve its ability to extract performance indicator data features and the accuracy of anomaly detection; Adopt the optimization algorithm based on gradient descent, stochastic gradient descent SGD, to optimize the control strategy parameters so that the control strategy can adapt to new working conditions and technical requirements; Apply the optimized algorithm and control strategy to the control loop, repeat the above steps, and continuously perform iterative optimization to cope with increasingly complex operating conditions and technological advances.

[0013] The present invention provides a thermal control loop quality monitoring method for a thermal power plant, which has many beneficial effects. First, through multi-dimensional data fusion analysis, big data analysis technology is used to mine patterns and laws in historical data, which can warn of potential faults or performance degradation in advance, reduce the probability of abnormal events, reduce downtime and maintenance costs caused by faults, and ensure the safety and reliability of thermal power plant operation. Furthermore, based on abnormal detection information and performance evaluation results, the control strategy can be adjusted automatically or manually, and the adjustment effect can be continuously monitored, so that the control loop can quickly adapt to changes in operating conditions, always maintain the optimal working state, and improve the operating efficiency and economy of the power plant. In addition, the regularly generated detailed performance reports and graphical interfaces display key performance indicators, which are convenient for operators to intuitively understand the control loop status, help to find problems in time and take measures, improve operational convenience and management efficiency, and at the same time, through continuous iteration and optimization of algorithms and control strategies, it can better cope with increasingly complex operating conditions and technological advances, so that the monitoring method has good adaptability and sustainability, and ensures the long-term stable and efficient operation of the thermal control system of the thermal power plant. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is the algorithm flow of the thermal control loop quality monitoring of the thermal power plant of the present invention; DETAILED DESCRIPTION Example 1 like Figure 1 As shown, a thermal control loop quality monitoring method for a thermal power plant comprises: S1, online real-time acquisition of all operating parameters related to the PID control loop, including input signals, output signals, set values, and actual measured values, and normalization of the acquired data; S2. Analyze the performance of the existing PID control loop using the minimum variance or generalized minimum variance performance evaluation method; calculate the actual performance index of the current PID control loop and compare it with the ideal or historical best performance index; S3. For control loops whose control performance is affected by unreasonable PID parameter settings, re-optimize these parameters through data-driven modeling methods; use adaptive control and predictive control advanced control technologies to adjust the proportional, integral and differential coefficients of the PID controller according to real-time data to improve control performance; S4. Set a reasonable threshold range. When the performance indicators of the PID control loop deviate from the normal range, trigger the alarm mechanism. For performance degradation problems in the long-term process, establish a trend prediction model to warn of possible failures or performance degradation in advance. S5. Automatically or manually adjust the control strategy based on the performance evaluation results and abnormal detection information to ensure that the control loop is always in the optimal working state; continuously monitor the effect of the adjustment and record it for subsequent analysis and further optimization; S6. Combine the influence of process, equipment and load factors to conduct multi-dimensional data fusion analysis to find out the key factors affecting the performance of the control loop; use big data analysis technology to explore the patterns and rules hidden in a large amount of historical data to provide reference for future control optimization; S7. Generate control loop performance reports regularly, including but not limited to performance indicator trend, comparison before and after optimization, and abnormal event records; display key performance indicators through a graphical interface so that operators can intuitively understand the status and health of the control loop; S8. Based on the results of the above steps, continuously iterate and optimize algorithms and control strategies to cope with increasingly complex operating conditions and technological advances.

[0015] Example 2 Further illustrate with reference to Example 1, Figure 1 The structure shown in FIG. 1 is shown in FIG. 1 , and the specific steps of step S2 are: S2.1. The data in step S1 is organized into a data file, and historical data and current data are extracted from the given data file and converted into a matrix form suitable for calculation; historical data matrix for Dimension, current moment data vector for dimension; Normalization: Normalize all data to ensure that the numerical ranges of different variables are consistent and eliminate the impact of dimensional differences on subsequent calculations; For the historical data matrix Elements in , the normalized elements The calculation formula is: ; This formula can unify the value range of data into the interval [0,1], so that data between different variables have the same "status" in subsequent operations, ensuring the rationality of the calculation results, especially in subsequent comparative operations such as distance, to avoid deviations caused by differences in the magnitude of the original data; For the current moment data vector Elements in , the normalized elements The calculation formula is the same: ; S2.2. Construct prediction model: According to the generalized minimum variance control theory, a prediction model of the controlled variable is established, taking into account the past control input and the actual value of the controlled variable The linear relationship between the two models; the least squares method is used to estimate the model parameters and ; Calculate performance indicators: Use the prediction model to calculate the generalized minimum variance performance indicator at the current moment , the formula is: ;in, represents the mathematical expectation operation, is the actual value of the controlled variable at the current moment, is the predicted value; this indicator can quantitatively measure the performance of the current PID control loop; S2.3. Calculate the similar interval corresponding to the historical best performance index: Select the historical best moment: Review historical data and select performance indicators The smallest front moments, forming the best historical data set ; Calculate the dynamic similarity distance: For each data vector at the best historical moment , calculate the data vector normalized with the current moment Dynamic similarity distance , the formula is: ; in, It is The dynamic weight of a variable is calculated as: ; It is an exponential parameter that adjusts the similarity distance calculation characteristics and takes a value greater than 0; S2.4. Comparison and evaluation: Determine the similarity interval: Based on the calculated dynamic similarity distance , find the similar interval corresponding to the nearest historical best moment ; Performance comparison: Calculate the average generalized minimum variance performance index of historical data in a similar interval , and with the current moment The calculation formula for comparison is: ; in, Similar interval The number of historical data points contained in , indicating that the current control loop performance is good; if , it indicates that the control strategy needs to be further optimized.

[0016] Here are the detailed steps based on the S2 you provided: Read the historical data related to the PID control loop and the data corresponding to the current moment from the given data file, and organize them into a suitable format, such as a matrix form, to facilitate subsequent calculations. Suppose the historical data matrix is , the dimension is ( is the number of historical data points, is the number of variables contained in each data point, such as control input, controlled quantity, etc.), the current data vector is , the dimension is . Normalize the data to avoid unreasonable impact of data of different magnitudes on subsequent calculations. The commonly used normalization formula is as follows: For the historical data matrix Elements in ( Indicates data points, Indicates variables), the normalized elements The calculation formula is: ; This formula can unify the value range of the data into the interval [0,1], so that the data between different variables have the same "status" in subsequent operations, ensuring the rationality of the calculation results, especially in subsequent comparison operations such as distance, to avoid deviations caused by differences in the magnitude of the original data. Elements in , the normalized elements The calculation formula is the same:

[0017] Step 2: Calculate the generalized minimum variance performance index: First, build a prediction model based on the generalized minimum variance control theory. Suppose the predicted value of the controlled variable is ( represents time) and past control input ( , is the lag order of the control input) and the actual value of the controlled variable in the past ( , is the lag order of the controlled variable) has the following relationship (a linear relationship is used here to simply illustrate, and the actual more complex nonlinear relationship can also be converted by appropriate methods and processed similarly): ; in, and are the model parameters to be determined, is the prediction error term, assuming that its mean is 0 and its variance is Normal distribution. We can use historical data to estimate these parameters by using methods such as least squares method (the detailed least squares derivation is omitted here, assuming that the parameter estimates have been obtained and ).

[0018] Calculate the generalized minimum variance performance index at the current moment , the formula is as follows: ; in, It means mathematical expectation operation, which means finding the average. It is the actual value of the controlled variable at the current moment. This indicator reflects the difference between the actual output of the current control loop and the output of the prediction model built based on historical data. The smaller the value, the better the control effect. This indicator can quantitatively measure the performance of the current PID control loop.

[0019] Step 3: Calculate the similarity interval corresponding to the historical best performance indicator: 1. Review the best performance indicators (generalized minimum variance performance indicators) in the historical data (for example, the previous performance indicator with the smallest value). a moment, It can be set according to the actual situation). Let the historical data vectors corresponding to these moments be (Each The dimension is , which is consistent with the normalized data format above).

[0020] 2. For each historical best moment data vector , calculate it and the normalized data vector at the current moment Dynamic similarity distance , here we use a custom dynamic weight similarity distance formula: ; in, It is The dynamic weight corresponding to each variable is calculated as follows: ; here represents the generalized minimum variance performance index The historical data matrix Variable columns of data (using Indicates that all rows correspond to The covariance of the variables and performance indicators is measured by the covariance to determine the weight. It is an exponential parameter that adjusts the similarity distance calculation characteristics (it can be adjusted according to the actual data distribution and experience, and generally takes a value greater than 0, such as 1.5). This dynamic similarity distance formula can comprehensively consider the weight of the impact of different variables on performance and better capture the similarity characteristics between data through exponential adjustment. By calculating the distance from the historical best performance moment data, it measures the similarity between the current state and the historical good state.

[0021] Step 4: Compare and evaluate 1. Based on the calculated dynamic similarity distance , find the similar interval corresponding to the best historical moment closest to the current moment data (for example, with the most recent moment as the center, expand the data interval to a certain range), set it as .

[0022] 2. Compare the generalized minimum variance performance index at the current moment Similar to this interval The average generalized minimum variance performance index corresponding to the historical data (Calculate the average performance index of all moments in the interval), the calculation formula is: ; in, Similar interval The number of historical data points contained in the .

[0023] 3. By comparison and The performance of the current PID control loop is evaluated by the relationship between the size of , indicating that the current control loop performance has reached or exceeded the similar state corresponding to the historical best performance, and the control effect is good. This indicates that the current control loop performance has room for improvement compared to the similar state corresponding to the historical best performance, and further analysis and adjustment of PID parameters are needed to optimize the control performance.

[0024] Through the above-mentioned generalized minimum variance fusion similarity evaluation algorithm based on dynamic weights, the generalized minimum variance performance index and the similarity analysis with the historical best performance data are comprehensively utilized to evaluate the performance of the PID control loop in a more comprehensive and targeted manner, providing a strong basis for subsequent control strategy adjustment and optimization, especially through some customized formulas to incorporate the characteristics of the data itself and the actual situation of the control loop, making it different from the common simple performance evaluation algorithm.

[0025] Example 3 Further illustrate with reference to Example 1, Figure 1 The structure shown in FIG. 1 is as follows: S3.1. Extracting various required data from a given data file, including the historical control input sequence. , the actual value sequence of historical controlled variables , historical PID parameter sequence, and current real-time control input and the actual value of the controlled variable ; Normalize all collected data to ensure fair “contribution” of data of different magnitudes in subsequent calculations; S3.2, build a dynamic fuzzy neural network DFNN model: determine the network structure of DFNN, including input layer, fuzzification layer, fuzzy rule layer, defuzzification layer and output layer; the input layer node selects the normalized control input value at the current moment, the actual value of the controlled variable and the relevant data of the previous moments; the Gaussian membership function is used to fuzzify the input, and the clear input data is converted into a membership representation with fuzzy semantics, which is more in line with the uncertainty in the actual control process; the fuzzy rules are combined and operated according to the fuzzified membership value, and the activation strength of each fuzzy rule is calculated by product reasoning; the output of the fuzzy rule layer is defuzzified to obtain a clear output value, that is, the predicted PID parameter value, which is calculated by weighted average method; S3.3, Combine multimodal fusion mechanism for model optimization: Establish a prediction model of the controlled variable based on historical data, and obtain the PID parameter values ​​recommended by the predictive control by solving the optimization problem according to the set control objectives ( , , ); Adopt the idea of ​​model reference adaptive control, build a reference model, compare the difference between the actual controlled variable and the reference model output, and use the adaptive law to dynamically adjust the PID parameters ( , , ), so that the control loop is as close to the performance of the reference model as possible; the PID parameters predicted by the DFNN model, the PID parameters recommended by the predictive control, and the PID parameters adjusted by the adaptive control are weighted and fused to obtain the final PID parameters ( , , ); Fusion weight , , Dynamically adjust according to different control stages and the degree of trust in each mode; S3.4, real-time update and feedback adjustment: apply the final fused PID parameters to the actual PID controller to obtain the real-time control input and actual value of the controlled variable at the next moment; add the newly acquired data to the historical data, re-perform data preprocessing, model training and parameter optimization to achieve adaptive PID parameter optimization and continuous improvement of control performance.

[0026] The following are detailed steps based on S3: Step 1: Data collection and preprocessing: Extract the required data from the given data file, including the historical control input sequence ( is the number of historical data points, Indicates control input value at each historical moment), historical controlled variable actual value sequence , and the corresponding historical PID parameters (proportional coefficient , integral coefficient , differential coefficient ) sequence, as well as the current real-time control input , Actual value of the controlled variable ( Indicates the current time) and other data.

[0027] 2. Normalize the collected data so that it falls within a suitable range for subsequent calculations and model training. Taking the control input value as an example, the normalization formula is as follows: ; Normalization is used to unify the scale of raw data of different magnitudes to avoid certain features being masked or having adverse effects on model training due to large differences in data magnitude during subsequent neural network training, thereby ensuring fair "contribution" of each data in subsequent calculations. Similarly, the actual values ​​of the controlled variables, PID parameters, etc. are also normalized in a similar manner.

[0028] Step 2: Construct a dynamic fuzzy neural network (DFNN) model: Determine the network structure of the DFNN, including the input layer, fuzzification layer, fuzzy rule layer, defuzzification layer, and output layer. The input layer node selects the normalized control input value at the current moment. , Actual value of the controlled variable and previous moments (e.g. past a moment, The normalized control input value and actual value of the controlled variable can be selected according to the actual situation, that is, the total The fuzzification layer uses the Gaussian membership function to fuzzify the input. Input variables (here represents one of the various input values ​​mentioned above), and its corresponding The membership function of a fuzzy subset is defined as: ; in, It is The variable The central value of the fuzzy subset, is the corresponding width value. These parameters can be randomly set initially and then adjusted adaptively during the training process. By fuzzification, the clear input data is converted into a membership representation with fuzzy semantics, which is more in line with the fuzziness and uncertainty in the actual control process and prepares for the subsequent reasoning based on fuzzy rules. The fuzzy rule layer combines and operates fuzzy rules according to the fuzzified membership value, and adopts the product reasoning method. The activation strength of the fuzzy rules The calculation formula is as follows: ; in, Indicates The fuzzy rule corresponding to The fuzzy subset number selected for the input variable. This formula determines the activation degree of each fuzzy rule under the current input condition, reflects the influence weight of different rules on the final output, and is based on this for subsequent fuzzy reasoning synthesis. The defuzzification layer defuzzifies the output of the fuzzy rule layer to obtain a clear output value. Using the weighted average method, the predicted PID parameter values ​​output by the output layer (respectively, the predicted proportional coefficient , prediction integral coefficient , predicted differential coefficient ) is calculated as follows: ; in, is the total number of fuzzy rules, , , They are The fuzzy rules correspond to the recommended proportional, integral, and differential coefficient values ​​(which can be set initially based on experience and subsequently adjusted and optimized). The fuzzy results obtained by fuzzy reasoning are converted into practical and clear PID parameter prediction values, so that they can be subsequently applied to the actual control loop for performance improvement.

[0029] Step 3: Combine multimodal fusion mechanism to optimize the model: In addition to the PID parameters predicted by the DFNN model, PID parameter adjustment suggestions based on predictive control and adaptive control technology are also introduced. For the predictive control part, a prediction model of the controlled variable is established based on historical data (assuming that an autoregressive moving average model is used, and the ARMA model is used as an example for simplification). Set the predicted value of the controlled variable The relationship between the past control input and the actual value of the controlled variable is as follows: ; in, , are the autoregressive order and the moving average order, respectively. , are the model parameters to be estimated (which can be estimated using historical data by methods such as the least squares method), is the prediction error term. Based on this prediction model, according to the set control target (such as making the controlled variable track a certain reference value ), by solving an optimization problem (e.g. minimizing the quadratic performance index of the prediction error) to obtain the PID parameter values ​​recommended by the predictive control (denoted as , , ), the specific optimization solution process is omitted here. Using the predictive ability of predictive control, the future change trend of the controlled variable is considered in advance, providing a reference basis for PID parameter adjustment based on predictive performance optimization. For the adaptive control part, the idea based on model reference adaptive control (MRAC) is adopted to build a reference model, and the controlled variable value output by the reference model is set Reference input corresponding to the desired control performance The ideal relationship is as follows: ;in, is the transfer function of the reference model. By comparing the actual controlled variable Output with reference model The adaptive law is used to adjust the PID parameters so that the actual control loop is as close to the performance of the reference model as possible. The PID parameters after adaptive adjustment are set as , , , for example, for the proportionality factor The adaptive adjustment formula is as follows: ; in, is the adaptive learning rate, With the help of the adaptive ability of adaptive control, the PID parameters are dynamically adjusted in real time according to the deviation between the actual control performance and the expected performance, so that the control loop can adapt to different working conditions and changes in system characteristics.

[0030] 3. Perform multimodal fusion, fuse the PID parameters predicted by the DFNN model, the PID parameters recommended by the predictive control, and the PID parameters adjusted by the adaptive control, and use a weighted fusion method to obtain the final PID parameters after fusion (respectively , , ) is calculated as follows: ; in, , , The advantages of different control technologies are comprehensively utilized, and the PID parameter suggestions obtained from each mode are integrated through reasonable weight allocation to obtain the final PID parameters with more adaptability, accuracy and robustness to better improve the control performance.

[0031] Step 4: Real-time update and feedback adjustment: The final PID parameters after fusion , , Applied to the actual PID controller, control the controlled object and obtain the real-time control input at the next moment and the actual value of the controlled variable . Add the newly acquired data to the historical data, return to step 1 for data preprocessing and other operations, iterate the entire algorithm process, and continuously update the parameters of the DFNN model (such as the central value and width value of the membership function, the correlation coefficient of the fuzzy rule, etc.) and the model parameters involved in each control technology (such as the coefficient of the predictive control model, the learning rate of the adaptive control, etc.) according to the real-time data, so as to achieve adaptive PID parameter optimization and continuous improvement of control performance. Through the above-mentioned PID parameter adaptive optimization algorithm based on dynamic fuzzy neural network and multimodal fusion, it fully combines advanced technologies such as data-driven modeling, adaptive control and predictive control, and uses a series of customized mathematical formulas to realize the complete process from data processing, model construction, multimodal fusion to real-time feedback adjustment. It aims to optimize the PID parameters in a complex and targeted way to improve the overall performance of the control loop in the case of unreasonable PID parameter settings affecting the control performance, so as to make it more adaptable to complex and changeable actual control scenarios. However, in actual applications, it is necessary to further adjust the parameters and formula details of each part according to the specific characteristics of the control object and data characteristics to achieve the best optimization effect.

[0032] Example 4 Further illustrate with reference to Example 1, Figure 1The structure shown in FIG. 1 is as follows: S4.1. Extracting the performance index data sequence related to the PID control loop from the given data file And the corresponding timestamp information , to clarify the time sequence of each data point; Variational mode decomposition: convert the performance index data sequence Decompose into IMF components , decomposition is achieved by solving variational problems; VMD can adaptively decompose complex data into multiple relatively stable modal components with different frequency characteristics, which helps to mine the variation information of different scales hidden in the data; characteristic statistics calculation: calculate the characteristic statistics of each IMF component, such as the mean , Standard Deviation , skewness and Kurtosis ,These statistics can describe the distribution characteristics of each modal component from different perspectives,,providing rich characteristic information for the subsequent construction of performance evaluation and,trend prediction models; S4.2. Set threshold range and alarm mechanism: Determine a reasonable threshold range based on the characteristic statistics of each IMF component corresponding to the performance indicator data under historical normal operation; , assuming its normal threshold range is , determined by the empirical formula: ; in, It is the first in the historical normal data The average of the means of the IMF components, is the corresponding standard deviation, is the number of historical normal data points selected, The corresponding confidence level is The standard normal distribution quantile of the current performance indicator data is monitored in real time. Once a characteristic statistic of an IMF component exceeds its corresponding threshold range, an alarm is triggered and the alarm time is recorded. and abnormal characteristic information, and send alarm prompts to operators to facilitate rapid location and troubleshooting of problems; S4.3. Build a trend prediction model: The extracted characteristic statistics of each IMF component and the corresponding timestamp information form an input sample set , the output is the predicted value of the performance index corresponding to the future moment ; The long short-term memory network LSTM is used to build a trend prediction model. The core calculation formula inside the LSTM unit includes the input gate, forget gate, output gate and cell state update; The LSTM network can handle the long-term dependencies in time series data well, and selectively remember or forget historical information through a complex gating mechanism to improve prediction accuracy; Use historical data to train the constructed LSTM model, and optimize the model parameters by minimizing the mean square error MSE loss function between the predicted value and the true value; The MSE calculation formula is: ; By minimizing this function, the model is prompted to continuously adjust parameters to improve prediction accuracy; S4.4. Early warning and fault prevention: Use the trained LSTM trend prediction model to predict the performance indicators at future moments and obtain the prediction result sequence Analyze the changing trend of performance indicators according to the prediction results, set the warning threshold, and issue warning information in advance if it is predicted that the performance indicators will exceed the warning threshold at a certain time in the future to remind relevant personnel to take corresponding fault prevention or control performance optimization measures. Adjust PID parameters and check related equipment to avoid or reduce the impact of possible faults on the entire system.

[0033] The following is the detailed method of S4 step: Step 1: Data collection and feature extraction: Extract the performance index data sequence related to the PID control loop from the incoming data file and set it as ( is the number of data points, Indicates The performance indicator value corresponding to the moment), and extract the corresponding timestamp information , to clarify the time sequence of each data point. Perform variational mode decomposition (VMD) and decompose it into The intrinsic mode function (IMF) components are denoted as The VMD algorithm achieves decomposition by solving the following variational problem: ; in, Indicates IMF components, is the corresponding center frequency, Indicates time The partial derivative of is the Dirac function, * represents the convolution operation, express VMD can adaptively decompose complex performance indicator data into multiple relatively stable modal components with different frequency characteristics, which helps to analyze data features more carefully and mine the change information of different scales hidden in the data. Compared with traditional methods such as empirical mode decomposition, it has better anti-modal aliasing ability and can extract data features more accurately. Calculate the characteristic statistics of each IMF component, the mean , Standard Deviation , skewness and Kurtosis The calculation formula is as follows (taking the mean as an example): ; in, Indicates The IMF component is The value of a moment.

[0034] These statistics can describe the distribution characteristics of each modal component from different perspectives, provide rich feature information for the subsequent construction of performance evaluation and trend prediction models, and help to more comprehensively grasp the changing patterns of performance indicator data.

[0035] Step 2: Set the threshold range and alarm mechanism: Determine a reasonable threshold range based on the characteristic statistics of each IMF component corresponding to the historical performance indicator data under normal operation (data from a period of stable operation and no failure can be selected). , assuming its normal threshold range is , determined by the following empirical formula: ; in, It is the first in the historical normal data The average of the means of the IMF components, is the corresponding standard deviation, is the number of historical normal data points selected, The corresponding confidence level is The formula can scientifically and reasonably determine the normal fluctuation range of the mean of each modal component based on the statistical characteristics of historical data. By analogy, the threshold range of other characteristic statistics such as standard deviation, skewness, kurtosis, etc. can be determined. When the data characteristics monitored in real time exceed these thresholds, the alarm mechanism can be triggered in time to indicate possible performance anomalies. Real-time monitoring of the characteristic statistics of each IMF component after the decomposition of the current performance indicator data. Once a characteristic statistic of a certain IMF component exceeds its corresponding threshold range, an alarm is triggered and the alarm time is recorded. As well as the corresponding abnormal feature information, an alarm prompt is sent to the relevant operators at the same time. The prompt content contains information such as the modal components that may be abnormal, the feature statistics that exceed the threshold, and the corresponding specific values, which is convenient for quick location and troubleshooting.

[0036] Step 3: Construct trend prediction model: The extracted characteristic statistics of each IMF component and the corresponding timestamp information form the input sample set. Let the input matrix be (Each of which is a vector containing the characteristic statistics of each IMF component, and its dimension is consistent with the number of extracted features), and the output is the predicted value of the performance indicator corresponding to the future moment (since we are concerned about the performance degradation trend, the predicted values ​​of the performance indicators at the next few moments can be used as the output here, and the prediction sequence is set to , is the prediction step length). The long short-term memory network (LSTM) is used to build the trend prediction model. The core calculation formula inside the LSTM unit is as follows: ; ; ; ; ; ; in, , , are the activation values ​​of the input gate, forget gate, and output gate, respectively. is the sigmoid activation function, , etc. are the corresponding weight matrices, , etc. are bias terms, represents the element-wise multiplication operation, is the hidden state at the current moment, is the cell state at the current moment, is the input vector at the current moment. The LSTM network can handle long-term dependencies in time series data very well. Through the complex gating mechanism mentioned above, it can selectively remember or forget historical information, thereby more accurately predicting the future trend of performance indicator data, especially for problems with time series characteristics such as performance degradation during long-term changes. Compared with traditional time series prediction methods, it has higher prediction accuracy and adaptability. Use historical data to train the constructed LSTM model, and optimize the model parameters by minimizing the mean square error (MSE) loss function between the predicted value and the true value. The MSE calculation formula is as follows: ;in, It is The sample corresponding to The actual performance indicator value at the prediction moment, is the corresponding predicted value. The mean square error loss function can quantify the difference between the predicted result and the actual result. By minimizing this function, the model is prompted to continuously adjust parameters to improve the accuracy of the prediction, so that the model can better capture the changing trend of performance indicator data, thereby achieving early warning of possible failures or performance degradation.

[0037] Step 4: Early warning and fault prevention: Use the trained LSTM trend prediction model to predict the performance indicators at future moments and obtain the prediction result sequence . According to the prediction results, the trend of performance indicators is analyzed and the warning threshold is set. If the performance indicators are predicted to exceed the warning threshold range at a certain moment in the future, the warning information is issued in advance. The warning information contains the expected time of performance degradation, the decline range and other details, reminding relevant personnel to take corresponding fault prevention or control performance optimization measures in advance, such as adjusting PID parameters, checking related equipment, etc., to avoid or reduce the impact of possible faults on the entire system. Through the above algorithm based on the fusion of variational mode decomposition and long short-term memory network, the variational mode decomposition is first used to extract and decompose the complex PID control loop performance indicator data, and then a reasonable threshold range is set based on the historical data characteristics to realize the alarm mechanism, and then the long short-term memory network is used to build a trend prediction model to predict the performance degradation trend and early warning. The overall algorithm is relatively complex and comprehensively uses some relatively uncommon but highly targeted data processing and modeling methods, aiming to more effectively monitor the performance status of the PID control loop, respond to possible faults or performance degradation problems in advance, and ensure the stable and reliable operation of the system. However, in actual applications, it is necessary to further optimize the parameters of each part and adjust the relevant thresholds according to the specific data characteristics and control scenarios to achieve the best actual effect. Example 5 Further illustrate with reference to Example 1, Figure 1The structure shown in FIG. 1 is as follows: S5.1. Extracting performance evaluation result data from a given data file , anomaly detection information data And the current values ​​of control strategy related parameters ; Standardization: Performance evaluation result data For standardization, a standardization formula based on data distribution adaptation is adopted: ; in, It is The mean of the performance evaluation indicators, is the corresponding variance, is a very small positive number. is an adaptive weight; for anomaly detection information and control strategy parameters , perform simple preprocessing operations such as normalization or discretization according to the actual situation; S5.2. Constructing a hybrid intelligent optimization model: Fusion of particle swarm optimization algorithm PSO and simulated annealing algorithm SA: Defining particle position vector and the particle velocity vector Through information sharing and position and speed update among particles, the control strategy parameter space is continuously searched to find the parameter combination that makes the performance better; the acceptance criterion of simulated annealing algorithm is introduced to improve the problem that PSO is prone to fall into local optimum; the comprehensive performance evaluation index function is defined , whose calculation formula is derived based on the performance evaluation result data and anomaly detection information, can comprehensively and quantitatively measure the overall performance under different control strategy parameter combinations, and provide a clear optimization goal for the hybrid intelligent optimization model; S5.3, Control strategy parameter adjustment: Using the constructed hybrid intelligent optimization model, based on the current performance evaluation results and anomaly detection information, start iterative optimization of control strategy parameters; initialize the position, speed, individual optimal, global optimal and other related parameters of the particle swarm, set the initial temperature and other parameters of the simulated annealing algorithm; in each iteration, update the position of the particle, calculate the comprehensive performance evaluation index function value corresponding to each particle , update the individual optimal and global optimal positions and the corresponding function values; continue to iterate until the preset number of iterations is reached Or if the convergence condition is met, the control strategy parameter value corresponding to the final global optimal position is used as the adjusted control strategy parameter ; S5.4. Implement adjustments and continuously monitor the effects: Applied to the actual control loop, it starts to continuously monitor the operation of the control loop, collects performance evaluation related data and abnormal detection information in real time, performs data preprocessing, and obtains new standardized performance evaluation result data and anomaly detection information vector ; Effect evaluation: Calculate the adjusted comprehensive performance evaluation index function value , use the same calculation formula as before, but substitute the adjusted data for calculation to quantitatively evaluate the control effect after adjustment; record the performance evaluation results, anomaly detection information, and comprehensive performance evaluation index function values ​​corresponding to each monitoring moment for further analysis; S5.5. Further optimization feedback based on monitoring results: Regularly analyze the recorded monitoring data to check whether the control loop still has room for performance improvement or abnormal conditions have not been effectively improved; If any optimizable problems are found, return to step S5.2 and reuse the hybrid intelligent optimization model, taking the latest monitoring data as input, continue to adjust the control strategy parameters, repeat steps 5.3 and 5.4 to achieve continuous optimization of the control strategy and ensure that the control loop is always in the optimal working state.

[0038] The following is a detailed method in step S5: Extract the performance evaluation result related data from the incoming data file and set it as a matrix (in represents different evaluation moments, Represents different performance evaluation indicator dimensions, such as stability indicator, accuracy indicator, etc. Indicates The moment performance evaluation index values), and anomaly detection information data, set as vector ( Indicates At the same time, the current values ​​of the control strategy related parameters are extracted and set as vector ( Indicates the number of different parameters contained in the control strategy, such as the proportional, integral, and differential coefficients in PID control. Indicates control strategy parameters). For standardization, a standardization formula based on data distribution adaptation is adopted: ; in, It is The average of performance evaluation indicators (obtained through historical data statistics), is the corresponding variance, is a very small positive number (to prevent the denominator from being 0), is the adaptive weight, which is calculated as follows: ; This standardization formula can not only normalize the performance evaluation index data of different magnitudes and distributions to a suitable range, which is convenient for subsequent comprehensive analysis, but also considers the impact of the fluctuation range of each index data on standardization through adaptive weights, so that the standardized data can more reasonably reflect the actual performance. For abnormal detection information and control strategy parameters, simple preprocessing operations such as normalization or discretization can be performed according to actual conditions to make their format suitable for subsequent steps.

[0039] Step 2: Build a hybrid intelligent optimization model 1. A hybrid intelligent optimization model is constructed by integrating the particle swarm optimization algorithm (PSO) and the simulated annealing algorithm (SA) to adjust the control strategy parameters. First, define the particle position vector in the particle swarm optimization algorithm: ( represents the number of particles, represents the number of iterations, Indicates The particle in The control strategy parameter value vector corresponding to the iteration has the same dimension as same), the particle velocity vector .

[0040] The particle update formula is as follows: ; Among them, the speed update formula is: ;here, is the inertia weight, and is the learning factor, and is a random number uniformly distributed in the interval $[0, 1]$, It is Particles to The optimal position of the individual up to the iteration, It is to The global optimal position of the population until the iteration. The particle swarm optimization algorithm continuously searches the control strategy parameter space through information sharing and position and speed update among particles to find a parameter combination that makes the performance better.

[0041] 2. The acceptance criterion of the simulated annealing algorithm is introduced to improve the problem that the particle swarm optimization algorithm is prone to fall into the local optimum. Suppose the objective function value corresponding to the current particle position (here it is measured by the comprehensive performance evaluation index function value, which will be specifically defined later) is , the objective function value corresponding to the new updated position is , then the probability calculation formula for accepting a new position with a certain probability is: ; in, is the temperature parameter in the simulated annealing algorithm, which changes with the number of iterations Decrease according to a certain cooling strategy (such as , is the cooling coefficient). Through this probability acceptance criterion, even if the objective function value at the new position may become worse, it is still accepted with a certain probability, which helps to jump out of the local optimum, expand the search range, and increase the possibility of finding the global optimal control strategy parameter combination.

[0042] 3. Define comprehensive performance evaluation index function , whose calculation formula is derived based on the performance evaluation result data and anomaly detection information (the following is an illustrative complex combination formula, which can be adjusted according to the actual control objectives and the emphasis on performance and anomalies): ; in, It is The weight corresponding to each performance evaluation indicator (reflecting its importance to the overall performance, which can be determined by methods such as hierarchical analysis based on historical data and actual control requirements), is the weight corresponding to the anomaly detection information, It is the weight corresponding to the change range of the control strategy parameters. Its purpose is to pursue good performance and reduce the occurrence of anomalies without excessively changing the original control strategy parameters. This comprehensive performance evaluation index function can comprehensively and quantitatively measure the overall performance under different control strategy parameter combinations, provide a clear optimization goal for the hybrid intelligent optimization model, and guide the search in the direction of making the control loop performance better and less abnormal.

[0043] Step 3: Control strategy parameter adjustment 1. Using the constructed hybrid intelligent optimization model, based on the current performance evaluation results and anomaly detection information, start iterative optimization of control strategy parameters. Initialize the position, speed, individual optimal, global optimal and other related parameters of the particle swarm, and set the initial temperature and other parameters of the simulated annealing algorithm.

[0044] 2. In each iteration, according to the update formula of the particle swarm optimization algorithm and the acceptance criterion of the simulated annealing algorithm, the position of the particle (i.e., the control strategy parameter value) is updated, and the comprehensive performance evaluation index function value corresponding to each particle is calculated. , update the individual optimal and global optimal positions and the corresponding function values.

[0045] 3. Continue iterating until the preset number of iterations is reached Or if the convergence condition is met (such as the change of the global optimal value for several consecutive iterations is less than a certain threshold), the control strategy parameter value corresponding to the final global optimal position is used as the adjusted control strategy parameter, recorded as .

[0046] Step 4: Implement adjustments and continue to monitor results 1. Adjust the control strategy parameters Apply it to the actual control loop, start to continuously monitor the operation of the control loop, collect performance evaluation related data and abnormal detection information in real time, and perform data preprocessing according to the method in step 1 to obtain new standardized performance evaluation result data ( represents the number of different monitoring moments after adjustment) and the anomaly detection information vector .

[0047] 2. Calculate the adjusted comprehensive performance evaluation index function value , use the same calculation formula as step 2, but substitute the adjusted data for calculation to quantitatively evaluate the control effect after adjustment.

[0048] 3. At the same time, record the performance evaluation results, anomaly detection information, and comprehensive performance evaluation index function values ​​corresponding to each monitoring moment, and store these data in a dedicated database or data file for further analysis, such as analyzing the long-term impact trend of control strategy adjustments on different performance indicators, changes in the frequency of occurrence of different abnormal situations, etc., in order to provide a basis for subsequent more in-depth optimization.

[0049] Step 5: Further optimization feedback based on monitoring results 1. Analyze the recorded monitoring data regularly (e.g., at regular intervals or after a certain amount of new monitoring data has been accumulated) to check whether the control loop still has room for performance improvement or abnormal conditions have not been effectively improved.

[0050] 2. If there are problems that can be optimized, go back to step 2 again, reuse the hybrid intelligent optimization model, use the latest monitoring data as input, continue to adjust the control strategy parameters, repeat steps 3 and 4, and achieve continuous optimization of the control strategy to ensure that the control loop is always in the optimal working state. Through the above algorithm based on the fusion of hybrid intelligent optimization and dynamic Bayesian network, the control strategy parameters are adjusted by using a complex hybrid intelligent optimization model to comprehensively consider the performance evaluation results and abnormal detection information, and then the effect of the adjustment is continuously monitored and the relevant data is recorded. Finally, feedback optimization is performed based on the monitoring results. The whole process uses multiple custom mathematical formulas to realize quantitative calculation and target guidance of each link, so that the algorithm can not only optimize the control strategy adaptively, but also continuously improve the performance of the control loop with the help of subsequent analysis feedback to ensure its stable and efficient operation. However, in actual applications, it is necessary to further fine-tune the formula parameters, optimization strategies, etc. of each part according to the specific characteristics of the control object and data characteristics to better adapt to the actual control scenario. Example 6 Further illustrate with reference to Example 1, Figure 1 The structure shown in FIG. 1 is as follows: S6.1. Extract process-related data from a given data file , equipment related data , load related data And control loop performance indicator data ; Remove obviously erroneous or missing data points, use a filling method based on data distribution similarity for missing data, and use the covariance relationship between process variables to reasonably fill in missing values ​​to ensure data integrity and consistency; S6.2. Integrate process, equipment and load data to build a fusion data matrix ; Multivariate Kernel Density Estimation MVKDE: Fusion Data Matrix Perform multivariate kernel density estimation to obtain the probability density distribution characteristics of the data and extract the eigenvector MVKDE can describe the distribution of data on multi-dimensional data. The extracted feature vectors can more comprehensively reflect the overall characteristics and local change information of the data, and capture the complex patterns and potential relationships in the data. S6.3. Key factor mining based on deep belief network DBN: DBN construction and training: Construct a deep belief network composed of multiple layers of restricted Boltzmann machines (RBM) and extract the fusion data feature vector The data is input into DBN as the input layer; the RBM is trained by the contrastive divergence (CD) algorithm so that the network can learn the potential patterns and rules in the input feature vector; Fine-tuning and importance score calculation: Based on the trained DBN, the control loop performance index data is introduced through the back propagation algorithm Fine-tune the network to further optimize the network's learning of the relationship between key factors and performance indicators; calculate the importance score of each input feature , sort the features according to the scores and find out the key factors affecting the control loop performance; S6.4. Big data analysis and optimization reference mining: Data segmentation: Divide all historical data into multiple data segments in chronological order ; Statistical feature calculation: For each data segment, calculate the average value of the control loop performance index and standard deviation , and the statistical characteristics of each process, equipment, and load variable; Sequential pattern mining: Build a model based on sequential pattern mining to find the pattern rules between the statistical characteristics of variables and the changes in control loop performance indicators between different data segments, and mine frequent pattern sets. ; These patterns can discover the influence of different variable combinations on control loop performance from the time series changes of a large amount of historical data, and provide a reference for the optimization of future control loops; S6.5. Result integration and visualization: Analysis report: Integrate the key factors affecting the control loop performance and the optimization reference mode obtained from big data analysis to form a detailed analysis report; Visual presentation: Heat map, bar chart, and line chart visualization techniques are used to intuitively present the importance scores of key factors, the relationship between variables and performance indicators under different modes, and other results, making it easier for engineers and technicians to understand and further analyze them, so as to better utilize these results in the actual control optimization process.

[0051] Here are the detailed steps for S6: Step 1: Data collection and preliminary organization 1. Extract process-related data from the incoming data file and set it as a matrix (in represents the number of data samples, represents the number of process-related variables, Indicates The sample process variable values); equipment-related data is set as a matrix ( Indicates the number of device-related variables, Indicates The sample device variable values); load-related data is set as vector ; and the control loop performance index data is set as vector .

[0052] 2. Clean all types of data to remove data points with obvious errors or missing data points. For missing data, a filling method based on data distribution similarity is used. Taking the missing value filling of a variable as an example, the calculation formula is as follows: ; Among them, the weight Based on the covariance calculation between variables, let the covariance matrix be ,but ; This formula uses the covariance relationship between process variables to reasonably fill in missing data based on the values ​​of other related variables, retaining the original information and internal relationships of the data as much as possible, making the data more complete and suitable for subsequent analysis.

[0053] Step 2: Multi-dimensional data fusion and feature extraction 1. Integrate process, equipment and load data to build a fusion data matrix , the dimension is .

[0054] 2. Fusion data matrix Perform multivariate kernel density estimation (MVKDE) to obtain the probability density distribution characteristics of the data and extract the eigenvector. Suppose the multivariate kernel function is (Multivariate Gaussian kernel function is used here), the kernel density estimation formula is as follows: ; in, represents the first Row data vector. For the multivariate Gaussian kernel function ( is the covariance matrix, which can be calculated based on the fused data).

[0055] The probability density value obtained by kernel density estimation and some derived statistics based on probability density (such as the gradient and curvature of probability density) are used as the feature vector of fused data, set as ( is the number of extracted features).

[0056] Multivariate kernel density estimation can describe the distribution pattern of data on multi-dimensional data. The extracted feature vectors can more comprehensively reflect the overall characteristics and local change information of the data. Compared with simple statistical features, it can better capture the complex patterns and potential relationships in the data, which helps to accurately explore the key factors affecting the performance of the control loop in the subsequent process.

[0057] Step 3: Key factor mining based on deep belief network (DBN) 1. Construct a deep belief network, the network structure consists of multiple layers of restricted Boltzmann machines (RBM). As input layer data is input into DBN.

[0058] 2. The energy function of RBM is defined as follows (based on Layer RBM as an example): ; in, represents the visible layer unit (corresponding to the input feature vector), represents the hidden layer unit, is the number of visible layer units (equal to ), is the number of hidden layer units, and are the biases of the visible and hidden layer units, respectively. are the weights connecting the visible and hidden layer units.

[0059] Based on the energy function, the joint probability distribution of visible layer units and hidden layer units is: ; in, is a normalization constant. The RBM is trained using the contrastive divergence (CD) algorithm to adjust the weights and bias , , which enables the network to learn the potential patterns and regularities in the input feature vector.

[0060] Through a multi-layer unsupervised learning structure, the deep belief network can automatically learn the deep representation of data from complex multi-dimensional feature data, explore the potential factor relationships hidden in the data, and thus find out the key factors related to the control loop performance indicators.

[0061] 3. Based on the trained DBN, the control loop performance indicator data is transmitted through the back propagation algorithm The network is introduced and fine-tuned to further optimize the network's learning of the relationship between key factors and performance indicators.

[0062] 4. Calculate the importance score of each input feature to the control loop performance indicator. The importance score calculation formula for a feature is as follows: ;in, is the number of layers of DBN, It is Layer network and The features are sorted according to their importance scores, and the process, equipment, and load factors corresponding to the features with high scores are the key factors affecting the performance of the control loop.

[0063] Step 4: Big data analysis and optimized reference mining 1. Divide all historical data (including process, equipment, load data and control loop performance index data) into multiple data segments in chronological order. ( is the number of data segments).

[0064] 2. For each data segment , calculate the average value of the control loop performance index within this section and standard deviation , as well as the statistical characteristics of each process, equipment, and load variable (such as mean, variance, maximum value, minimum value, etc.).

[0065] 3. Build a model based on sequence pattern mining to find the pattern between the statistical characteristics of variables and the changes in control loop performance indicators between different data segments. ( represents the relationship between a specific combination of variable statistical features and performance index changes), the support calculation formula of the pattern is as follows: ; in, Represents the cardinality of a set. By setting the minimum support threshold , mining frequent pattern sets This pattern mining process can discover the influence of different variable combinations on control loop performance from the time series changes of a large amount of historical data. These rules can provide references for future control loop optimization. For example, when specific processes, equipment, and load conditions change, the possible change trend of control loop performance can be predicted and optimization measures can be taken in advance.

[0066] Step 5: Results integration and visualization The key factors affecting the performance of the control loop and the optimization reference mode obtained by big data analysis are integrated to form a detailed analysis report. Visualization techniques such as heat maps, bar charts, and line charts are used to intuitively present the importance scores of key factors, the relationship between variables and performance indicators under different modes, and other results, which are convenient for engineers and technicians to understand and further analyze, so as to better utilize these results in the actual control optimization process. Through the above-mentioned algorithm based on the fusion of multivariate kernel density estimation and deep belief network, starting from data collection and collation, through multi-dimensional data fusion and feature extraction, using deep belief network to mine key factors, and then mining optimization reference mode by big data analysis, and finally integrating the results and visualizing them, the whole process uses multiple complex mathematical formulas derived from data characteristics, aiming to comprehensively and deeply combine process, equipment, and load factors, and mine the key factors affecting the performance of the control loop, and obtain reference information for control optimization from a large amount of historical data to improve the overall performance and adaptability of the control loop. In practical applications, the parameters and model structure in the algorithm can be further adjusted according to the specific data characteristics and control requirements to achieve better results.

[0067] Example 7 Further illustrate with reference to Example 1, Figure 1 The structure shown in FIG. 1 is shown in FIG. 1 , and the specific steps of steps S7-S8 are as follows: A1. Extract control loop related performance index data PI, optimization operation record data OR and abnormal event data AE from the given data file; perform weighted moving average processing on the performance index data, adopt an exponential decreasing weight formula to reduce noise interference in the data, make the change trend of the performance index clearer, and emphasize the importance of recent data; A2. Build an autoencoder neural network, which includes an input layer, a hidden layer, and an output layer. By minimizing the reconstruction error, the autoencoder learns the characteristic pattern of normal data. Calculate the anomaly score of each data point, quantify the degree of anomaly, and determine the time and severity of the abnormal event by comparing it with the set anomaly threshold; A3. Use the time series decomposition method to decompose the performance indicator data into trend components, seasonal components and residual components, and intuitively display the long-term change trend of the performance indicators over time; for each optimization operation, calculate the degree of improvement of the performance indicators after optimization, and quantify the improvement effect of the optimization operation on the control loop performance; Combine abnormal event data, abnormal score and corresponding time information to form a detailed abnormal event report, including the time when the abnormality occurred, abnormal score, possible cause, etc. Organize the performance indicator change trend, comparison results before and after optimization, abnormal event records and other information into a complete control loop performance report to facilitate engineers and technicians to understand and further analyze; A4. Select several key performance indicators for visualization, draw a line graph of changes over time, mark the time point when abnormal events occur in the graph, and help operators intuitively see the relationship between abnormal events and changes in performance indicators; draw a box plot of performance indicators to show their distribution characteristics, and help operators understand the overall distribution and discreteness of performance indicators, so as to judge the stability and reliability of the control loop; A5. Iterative optimization algorithm and control strategy: Based on the information in the control loop performance report, analyze the problems and deficiencies of the current algorithm and control strategy, and optimize them in a targeted manner; adjust the structure and parameters of the autoencoder to improve its ability to extract performance indicator data features and the accuracy of anomaly detection; Adopt the optimization algorithm based on gradient descent, stochastic gradient descent SGD, to optimize the control strategy parameters so that the control strategy can adapt to new working conditions and technical requirements; Apply the optimized algorithm and control strategy to the control loop, repeat the above steps, and continuously perform iterative optimization to cope with increasingly complex operating conditions and technological advances.

[0068] The specific steps of steps S7-S8 are: Step 1: Data collection and arrangement 1. Extract control loop related data from the data file, including performance indicator data series ( is the number of data points, Indicates The performance index value at each moment), the optimization operation record data (recording the time of each optimization, the parameter values ​​before and after optimization, etc.) is set as a matrix ( To optimize the number of operations, Record the information dimensions for each optimization operation, such as optimization time, original parameter value, new parameter value, etc. Indicates The first optimization operation record information value), the abnormal event data is set as vector ( Indicates Whether an abnormal event occurs at a certain moment, represented by 0 or 1).

[0069] 2. Smooth the performance indicator data using the weighted moving average formula: ; in, is the window size of the moving average, weight Assign data based on their timeliness and importance, for example using exponentially decreasing weights: ; here is the decay factor, which is used to control the speed at which the weight decreases.

[0070] Smoothing can reduce noise interference in the data and make the changing trend of performance indicators clearer. The exponentially decreasing weight emphasizes the greater impact of recent data on the current smoothed value, which is more in line with the relatively high importance of recent data in actual situations.

[0071] Step 2: Build an autoencoder model for feature extraction and anomaly detection 1. Construct an autoencoder neural network, whose structure includes input layer, hidden layer and output layer. The number of nodes in the input layer is equal to the dimension of the performance indicator data, and the number of nodes in the output layer is also the same. The middle hidden layer is used to extract the feature representation of the data.

[0072] 2. The training goal of the autoencoder is to minimize the reconstruction error. The reconstruction error formula is as follows: ; in, is the performance indicator data of the autoencoder for the input The reconstruction output is express The autoencoder is trained with a large amount of normal performance indicator data so that it can learn the characteristic patterns of normal data.

[0073] The autoencoder can automatically learn the intrinsic feature representation of the data. By minimizing the reconstruction error, it can capture the main information patterns in the data. When encountering abnormal data, the reconstruction error will increase significantly, thereby realizing the anomaly detection function.

[0074] 3. Calculate the anomaly score of the performance indicator data using the following formula: ; The anomaly score can quantify the degree of abnormality of each data point, which is convenient for the subsequent analysis and recording of abnormal events. By comparing with the set abnormal threshold, the time and severity of the abnormal event can be determined.

[0075] Step 3: Generate a control loop performance report 1. Calculate the changing trend characteristics of performance indicators and decompose the performance indicator data into trend components using a time series decomposition method , seasonal component (if any) and residual component. The time series decomposition formula is as follows: ; Among them, the trend component can be fitted by methods such as local weighted regression, for example, the LOWESS (Locally Weighted Scatterplot Smoothing) method: ; Among them, the weight Based on the data points and The closer the distance, the greater the weight. The Gaussian kernel function can be used to calculate the weight: ; here is the bandwidth parameter, which controls the decay speed of the weight.

[0076] The trend components obtained through decomposition can intuitively display the long-term change trend of performance indicators over time, which helps to analyze the overall trend of control loop performance and provide direction for subsequent optimization.

[0077] 2. Compare the performance indicators before and after optimization. For each optimization operation, calculate the degree of improvement of the performance indicators after optimization. The formula is as follows: ; in, It is The average value of performance indicators before the optimization operation, is the average value of the performance indicator after optimization.

[0078] This formula can quantify the effect of each optimization operation on the control loop performance, which is convenient for evaluating the effectiveness of different optimization strategies and summarizing optimization experience.

[0079] 3. Integrate abnormal event records, combine abnormal event data AE with abnormal score AS and corresponding time information to form a detailed abnormal event report, including the time of abnormal occurrence, abnormal score, possible abnormal cause and other information.

[0080] 4. Organize the performance indicator change trend, comparison results before and after optimization, abnormal event records and other information into a complete control loop performance report. The report can be in the form of a combination of text and tables to clearly display various information.

[0081] Step 4: Graphical interface displays key performance indicators 1. Select several key performance indicators (such as stability indicators, accuracy indicators, etc.) for visualization. For each key performance indicator, draw a line graph over time, with time as the horizontal axis and the performance indicator value as the vertical axis. At the same time, mark the time point when the abnormal event occurs in the graph and highlight it with markers of different colors or shapes so that operators can intuitively see the relationship between the abnormal event and the change of performance indicators.

[0082] 2. Draw a box plot of the performance indicators to show their distribution characteristics, including the median, quartiles, upper and lower edges, etc., to help operators understand the overall distribution and dispersion of the performance indicators, so as to judge the stability and reliability of the control loop.

[0083] Step 5: Iterative optimization algorithm and control strategy According to the information in the control loop performance report, analyze the problems and deficiencies of the current algorithm and control strategy. For example, if it is found that the performance index continues to decline under a certain working condition, or a certain abnormal event occurs frequently, targeted optimization is required. Adjust the structure and parameters of the autoencoder, such as increasing the number of hidden layer nodes, changing the activation function, etc., to improve its ability to extract the characteristics of performance indicator data and the accuracy of anomaly detection. The adjustment formula is as follows: ;in, is the new number of hidden layer nodes, is the original number of hidden layer nodes, The node number adjustment is determined based on the analysis results. The appropriate value can be determined through experiments and performance evaluation. Optimize the control policy parameters using a gradient descent-based optimization algorithm, such as stochastic gradient descent (SGD). Suppose the control policy parameter vector is , the loss function is , then the parameter update formula is as follows: ;in, is the learning rate, is the loss function with parameters The gradient at. By continuously iteratively updating parameters, the control strategy can adapt to new operating conditions and technical requirements. Apply the optimized algorithm and control strategy to the control loop, repeat steps one to five, and continuously perform iterative optimization to cope with increasingly complex operating conditions and technological advances. This algorithm integrates autoencoders with hierarchical clustering to achieve a complete process from data collection and collation, feature extraction and anomaly detection, performance report generation, graphical interface display to algorithm and control strategy iterative optimization. The custom mathematical formulas in each step are designed according to the data characteristics and the needs of control loop monitoring and optimization, which helps to improve the accuracy of control loop performance monitoring and the effectiveness of control strategy optimization, so that the control loop can maintain a good operating state under complex and changeable working conditions and continuously adapt to the requirements of technological development. In practical applications, it is necessary to further adjust the parameters in the formula and the details of the algorithm according to the specific data characteristics and control requirements to achieve the best effect.

Claims

1. A method for monitoring the quality of a thermal control loop in a thermal power plant, characterized by: The method includes: S1. Collect all operating parameters related to the PID control loop online in real time, including input signal, output signal, set value, actual measurement value, and normalize the collected data; S2. Analyze the performance of the existing PID control loop using the minimum variance or generalized minimum variance performance evaluation method; Calculate the actual performance of the current PID control loop and compare it with the ideal or historical best performance; S3. For control loops whose control performance is affected by unreasonable PID parameter settings, these parameters are re-optimized through data-driven modeling methods; Use adaptive control and predictive control advanced control technology to adjust the proportional, integral and differential coefficients of the PID controller according to real-time data to improve control performance; S4. Set a reasonable threshold range. When the performance index of the PID control loop deviates from the normal range, the alarm mechanism is triggered; For performance degradation issues during long-term changes, a trend prediction model is established to provide early warning of possible failures or performance degradation; S5. Automatically or manually adjust the control strategy based on the performance evaluation results and abnormal detection information to ensure that the control loop is always in the optimal working state; Continuously monitor the effects of adjustments and record them for subsequent analysis and further optimization; S6. Combine the influence of process, equipment and load factors to conduct multi-dimensional data fusion analysis to find out the key factors affecting the performance of the control loop; Use big data analysis technology to discover patterns and rules hidden in large amounts of historical data, providing reference for future control optimization; S7. Generate control loop performance reports regularly, including but not limited to performance indicator change trends, comparisons before and after optimization, and abnormal event records; Key performance indicators are displayed through a graphical interface, allowing operators to intuitively understand the status and health of the control loop; S8. Based on the results of the above steps, continuously iterate and optimize algorithms and control strategies to cope with increasingly complex operating conditions and technological advances.

2. A method for monitoring the quality of a thermal control loop in a thermal power plant according to claim 1, characterized in that: The specific steps of step S2 are: S2.

1. The data in step S1 is organized into a data file, and historical data and current data are extracted from the given data file and converted into a matrix form suitable for calculation; historical data matrix for Dimension, current moment data vector for dimension; Normalization: Normalize all data to ensure that the numerical ranges of different variables are consistent and eliminate the impact of dimensional differences on subsequent calculations; For the historical data matrix Elements in , the normalized elements The calculation formula is: ； This formula can unify the value range of data into the interval [0,1], so that data between different variables have the same "status" in subsequent calculations, ensuring the rationality of the calculation results, especially in subsequent comparative operations such as distance, to avoid deviations caused by differences in the magnitude of the original data; For the current moment data vector Elements in , the normalized elements The calculation formula is the same: ； S2.

2. Construct prediction model: According to the generalized minimum variance control theory, a prediction model of the controlled variable is established, taking into account the past control input and the actual value of the controlled variable The linear relationship between the two models; the least squares method is used to estimate the model parameters and ; Calculate performance indicators: Use the prediction model to calculate the generalized minimum variance performance indicator at the current moment , the formula is: ； in, represents the mathematical expectation operation, is the actual value of the controlled variable at the current moment, is the predicted value; this indicator can quantitatively measure the performance of the current PID control loop; S2.

3. Calculate the similar interval corresponding to the historical best performance index: Select the historical best moment: Review historical data and select performance indicators The smallest front moments, forming the best historical data set ; Calculate the dynamic similarity distance: For each data vector at the best historical moment , calculate the data vector normalized with the current moment The dynamic similarity distance , the formula is: ； in, It is The dynamic weight of a variable is calculated as: ； It is an exponential parameter that adjusts the similarity distance calculation characteristics and takes a value greater than 0; S2.

4. Comparison and evaluation: Determine the similarity interval: Based on the calculated dynamic similarity distance , find the similar interval corresponding to the nearest historical best moment ; Performance comparison: Calculate the average generalized minimum variance performance index of historical data in a similar interval , and with the current moment The calculation formula for comparison is: ； in, Similar interval The number of historical data points contained in like , indicating that the current control loop performance is good; like , it indicates that the control strategy needs to be further optimized.

3. The method for monitoring the quality of a thermal control loop in a thermal power plant according to claim 1, characterized in that: The specific steps of step S3 are: S3.

1. Extract the required data from the given data file, including the historical control input sequence , the actual value sequence of historical controlled variables , historical PID parameter sequence, and current real-time control input and the actual value of the controlled variable ; Normalize all collected data to ensure fair "contribution" of data of different magnitudes in subsequent calculations; S3.2, build dynamic fuzzy neural network DFNN model: determine the network structure of DFNN, including input layer, fuzzification layer, fuzzy rule layer, defuzzification layer and output layer; the input layer node selects the normalized control input value at the current moment, the actual value of the controlled variable and the relevant data of the previous moments; The Gaussian membership function is used to fuzzify the input, converting the clear input data into a membership representation with fuzzy semantics, which is more in line with the uncertainty in the actual control process; According to the fuzzified membership values, fuzzy rules are combined and operated, and the activation strength of each fuzzy rule is calculated by product reasoning. The output of the fuzzy rule layer is defuzzified to obtain a clear output value, i.e., the predicted PID parameter value, which is calculated using the weighted average method; S3.3, Combine multimodal fusion mechanism for model optimization: Establish a prediction model of the controlled variable based on historical data, and obtain the PID parameter values ​​recommended by the predictive control by solving the optimization problem according to the set control objectives ( , , ); The idea of ​​model reference adaptive control is adopted to build a reference model. By comparing the difference between the actual controlled variable and the reference model output, the PID parameters are dynamically adjusted using the adaptive law. , , ), so that the control loop is as close as possible to the performance of the reference model; The PID parameters predicted by the DFNN model, the PID parameters recommended by the predictive control, and the PID parameters adjusted by the adaptive control are weighted and fused to obtain the final PID parameters ( , , ); Fusion weight , , Dynamically adjust according to different control stages and the degree of trust in each mode; S3.4, real-time update and feedback adjustment: apply the final fused PID parameters to the actual PID controller to obtain the real-time control input and actual value of the controlled variable at the next moment; The newly acquired data is added to the historical data, and data preprocessing, model training and parameter optimization are re-performed to achieve adaptive PID parameter optimization and continuous improvement of control performance.

4. The method for monitoring the quality of a thermal control loop in a thermal power plant according to claim 1, characterized in that: The specific steps of step S4 are: S4.

1. Extract the performance index data sequence related to the PID control loop from the given data file And the corresponding timestamp information , to clarify the time sequence of each data point; Variational mode decomposition: transform the performance index data series Decompose into IMF components , decomposition is achieved by solving variational problems; VMD can adaptively decompose complex data into multiple relatively stable modal components with different frequency characteristics, which helps to mine the variation information of different scales hidden in the data; Characteristic statistics calculation: Calculate the characteristic statistics of each IMF component, such as the mean , Standard Deviation , skewness and Kurtosis ,These statistics can describe the distribution characteristics of each modal component from different perspectives,,providing rich characteristic information for the subsequent construction of performance evaluation and,trend prediction models; S4.

2. Set threshold range and alarm mechanism: Determine a reasonable threshold range based on the characteristic statistics of each IMF component corresponding to the performance indicator data under historical normal operation; , assuming its normal threshold range is , determined by the empirical formula: ； in, It is the first in the historical normal data The average of the means of the IMF components, is the corresponding standard deviation, is the number of historical normal data points selected, The corresponding confidence level is The standard normal distribution quantile of ; Real-time monitoring of the characteristic statistics of each IMF component after the decomposition of the current performance indicator data. Once a characteristic statistic of a certain IMF component exceeds its corresponding threshold range, an alarm is triggered and the alarm time is recorded. and abnormal characteristic information, and send alarm prompts to operators to facilitate rapid location and troubleshooting of problems; S4.

3. Build a trend prediction model: The extracted characteristic statistics of each IMF component and the corresponding timestamp information form an input sample set , the output is the predicted value of the performance index corresponding to the future moment ; The long short-term memory network (LSTM) is used to build a trend prediction model. The core calculation formula inside the LSTM unit includes the input gate, forget gate, output gate and cell state update. The LSTM network can handle the long-term dependencies in time series data well, and selectively remember or forget historical information through a complex gating mechanism to improve prediction accuracy. Use historical data to train the constructed LSTM model and optimize the model parameters by minimizing the mean square error (MSE) loss function between the predicted value and the true value. The MSE calculation formula is: ； By minimizing this function, the model is prompted to continuously adjust parameters to improve prediction accuracy; S4.

4. Early warning and fault prevention: Use the trained LSTM trend prediction model to predict the performance indicators at future moments and obtain the prediction result sequence ; Analyze the changing trend of performance indicators based on the prediction results and set warning thresholds. If it is predicted that the performance indicators at a certain moment in the future will exceed the warning threshold range, issue a warning message in advance to remind relevant personnel to take corresponding fault prevention or control performance optimization measures. Adjust PID parameters and check related equipment to avoid or reduce the impact of possible faults on the entire system.

5. The method for monitoring the quality of a thermal control loop in a thermal power plant according to claim 1, characterized in that: The specific steps of step S5 are: S5.

1. Extract performance evaluation result data from the given data file , anomaly detection information data And the current values ​​of control strategy related parameters ; Standardization: Performance evaluation result data For standardization, a standardization formula based on data distribution adaptation is adopted: ； in, It is The mean of the performance evaluation indicators, is the corresponding variance, is a very small positive number. is an adaptive weight; for anomaly detection information and control strategy parameters , perform simple preprocessing operations such as normalization or discretization according to the actual situation; S5.

2. Constructing a hybrid intelligent optimization model: Fusion of particle swarm optimization algorithm PSO and simulated annealing algorithm SA: Defining particle position vector and the particle velocity vector Through information sharing and position and speed update among particles, the control strategy parameter space is continuously searched to find the parameter combination that makes the performance better; the acceptance criterion of simulated annealing algorithm is introduced to improve the problem that PSO is prone to fall into local optimum; Define comprehensive performance evaluation index function , whose calculation formula is derived based on the performance evaluation result data and anomaly detection information, can comprehensively and quantitatively measure the overall performance under different control strategy parameter combinations, and provide a clear optimization goal for the hybrid intelligent optimization model; S5.3, Control strategy parameter adjustment: Using the constructed hybrid intelligent optimization model, based on the current performance evaluation results and anomaly detection information, start iterative optimization of control strategy parameters; initialize the position, speed, individual optimal, global optimal and other related parameters of the particle swarm, set the initial temperature and other parameters of the simulated annealing algorithm; in each iteration, update the position of the particle, calculate the comprehensive performance evaluation index function value corresponding to each particle , update the individual optimal and global optimal positions and the corresponding function values; continue to iterate until the preset number of iterations is reached Or if the convergence condition is met, the control strategy parameter value corresponding to the final global optimal position is used as the adjusted control strategy parameter ; S5.

4. Implement adjustments and continuously monitor the effects: Applied to the actual control loop, it starts to continuously monitor the operation of the control loop, collects performance evaluation related data and abnormal detection information in real time, performs data preprocessing, and obtains new standardized performance evaluation result data and anomaly detection information vector ; Effect evaluation: Calculate the adjusted comprehensive performance evaluation index function value , use the same calculation formula as before, but substitute the adjusted data for calculation to quantitatively evaluate the control effect after adjustment; record the performance evaluation results, anomaly detection information, and comprehensive performance evaluation index function values ​​corresponding to each monitoring moment for further analysis; S5.

5. Further optimization feedback based on monitoring results: Regularly analyze the recorded monitoring data to check whether the control loop still has room for performance improvement or abnormal conditions have not been effectively improved; If any optimizable problems are found, return to step S5.2 and reuse the hybrid intelligent optimization model, taking the latest monitoring data as input, continue to adjust the control strategy parameters, repeat steps 5.3 and 5.4 to achieve continuous optimization of the control strategy and ensure that the control loop is always in the optimal working state.

6. The method for monitoring the quality of a thermal control loop in a thermal power plant according to claim 1, characterized in that: The specific steps of step S6 are: S6.

1. Extract process-related data from a given data file , equipment related data , load related data And control loop performance indicator data ; Remove obviously erroneous or missing data points, use a filling method based on data distribution similarity for missing data, and use the covariance relationship between process variables to reasonably fill in missing values ​​to ensure data integrity and consistency; S6.

2. Integrate process, equipment and load data to build a fusion data matrix ; Multivariate Kernel Density Estimation MVKDE: Fusion Data Matrix Perform multivariate kernel density estimation to obtain the probability density distribution characteristics of the data and extract the eigenvector MVKDE can describe the distribution of data on multi-dimensional data. The extracted feature vectors can more comprehensively reflect the overall characteristics and local change information of the data, and capture the complex patterns and potential relationships in the data. S6.

3. Key factor mining based on deep belief network DBN: DBN construction and training: Construct a deep belief network composed of multiple layers of restricted Boltzmann machines (RBM) and extract the fusion data feature vector The data is input into DBN as the input layer; the RBM is trained by the contrastive divergence (CD) algorithm so that the network can learn the potential patterns and rules in the input feature vector; Fine-tuning and importance score calculation: Based on the trained DBN, the control loop performance index data is introduced through the back propagation algorithm Fine-tune the network to further optimize its learning of the relationship between key factors and performance indicators; Calculate the importance score for each input feature , sort the features according to the scores and find out the key factors affecting the control loop performance; S6.

4. Big data analysis and optimization reference mining: Data segmentation: Divide all historical data into multiple data segments in chronological order ; Statistical feature calculation: For each data segment, calculate the average value of the control loop performance index and standard deviation , and the statistical characteristics of each process, equipment, and load variable; Sequential pattern mining: Build a model based on sequential pattern mining to find the pattern rules between the statistical characteristics of variables and the changes in control loop performance indicators between different data segments, and mine frequent pattern sets. ; These patterns can discover the influence of different variable combinations on control loop performance from the time series changes of a large amount of historical data, and provide a reference for the optimization of future control loops; S6.

5. Result integration and visualization: Analysis report: Integrate the key factors affecting the control loop performance and the optimization reference mode obtained from big data analysis to form a detailed analysis report; Visual presentation: Heat map, bar chart, and line chart visualization techniques are used to intuitively present the importance scores of key factors, the relationship between variables and performance indicators under different modes, and other results, making it easier for engineers and technicians to understand and further analyze them, so as to better utilize these results in the actual control optimization process.

7. The method for monitoring the quality of a thermal control loop in a thermal power plant according to claim 1, characterized in that: The specific steps of steps S7-S8 are: A1. Extract control loop related performance index data PI, optimization operation record data OR and abnormal event data AE from the given data file; perform weighted moving average processing on the performance index data, adopt an exponential decreasing weight formula to reduce noise interference in the data, make the change trend of the performance index clearer, and emphasize the importance of recent data; A2. Build an autoencoder neural network, which includes an input layer, a hidden layer, and an output layer. By minimizing the reconstruction error, the autoencoder learns the characteristic pattern of normal data. Calculate the anomaly score of each data point, quantify the degree of anomaly, and determine the time and severity of the abnormal event by comparing it with the set anomaly threshold; A3. Use the time series decomposition method to decompose the performance indicator data into trend components, seasonal components and residual components, and intuitively display the long-term change trend of the performance indicators over time; For each optimization operation, calculate the degree of improvement of the performance indicators after optimization, and quantify the improvement effect of the optimization operation on the control loop performance; Combine abnormal event data, abnormal score and corresponding time information to form a detailed abnormal event report, including the time when the abnormality occurred, abnormal score, possible cause, etc. Organize the performance indicator change trend, comparison results before and after optimization, abnormal event records and other information into a complete control loop performance report to facilitate engineers and technicians to understand and further analyze; A4. Select several key performance indicators for visualization, draw a line graph of changes over time, mark the time point when abnormal events occur in the graph, and help operators intuitively see the relationship between abnormal events and changes in performance indicators; draw a box plot of performance indicators to show their distribution characteristics, and help operators understand the overall distribution and discreteness of performance indicators, so as to judge the stability and reliability of the control loop; A5. Iterative optimization algorithm and control strategy: Based on the information in the control loop performance report, analyze the problems and deficiencies of the current algorithm and control strategy, and optimize them in a targeted manner; adjust the structure and parameters of the autoencoder to improve its ability to extract performance indicator data features and the accuracy of anomaly detection; Adopt the optimization algorithm based on gradient descent, stochastic gradient descent SGD, to optimize the control strategy parameters so that the control strategy can adapt to new working conditions and technical requirements; Apply the optimized algorithm and control strategy to the control loop, repeat the above steps, and continuously perform iterative optimization to cope with increasingly complex operating conditions and technological advances.

Images