A thermal power unit initial pressure dynamic optimization control system based on big data

Abstract

The application discloses a kind of based on big data's thermal power unit initial pressure dynamic optimization control system, it is related to industrial automation technical field, the system includes data preprocessing module, for obtaining multidimensional parameter record from industrial unit historical operation database, using standardization method to original data is dimension uniform treatment, while reducing data noise interference by wavelet transform denoising technology, if missing data is found, missing point is completed by linear interpolation method, and the complete historical data set after processing is obtained;The based on big data's thermal power unit initial pressure dynamic optimization control system can effectively predict industrial unit operation risk, improve early warning accuracy, and continuously improve prediction model performance by adaptive optimization, provide strong guarantee for the safe and stable operation of industrial equipment.

Description

Technical Field

[0001] The present invention relates to the field of industrial automation technology, and in particular to a dynamic optimization control system for initial pressure of a thermal power unit based on big data. Background Art

[0002] As core equipment in modern manufacturing and energy systems, the safe and stable operation of industrial units is directly related to production efficiency, economic benefits, and personnel safety. With the continuous advancement of industrial automation, the operating environment of units is becoming increasingly complex, and various potential risk factors are intertwined, posing a serious challenge to traditional passive safety management models. Current mainstream risk management methods rely primarily on a passive approach of regular inspections and post-fault repairs, which has significant shortcomings. Traditional methods are often based on fixed maintenance cycles and empirical judgments, lacking in-depth analysis of real-time operating conditions, making it difficult to accurately predict the timing and extent of risk. Furthermore, existing monitoring systems often use single-parameter threshold alarms, which cannot effectively identify complex risk patterns arising from the interaction of multiple parameters. This leads to delayed risk identification and untargeted response measures. A core challenge facing unit risk management stems from the difficulty in accurately extracting and identifying risk signature patterns embedded in historical operating data. Vast amounts of historical operating data record the performance of units under different operating conditions, but the high-risk operating patterns hidden within this data often have multidimensional and nonlinear characteristics, making it difficult for traditional analysis methods to effectively identify them. This lack of risk identification capability directly leads to issues with the timeliness of risk warnings. Even when potential risks are identified, there is often insufficient lead time to implement effective preventive measures. This lack of timely risk warnings further exposes the problem of delayed control strategy adjustments. By the time the system detects a risk signal, the unit may have already entered or is approaching a dangerous state. At this point, adjusting control parameters is often too late to fundamentally prevent the risk from occurring. Summary of the Invention

[0003] The purpose of the present invention is to provide a dynamic optimization control system for the initial pressure of thermal power units based on big data, to construct a risk identification system based on deep mining of historical data, to achieve accurate prediction of high-risk operating conditions, to actively adjust the control strategy at the incipient stage of risks, and to establish a rapid response risk response mechanism.

[0004] To achieve the above-mentioned object, the present invention provides the following technical solution: a dynamic optimization control system for initial pressure of a thermal power unit based on big data, the system comprising:

[0005] The data preprocessing module is used to obtain multi-dimensional parameter records from the historical operation database of industrial units, and uses standardized methods to unify the dimensions of the original data. At the same time, wavelet transform denoising technology is used to reduce data noise interference. If data is missing, the missing points are supplemented by linear interpolation to obtain a complete historical data set after processing;

[0006] The risk modeling and processing module is used to obtain an updated risk prediction model based on the processed complete historical data set. For the updated risk prediction model, the latest operating parameter data is obtained and the parameter change rate and fluctuation range are calculated using sliding window technology. If the fluctuation range of the vibration or pressure parameters exceeds the preset safety range, the abnormal pattern screening technology is used again to identify whether the current operating state belongs to the risk mode and determine the current risk level.

[0007] The dynamic control module is used to dynamically adjust the unit operating parameters based on the current risk level using a real-time control system. The adjustment priority is determined based on the parameter weight distribution results. If the risk probability value continues to drop below the preset safety threshold, the current control parameters are maintained to achieve a stable operating state.

[0008] The operation optimization recording module is used to continuously record the detailed information of each parameter adjustment for a stable operation state, store the adjustment data through the update log recording technology, and analyze the changing trend of the operation mode distribution diagram to determine the long-term operation mode optimization direction.

[0009] Preferably, the risk modeling and processing module obtains an updated risk prediction model based on the processed complete historical data set, including:

[0010] Based on the processed complete historical data set, a clustering algorithm is used to classify and identify the operating modes under different working conditions. Specifically, the operating condition feature extraction technology is used to separate the characteristic values ​​of multi-dimensional parameters including temperature, pressure, and vibration from the data, and the Euclidean distance is used as the distance metric to calculate the differences between the operating modes. At the same time, the maximum number of iterations and the center point offset threshold are set as the iterative convergence conditions to generate a set of operating modes with pattern classification labels.

[0011] Preferably, the risk modeling and processing module obtains an updated risk prediction model based on the processed complete historical data set, further comprising:

[0012] For the generated operating mode set, the parameter weight distribution method is used to determine the influence of each parameter on mode differentiation. The abnormal mode screening technology is combined to identify abnormal operating modes with a small sample size, and an operating mode distribution map is generated as the result visualization output to determine the characteristic parameter combination of the risky operating mode.

[0013] Preferably, the risk modeling and processing module obtains an updated risk prediction model based on the processed complete historical data set, further comprising:

[0014] According to the determined risk operation mode characteristic parameter combination, a time series-based risk prediction model is constructed. Specifically, the operation data of the past period is extracted as the training basis through the historical data backtracking method. If the similarity between the current operation parameter combination and the characteristic value of the high-risk mode exceeds the preset threshold, a risk warning is triggered, and the probability of risk occurrence and the expected time window of occurrence are obtained.

[0015] Preferably, the risk modeling and processing module obtains an updated risk prediction model based on the processed complete historical data set, further comprising:

[0016] For the triggered risk warning and the expected occurrence time window, the real-time error calculation method is used to compare the deviation between the predicted risk value and the actual risk value, and the time interval for updating the model parameters is determined through the feedback cycle setting technology. At the same time, the matching percentage between the model prediction results and the actual results is calculated using the prediction accuracy evaluation method to obtain the performance indicators of the current risk prediction model.

[0017] Preferably, the risk modeling and processing module obtains an updated risk prediction model based on the processed complete historical data set, further comprising:

[0018] According to the performance indicators of the current risk prediction model, if the prediction accuracy is lower than the preset threshold, the model parameter weights are dynamically updated through the weight adjustment step technology. At the same time, the warning threshold interval is adjusted in combination with the threshold dynamic range method, and the model is optimized based on the adjustment convergence target to obtain the updated risk prediction model.

[0019] Preferably, the specific formula for analyzing the change trend of the operation mode distribution diagram is:

[0020] ;

[0021] in, Indicates the predicted value of the dynamic change trend of the operating status parameters of the thermal power unit, Indicates the overall sensitivity of the model output to input disturbances, Indicates time The actual disturbance signal value, Indicates time Abnormal concentration, Indicates the current moment, represents the historical integral variable, Represents the base of natural logarithms.

[0022] Preferably, the overall sensitivity of the model output to the input disturbance The calculation formula is: ;

[0023] in, Indicates the overall sensitivity of the model output to input disturbances, Indicates the current actual disturbance signal value, Represents the mean of historical disturbance signals.

[0024] Preferably, the time Abnormal concentration The specific calculation formula is:

[0025] ;

[0026] in, Indicates time Abnormal concentration, Indicates time The actual disturbance signal value, Represents the median of the disturbance values ​​within the past time window.

[0027] Preferably, the standardization method in the data preprocessing module is a Z-Score standardization method.

[0028] It can be seen from the above technical solution that the present invention has the following beneficial effects:

[0029] This big data-based dynamic optimization control system for the initial pressure of thermal power units standardizes and denoises historical operating data, uses a clustering algorithm to identify operating modes under different working conditions, and determines the characteristic parameter combination of high-risk modes. Based on this, a time series risk prediction model is constructed, and an early warning is triggered when the similarity between the operating parameters and the high-risk mode exceeds a threshold. The present invention continuously optimizes the prediction accuracy through real-time error calculation and dynamic updating of model parameters. At the same time, it combines sliding window technology to monitor parameter changes, judge the risk level in real time, and make dynamic adjustments, ultimately achieving stable operation of the unit. The present invention can effectively predict the operating risks of industrial units, improve the accuracy of early warnings, and continuously improve the performance of the prediction model through adaptive optimization, providing a strong guarantee for the safe and stable operation of industrial equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a connection diagram of the system modules of the present invention. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0032] like Figure 1As shown, the present invention provides a technical solution: a dynamic optimization control system for initial pressure of a thermal power unit based on big data, the system comprising:

[0033] The data preprocessing module is used to obtain multi-dimensional parameter records from the historical operation database of industrial units, and uses standardized methods to unify the dimensions of the original data. At the same time, wavelet transform denoising technology is used to reduce data noise interference. If data is missing, the missing points are supplemented by linear interpolation to obtain a complete historical data set after processing;

[0034] The risk modeling and processing module is used to obtain an updated risk prediction model based on the processed complete historical data set. For the updated risk prediction model, the latest operating parameter data is obtained and the parameter change rate and fluctuation range are calculated using sliding window technology. If the fluctuation range of the vibration or pressure parameters exceeds the preset safety range, the abnormal pattern screening technology is used again to identify whether the current operating state belongs to the risk mode and determine the current risk level.

[0035] The dynamic control module is used to dynamically adjust the unit operating parameters based on the current risk level using a real-time control system. The adjustment priority is determined based on the parameter weight distribution results. If the risk probability value continues to drop below the preset safety threshold, the current control parameters are maintained to achieve a stable operating state.

[0036] The operation optimization recording module is used to continuously record the detailed information of each parameter adjustment for a stable operation state, store the adjustment data through the update log recording technology, and analyze the changing trend of the operation mode distribution diagram to determine the long-term operation mode optimization direction.

[0037] The core of this implementation lies in the integration of big data analysis, multi-dimensional parameter modeling, and intelligent dynamic control technologies to achieve dynamic optimization and regulation of the initial pressure of thermal power units. During the initial system operation, the data preprocessing module extracts historical operating parameter data on multiple dimensions, including temperature, pressure, flow, vibration, current, and voltage, from the DCS or historical data storage server within the industrial control system. Considering that data may originate from different time periods, different sensor types, or have inconsistent sampling frequencies, the data is first normalized using a Z-score to ensure comparability among parameters while preserving the data's fluctuation characteristics.

[0038] To eliminate the impact of high-frequency noise caused by sensor errors, electromagnetic interference, and other factors on subsequent modeling, the system uses wavelet packet transforms (such as Daubechies wavelets) to perform multi-scale decomposition and reconstruction of the data, identifying and filtering out high-frequency interference components. For missing points in historical data caused by communication interruptions, equipment maintenance, and other factors, the system automatically uses linear or spline interpolation methods to fill in the gaps, ensuring the continuity and integrity of the underlying modeling data, ultimately forming a high-quality, complete dataset.

[0039] Based on a complete historical dataset, the risk modeling and processing module uses machine learning algorithms such as support vector machines (SVMs), extreme gradient boosting trees (XGBoost), and long short-term memory networks (LSTMs) to train risk prediction models. These models are used to identify potential system instability risks under different operating conditions. The system collects real-time parameter data for the current operating cycle and uses a sliding window mechanism (e.g., over the past 5 minutes, 30 minutes, 2 hours, and other granularities) to calculate fluctuation indicators such as the rate of change, standard deviation, and peak-to-peak value for each parameter.

[0040] If any key parameter (such as main steam pressure or high-pressure cylinder vibration) is detected to exceed a dynamically set safety threshold (for example, within a ±3σ range), the system immediately activates the anomaly identification submodule. This module uses a K-means clustering algorithm or Mahalanobis distance-based outlier detection method to determine whether the current state is in an abnormal operating mode defined in the training set. It also uses historical event labels to determine the risk level (such as Level 1 alert, Level 2 warning, etc.).

[0041] After identifying the risk level, the dynamic control module activates and prioritizes operational parameters (such as reheater outlet temperature and feedwater pump flow) based on their impact on system stability (e.g., weights determined by principal component analysis (PCA) or the Analytic Hierarchy Process (AHP)). The control strategy utilizes model predictive control (MPC) or PID adaptive control algorithms, issuing real-time commands to adjust actuators (such as throttles, electric valves, and speed governors) to minimize system dynamic offsets.

[0042] The control adjustment process continuously tracks the risk probability output curve (such as the probability value output by the LSTM model). When the risk probability value is lower than the set safety threshold (for example, 0.05) for several consecutive cycles, the system determines that the operating status has returned to the safe area. At this time, the control parameters remain unchanged and enter the stable maintenance state.

[0043] To achieve long-term operational optimization, the operation optimization recording module stores information such as input and output status, timestamps, and risk level changes for each adjustment in a structured manner in an optimization log database. Furthermore, through association analysis (Apriori), operation mode map construction, and change detection techniques, it extracts the evolution patterns of typical operation modes, laying the data foundation for the subsequent construction of a self-learning control model.

[0044] The core principles are mainly divided into the following aspects:

[0045] First, the system extracts operating parameter data from the unit's historical database, covering multiple dimensions, including but not limited to main steam pressure, reheat steam temperature, feedwater flow, main steam valve opening, and unit vibration levels. Because these data are not uniform in their original dimensions and may be affected by differences in sensor type and sampling time, the system standardizes all data for uniform processing. Specifically, each parameter is subtracted from its historical mean and then divided by its standard deviation, ensuring consistency in the relative scale of change across all parameters.

[0046] After data standardization, in order to eliminate the interference of high-frequency noise (such as abnormal sampling values ​​caused by electromagnetic interference and mechanical vibration) on subsequent analysis, the system uses a hierarchical signal decomposition technology (such as wavelet decomposition) to break down the original signal into multiple frequency bands, and then remove the high-frequency interference part, retaining only the low-frequency part that is representative of the system trend, and then reconstruct the denoised operating data.

[0047] During the actual data collection process, some data is often missing due to communication interruptions or equipment maintenance. For these missing parts, the system will find the adjacent valid data points and use linear interpolation to make reasonable calculations on the time axis to fill in the gaps, thus obtaining a complete and continuous historical data set.

[0048] Next, during each operating cycle, the system calculates the rate of change and the fluctuation range between the maximum and minimum values ​​of real-time parameters within a set time window (for example, the past 5 minutes or the past 30 minutes). If the fluctuation range of a key parameter exceeds the safe range, the system's current operating state is deemed risky. For example, if the difference between the maximum and minimum values ​​of the main steam pressure within the sliding time period exceeds 1.5 MPa (this value is derived from national thermal power safety operation standards and historical accident data), it is considered an abnormal pressure fluctuation.

[0049] After identifying a parameter anomaly, the system calls a pre-trained risk identification model to further classify the current status. This model, built on extensive historical data, outputs the probability that the current status represents a risk. If this probability exceeds a set risk threshold, for example, greater than 50%, the corresponding risk warning level is triggered.

[0050] Once a risk is identified, the system uses a weighted priority control mechanism to determine which operating parameters require prioritization. For example, if both reheat steam temperature and high-pressure cylinder vibration are abnormal, and historical analysis indicates that the former is more likely to cause system instability, the temperature parameter will be adjusted first. This control and adjustment process uses a gradual trial-and-error approach: after making small adjustments to a parameter, the system observes whether the risk probability decreases. If a significant downward trend is observed, the system continues to fine-tune the current parameter until the risk value drops below a safe threshold (for example, below 5%).

[0051] If the risk value remains below the safety threshold for multiple consecutive operating cycles, the system determines the current operating state is stable and no further parameter changes will be made. During this stable phase, the system continuously records the details of each parameter adjustment, including the value change before and after the adjustment, the time point, the adjustment range, and the corresponding risk level, forming a detailed operation log.

[0052] This log data is further analyzed to identify the parameter combinations most conducive to long-term stable system operation, creating a map of multiple operating modes. Based on the evolving trends of these maps, the system can autonomously update risk models and optimize control strategies, achieving long-term self-learning and self-optimization.

[0053] This implementation method can expand the scale and intelligence of optimized control through modular division of labor and cooperation, build a more accurate risk model through processed data, and greatly improve the assessment effect; respond to risk levels through dynamic control and real-time adjustment technology to reduce accident risks; recording and analysis patterns can further promote the standardization of operating modes and improve unit operating efficiency.

[0054] The risk modeling and processing module obtains an updated risk prediction model based on the processed complete historical data set, including classifying and identifying the operating modes under different working conditions using a clustering algorithm. Specifically, the operating condition feature extraction technology is used to separate the characteristic values ​​of multi-dimensional parameters including temperature, pressure, and vibration from the data, and the Euclidean distance is used as the distance metric to calculate the differences between the operating modes. At the same time, the maximum number of iterations and the center point offset threshold are set as iterative convergence conditions to generate an operating mode set with pattern classification labels.

[0055] The key to this implementation lies in automatically identifying characteristic patterns of thermal power units under different operating conditions in historical data through unsupervised learning and cluster analysis. First, the system invokes the operating condition feature extraction module to extract representative variables from the standardized and noise-removed historical data, including parameters such as main steam temperature, reheat steam temperature, main steam pressure, and high-pressure cylinder vibration. For each time period or operating cycle, the statistical characteristic values ​​of these variables, such as the mean, standard deviation, and maximum and minimum values, are calculated to form a multi-dimensional operating condition feature vector.

[0056] Subsequently, the system inputs all feature vectors into a clustering algorithm (such as K-means or Gaussian mixture model). The core of clustering is to divide samples according to the distance metric. Here, Euclidean distance is used to calculate the spatial distance between two samples, and the clustering results are iteratively optimized by minimizing the sum of the squares of the distances between the samples and the corresponding cluster centers. In order to prevent the algorithm from falling into an infinite loop or overfitting, the system sets two termination conditions: one is to reach the set maximum number of iterations (such as 100 times), and the other is that the moving distance of the cluster center between two consecutive iterations is less than the preset offset threshold (such as 0.01), indicating that the cluster center has stabilized. Finally, the algorithm outputs a set of operating modes with category labels. Each category label corresponds to a specific operating condition, such as "steady-state high pressure", "load jump", "abnormal vibration", etc., providing a priori knowledge basis for subsequent risk identification models.

[0057] In risk prediction modeling, this implementation first classifies and identifies operating modes under different working conditions based on a complete historical data set to form a basic operating status model with structural labels to support subsequent risk identification. The specific technical process and key calculations involved include the following aspects:

[0058] Working condition feature vector construction

[0059] Key parameters such as temperature, pressure, and vibration are extracted from the processed historical data set. For each period (e.g., every 30 minutes), the system calculates the statistical characteristics of each parameter within that period, including the mean, standard deviation, and extreme values. For a temperature parameter within a given period, the mean is defined as the sum of all sample values ​​divided by the number of samples; the standard deviation is the square root of the average of the squared differences between all data points and the mean; and the maximum and minimum values ​​are directly taken as the extreme values. These statistical values ​​form a feature vector, which serves as input for subsequent cluster analysis.

[0060] Clustering algorithm construction

[0061] The K-means clustering algorithm is used to perform unsupervised learning classification on the extracted feature vectors. The objective function of the K-means algorithm is to minimize the sum of the squares of the distances between all sample points and the center point of their class. Its mathematical expression is:

[0062] "Minimize the sum of the squares of the Euclidean distances between each cluster point and its cluster center";

[0063] The distance between each point and the center point is: the square root of the sum of the squares of all dimensional differences.

[0064] Parameter definition: The multidimensional eigenvalue of each sample point is recorded as a vector; the cluster center point is also in vector form; the number of categories K is determined by methods such as the silhouette coefficient or the elbow rule.

[0065] Calculation of Euclidean distance

[0066] The distance between two sample points is measured using the Euclidean distance, which is the square root of the sum of the squared differences in each dimension. This metric is applicable to data of the same dimension or that has been normalized. This metric reflects the absolute proximity of two points in multidimensional space and is suitable for measuring differences in operating conditions.

[0067] Clustering Iteration and Convergence Criteria

[0068] Cluster center point updates follow the following rule: After each iteration, the average value of all samples in each cluster is recalculated as the new center point, and all samples are then redistributed to the cluster corresponding to the closest center point. The maximum number of iterations is set to, for example, 100. The convergence condition is also set to: the sum of the squares of the distances moved by all cluster centers between two consecutive iterations must be below a preset threshold, for example, less than 0.01. This threshold can be selected by fitting and verifying the initial dataset's center point perturbation experiments.

[0069] Generation of run mode labels

[0070] After clustering, the system assigns unique labels to each cluster, such as "stable high-load operation," "low-load fluctuation," and "vibration over-limit abnormality." These labels can be manually named based on the mean characteristics of the operating conditions represented by the cluster center, or automatically named using an auxiliary recognition model trained based on historical event labels.

[0071] This implementation utilizes clustering-based operating mode recognition to classify complex operating conditions without manual labeling, significantly improving the automation of risk modeling. The Euclidean distance metric and iterative convergence mechanism ensure stable clustering convergence, avoiding classification fluctuations caused by random initialization. The operating mode set, as an intermediate result, provides a structured input basis for the risk prediction model, improving the model's generalization and recognition accuracy. Furthermore, this method can dynamically adapt to new data, supporting continuous iterative updates of the model.

[0072] The risk modeling and processing module obtains an updated risk prediction model based on the processed complete historical data set. It also includes determining the influence of each parameter on mode differentiation through the parameter weight distribution method for the generated operating mode set, combining abnormal mode screening technology to identify abnormal operating modes with a small sample size, and generating an operating mode distribution map as a result visualization output to determine the characteristic parameter combination of the risk operating mode.

[0073] Based on the completion of operating mode clustering, this implementation further identifies the importance of each parameter to mode differentiation through parameter weight allocation technology, and introduces an abnormal mode screening mechanism to assist in identifying low-frequency and high-risk states. By constructing an operating mode distribution map, the mode structure and boundaries are visualized to support risk identification and control strategy formulation.

[0074] Parameter weight allocation method

[0075] This system uses the Fisher discriminant criterion based on the ratio of between-class variance to within-class variance to analyze the discriminant contribution of each parameter. This criterion measures the ability of each parameter to distinguish between different clusters (i.e., operating modes). The weight calculation formula is:

[0076] The weight of each parameter = the average value of the square of the difference of each cluster center on the parameter / the average value of the square of the standard deviation of the parameter in all clusters;

[0077] Among them, the cluster center refers to the average value of the parameter in each category;

[0078] The within-class standard deviation reflects the volatility of the parameter within the category.

[0079] The higher the weight, the more significantly the parameter distinguishes different operating modes. The weighted results are used to generate a weight ranking chart and select the top 20% of parameters as the "key parameter set for mode discrimination."

[0080] Abnormal pattern screening technology

[0081] In all clustering results, the system determines the number of samples in each category and calculates the average number of samples in the total number of categories:

[0082] Average number of samples = total number of samples ÷ number of cluster categories;

[0083] If the number of samples in a certain category is less than 20% of the average value, that is, the number of samples in this category is less than 0.2×the average number of samples, it will be marked as "sparse mode" and enter the anomaly analysis module.

[0084] Furthermore, the system calculates the Mahalanobis distance for the samples within the category to evaluate the multidimensional anomaly degree from the sample point to the center of the category:

[0085] Mahalanobis distance = the weighted square distance of the multidimensional difference between the sample point and the category center point under the category covariance matrix;

[0086] If the Mahalanobis distance of a sample point exceeds the 95th percentile of the Mahalanobis distances of all samples in the category, the system will mark it as a "local outlier"; and use the local outlier factor (LOF) algorithm to further screen for group anomalies based on neighborhood density.

[0087] The LOF value indicates the relative sparseness of the density of a point in its neighborhood. When: the LOF value is > 1.5 (the empirically recommended threshold);

[0088] This point can be considered a "global low-density anomaly," indicating a potential operational risk. This threshold is determined through experimental evaluation. Specifically, the recognition accuracy at different LOF thresholds is tested on historical labeled accident samples, and the optimal F1-score point is selected.

[0089] Operation mode distribution map construction and visualization output

[0090] To visually display the boundaries and overlaps of operating modes, the system uses principal component analysis (PCA) or t-SNE algorithms to compress high-dimensional features into two-dimensional or three-dimensional space:

[0091] PCA calculates the main direction of the feature covariance matrix to achieve the projection in the direction of maximum retained variance;

[0092] t-SNE is more suitable for preserving local structures by constructing neighbor probability distributions and then iteratively optimizing low-dimensional embeddings.

[0093] After embedding, the system displays different cluster categories using color coding, marks local outliers with special graphics (such as red crosses), and highlights the boundaries of the cluster regions containing outliers. The final output is a "running mode distribution map."

[0094] Risk model feature parameter combination extraction

[0095] The system extracts the key parameter value ranges of samples from the identified abnormal category and compares them with the parameter ranges of the normal category to identify parameter combinations with statistically significant differences. The significance is determined based on non-parametric testing methods (such as the Mann–Whitney U test). If the p-value is less than 0.05, the parameter is considered to show a significant deviation in the abnormal category and is included in the "risk signature parameter set."

[0096] The risk modeling and processing module obtains an updated risk prediction model based on the processed complete historical data set. It also includes building a time series-based risk prediction model based on the determined risk operation mode characteristic parameter combination. Specifically, the operation data of the past period is extracted as the training basis through the historical data backtracking method. If the similarity between the current operation parameter combination and the characteristic value of the high-risk mode exceeds the preset threshold, a risk warning is triggered to obtain the probability of risk occurrence and the expected time window of occurrence.

[0097] This implementation builds a time-series-based risk prediction model to dynamically identify impending risk events within operational data, enhancing the system's proactive control capabilities. The implementation process includes: constructing feature parameters, organizing time-series data, defining the model structure, calculating similarity, and implementing a risk triggering mechanism.

[0098] Risk characteristic parameter construction

[0099] The system first identifies the characteristic parameter combinations of "high-risk operating modes" identified in the previous module, which are recorded as risk feature vectors. Assuming that the risk parameters include multi-dimensional indicators such as main steam pressure, reheater outlet temperature, shaft vibration, and main steam valve opening, the system extracts the typical value ranges of these parameters under abnormal operating conditions from historical data to serve as the basis for risk labels in model training.

[0100] Time series sample construction

[0101] When constructing supervised training samples, the system uses a sliding time window approach. Each training sample contains the operating status of T consecutive time steps (called a time slice) and corresponds to a label indicating whether a risk will occur in the future window. For example, for a window length of T = 30 (sampling once per minute, representing the last 30 minutes), if a risk occurs within 30 minutes of the sample starting point, the sample label is 1; otherwise, it is 0.

[0102] The input of a sample can be represented as a two-dimensional tensor: sample = parameter vector at each moment × number of time steps.

[0103] Model structure and training objectives

[0104] An LSTM network is used as the underlying structure of the time series model. The model recursively learns the dynamic evolution of historical states through hidden states. Its output is the probability of a risk occurring in a future time period, recorded as the "predicted risk probability."

[0105] The model training goal is to minimize the cross entropy loss function, which is expressed as: loss function = the negative logarithm sum between the true value of all samples and the model predicted probability value.

[0106] Among them: the true value is 0 or 1; the model prediction value is between 0 and 1, which represents the probability that the current sequence is predicted to be in a risk state.

[0107] Iterative training is performed using the Adam optimizer. The learning rate is usually set to 0.001, the batch size is set to 64, and the training cycle is generally set to 100 rounds.

[0108] Similarity judgment and risk trigger mechanism

[0109] In actual operation, the system extracts a time window of length T from the parameter sequence at each time point and calculates the similarity with the standard sample in the high-risk operation mode. The similarity measurement uses cosine similarity or dynamic time warping (DTW) algorithm:

[0110] Cosine similarity is defined as: the dot product of the current sequence vector and the standard risk sequence divided by the product of their modulo lengths. The value range is [-1, 1]. The larger the value, the closer it is.

[0111] Dynamic time warping measures the minimum nonlinear pairing cost between the current time series and the historical risk series at different time steps, and can handle local misalignment on the time axis.

[0112] When the similarity between the current time series and the risk template is greater than the preset threshold, that is:

[0113] If using cosine similarity: similarity>0.90;

[0114] If DTW is used: distance < the set distance threshold (for example, 30);

[0115] The system believes that the current state is very likely to evolve into a high-risk operating mode and immediately triggers an early warning.

[0116] Basis for determining threshold

[0117] Similarity Threshold: Based on a large number of historical accident samples, the system constructs an operating sequence from 30 minutes to one hour before an accident in the training set. This sequence is then compared with the normal operating state sequence and evaluated using metrics such as F1 score, recall rate, and false alarm rate. Ultimately, the similarity threshold that maximizes the F1 score while achieving the optimal recall rate is selected as the preset standard. For example, in the training set, a cosine similarity greater than 0.88 covers more than 85% of true risk events, while the false alarm rate is less than 10%. Therefore, a practical threshold of 0.90 is selected.

[0118] Risk probability threshold: The predicted probability value output by the model is usually set to 0.5 as the default warning line. It can also be dynamically set to 0.6 or 0.7 based on user needs for hierarchical management and supports hierarchical alarms.

[0119] Output

[0120] When the trigger conditions are met, the system outputs the following information: the predicted risk probability of the current operating status (such as 85%); the time range in which the risk event is most likely to occur (such as the next 10 to 30 minutes); the high-similarity reference template number; the main characteristic parameters involved in the judgment and the comparison of their current values ​​with the standard values.

[0121] The risk modeling and processing module obtains an updated risk prediction model based on the processed complete historical data set. It also includes the triggered risk warning and the expected occurrence time window, and uses a real-time error calculation method to compare the deviation between the predicted risk value and the actual risk value. It also determines the time interval for updating the model parameters through the feedback cycle setting technology, and calculates the matching percentage between the model prediction results and the actual results by the prediction accuracy evaluation method to obtain the performance indicators of the current risk prediction model.

[0122] To ensure the stability and accuracy of the risk prediction model over the long term, this implementation incorporates a performance evaluation and model update mechanism based on actual operational data feedback. This mechanism includes three core components: prediction error calculation, feedback update cycle determination, and model accuracy assessment.

[0123] Forecast error calculation mechanism

[0124] During each risk prediction process, the model outputs two results: one is the probability that the current operating state is high risk, typically ranging from 0 to 1; the other is the predicted time period when the risk is most likely to occur (for example, in the next 10 to 30 minutes). During actual operation, the system continuously monitors whether a risk event actually occurs within this time period.

[0125] To assess the accuracy of predictions, the system records the output probability value of each prediction and compares it with the actual outcome. If the prediction indicates a high probability of a risk, but the risk does not actually occur, it is considered a "false positive"; if the predicted probability is low, but the risk does occur, it is considered a "missed negative."

[0126] The system calculates prediction errors in two ways: one is to calculate the absolute value of the difference between the probability value predicted by the model and the actual result (1 means the risk occurs, 0 means it does not occur); the other is to square the differences of several predictions and then find the average value, that is, to calculate the "mean square error".

[0127] These errors can measure the overall deviation trend of the model over time. For example, if the model consistently overestimates risk, but the actual system is stable, it indicates that the model has systematic errors and needs to be adjusted.

[0128] Model feedback update mechanism

[0129] The system sets up a sliding time window to record the errors of the most recent several (e.g., 100) predictions. After each prediction cycle, the error record for this window is updated and its average is calculated. If the average error of two consecutive sliding windows exceeds a set tolerance threshold (usually 0.2, indicating a prediction deviation exceeding 20%), the model performance may be deteriorating.

[0130] At this point, the system triggers the "model parameter update mechanism." This mechanism doesn't completely retrain the model, but instead employs an incremental update approach, adjusting only high-level prediction layer parameters or the memory portion of the time series model to reduce update costs. The update cycle can be initially set to 48 hours or 100 prediction cycles. If the system continues to perform stably, the cycle is automatically extended; if error fluctuations increase, the cycle is shortened.

[0131] The method for setting the tolerance error threshold is as follows: during the model training phase, the system compares the predicted results of the verification samples with the actual situation, counts the distribution of all errors, and selects the upper limit value within 95% of the sample errors as the error threshold, thereby ensuring that updates are triggered only when there is a large-scale deviation, avoiding frequent disturbances of the model due to individual abnormal points.

[0132] Model prediction accuracy evaluation method

[0133] The system regularly evaluates the model's overall performance. Evaluation metrics include: Precision: How many of the samples the model identified as "likely to occur" actually experienced risk? Recall: How many of the actual risk scenarios were accurately identified by the model? The comprehensive F1 score: A weighted average of the two metrics above, reflecting the model's overall effectiveness.

[0134] The system recalculates these evaluation metrics every 500 predictions or every 24 hours of operation. If the F1 value falls below 0.7 for two consecutive rounds of evaluation, the model's recognition capability is considered to have declined, triggering a "Model Performance Degradation Alert," prompting operations personnel to conduct a manual inspection or the system to automatically initiate a full data retraining process.

[0135] Among them, the sliding window length (such as set to 100): is based on the principle of balancing the stability and timeliness of the evaluation; the average error tolerance threshold (such as 0.2): is selected based on the historical prediction performance distribution and can be adjusted according to the tolerance for false alarms in actual operation; the prediction probability threshold (default 0.5): can be adjusted according to the actual system safety level, such as setting it to 0.4 or 0.3 for important equipment to increase sensitivity; the model update cycle (such as 48 hours): needs to be set comprehensively based on the model calculation cost, prediction frequency and control system response requirements.

[0136] This implementation method, by introducing a real-time error assessment and feedback adjustment mechanism, enables dynamic performance monitoring and adaptive optimization of the risk prediction model during operation, thereby effectively improving the stability and prediction accuracy of the model. By calculating the deviation between the predicted probability and the actual result in real time, the system can quickly identify the trend of declining model performance and adjust the model parameters in a timely manner to avoid the risk of false alarms or omissions due to error accumulation. At the same time, the introduction of a sliding window mechanism and a tolerance error threshold ensures the flexibility and rhythm control of model updates, which not only improves the response sensitivity but also avoids frequent interference with system stability. In addition, the model effect is comprehensively evaluated through multi-dimensional indicators such as accuracy, recall rate and F1 value, providing a quantitative evaluation basis for risk management, making the prediction system more interpretable and a reference value for operation and maintenance decision-making. Overall, this approach enhances the model's adaptability and continuous prediction capabilities under complex operating conditions, effectively supporting the intelligent evolution of the primary pressure control system of thermal power units.

[0137] The risk modeling and processing module obtains an updated risk prediction model based on the processed complete historical data set, which also includes the performance indicators of the current risk prediction model. If the prediction accuracy is lower than the preset threshold, the model parameter weights are dynamically updated through the weight adjustment step technology. At the same time, the warning threshold interval is adjusted in combination with the threshold dynamic range method, and the model is optimized based on the adjustment convergence target to obtain an updated risk prediction model.

[0138] Building on the aforementioned prediction model evaluation, this implementation further addresses scenarios where model performance degrades by constructing a model self-optimization strategy based on dynamic weight adjustment and adaptive threshold adjustment. The core process includes four components: prediction performance assessment, weight step adjustment, dynamic adjustment of warning thresholds, and model convergence optimization.

[0139] Performance indicator judgment mechanism

[0140] The system first calculates the prediction accuracy of the model output in the current period. The calculation formula is:

[0141] Prediction accuracy = number of correctly predicted samples ÷ total number of predicted samples;

[0142] The set accuracy threshold is usually 85%, which means that if the model prediction accuracy is continuously lower than 85%, it is considered that the current model has degraded generalization ability and needs to start the adaptive adjustment process.

[0143] Weight adjustment step size technique

[0144] When the model is judged to have degraded performance, the system activates a dynamic weight adjustment mechanism to slightly adjust key parameters in the model (such as the state transfer weight of the time series layer and the output mapping weight of the fully connected layer).

[0145] The adjustment formula is: new parameter value = old parameter value + current error × dynamic step factor;

[0146] The current error is the difference between the predicted value and the actual label. The initial value of the dynamic step factor is set to 0.01. If the error continues to increase, the step size increases by 10% per round, up to a maximum of 0.1. If the error decreases and stabilizes, the step size decays by 20% per round, with a minimum of 0.001.

[0147] This mechanism can avoid drastic fluctuations in parameters and ensure the gradualness of the update process.

[0148] Warning threshold dynamic range adjustment mechanism

[0149] To adapt to changes in operating conditions, the system introduces a dynamic adjustment strategy for the warning threshold. The initial risk probability warning threshold is set to 0.5, which means that an alarm is triggered when the risk probability output by the model is higher than 50%. The system dynamically adjusts the threshold as follows:

[0150] If the false alarm rate in the recent period is higher than 20%, the threshold will be raised to 0.6 or higher;

[0151] If the false negative rate is higher than 15%, the threshold is lowered to 0.4 or lower;

[0152] The threshold adjustment range is performed in steps of 0.05, and the upper and lower limits of adjustment are 0.3 and 0.9 respectively.

[0153] The false alarm rate is calculated as: the number of samples that incorrectly issue alarms ÷ the total number of alarms;

[0154] The false negative rate is calculated as: the number of risk samples that were not successfully identified ÷ the total number of actual risk samples.

[0155] Through the above adjustments, the system gradually finds the optimal risk boundary suitable for the current data status.

[0156] Convergence target setting and model optimization mechanism

[0157] The system sets the convergence target for model optimization as "F1 value above 0.8 and prediction error below 0.15 for the last three consecutive evaluation cycles." After each weight or threshold adjustment, the system automatically performs a small batch evaluation process and monitors whether the indicators are gradually improving.

[0158] If there is no improvement or the indicators deteriorate for three consecutive rounds, the system will reset to the historical best parameter state and extend the feedback interval to avoid frequent oscillations.

[0159] This implementation method achieves flexible fine-tuning of model parameters by combining prediction accuracy evaluation with a weight adjustment step mechanism, effectively alleviating performance degradation problems caused by changes in operating conditions or data distribution drift. The dynamic step adjustment strategy ensures that model updates are stable and progressive, avoiding overfitting or learning oscillations. At the same time, through the dynamic adjustment mechanism of the warning threshold under the bidirectional control of false alarm rate and missed alarm rate, the model can avoid alarm flooding while maintaining sensitivity, thereby improving the actual use value of the system. Convergence target control and automatic fallback strategy further enhance the robustness and controllability of the model optimization process, thereby achieving continuous improvement in the operational safety and intelligent optimization capabilities of thermal power units while ensuring the stability of risk prediction.

[0160] The specific formula for analyzing the changing trend of the operation mode distribution diagram is:

[0161] ;

[0162] in, Indicates the predicted value of the dynamic change trend of the operating status parameters of the thermal power unit, Indicates the overall sensitivity of the model output to input disturbances, Indicates time The actual disturbance signal value, Indicates time Abnormal concentration, Indicates the current moment, represents the historical integral variable, Represents the base of natural logarithms.

[0163] This implementation method achieves accurate modeling of the changing trends in the operating status of thermal power units by introducing a dynamic integral prediction model. The system is time-based and predicts the dynamic response of state parameters over time by performing a weighted integration of the second-order derivative of the disturbance signal. Specifically, the model defines a time function to represent the predicted value of the unit's operating state parameters. This function consists of two main components: the acceleration term of the disturbance signal, which is the second-order rate of change of the disturbance signal with respect to time, representing the severity of the external input disturbance; and the time attenuation factor, which reflects the concentration of anomalies. The larger this factor, the more likely the system is to be unstable or highly abnormal at the current moment.

[0164] When making predictions, the system first extracts the changes in the disturbance signal within the historical time interval and calculates its second-order derivative with respect to time to reflect the dynamic trend of the change in the disturbance intensity. Subsequently, an exponential decay weight with the degree of abnormal aggregation as the independent variable is introduced to perform weighted integration processing on the above disturbance terms in the time domain. Finally, the integral result is multiplied by an overall sensitivity coefficient, which characterizes the response sensitivity of the model to the input disturbance, thereby obtaining the predicted value of the state parameter change trend corresponding to that time point. In this way, the model not only takes into account the instantaneous intensity of the disturbance, but also comprehensively introduces its time-weighted influence on the overall stability of the system, thereby enhancing the foresight and robustness of the state prediction.

[0165] This implementation method significantly improves the ability to sensitively capture the changing trends of the operating status of thermal power units by constructing an integral prediction model based on the coupling of the change rate of the disturbance signal and the time attenuation function. Compared with the traditional trend assessment method that relies on a fixed time window or static weight, this method can identify changes in the intensity of disturbances in real time and dynamically weight them according to the degree of abnormality to achieve more accurate state prediction. In addition, by introducing the abnormal concentration function and embedding its exponential form into the prediction integral term, the disturbance interference in non-critical time periods is effectively suppressed, and the noise resistance and stability of the prediction are improved. Overall, this method can identify potential unstable trends in advance, guide the subsequent control modules to intervene in a timely manner, and help improve the system operation safety and energy-saving optimization level.

[0166] The overall sensitivity of the model output to input disturbances The calculation formula is:

[0167] ;

[0168] in, Indicates the overall sensitivity of the model output to input disturbances, Indicates the current actual disturbance signal value, Represents the mean of historical disturbance signals.

[0169] This implementation introduces an input disturbance sensitivity factor to quantitatively model the disturbance response strength of thermal power units, enhancing the system's ability to discern fluctuations in actual operating conditions. This sensitivity factor, b, reflects the overall sensitivity of the model output to changes in input disturbances and serves as a crucial foundation for adjusting dynamic prediction accuracy and setting risk assessment thresholds. Its core calculation method is based on the absolute deviation between the current disturbance signal value and the historical average disturbance value, calculated by adding one to the logarithm of this deviation.

[0170] Specifically, the system collects the current disturbance signal value during each prediction cycle, representing the impact of the control disturbance, environmental disturbance, or load change on the unit. At the same time, the system also performs a statistical analysis of the disturbance signal during the past stable period and calculates its mean value as a reference baseline. The greater the deviation between the two, the more significant the current disturbance is compared to the historical state, and the model believes that the impact on the system is greater. At this time, by calculating the logarithm of the deviation plus one, the impact of extreme outliers can be effectively suppressed while maintaining sensitivity to disturbance changes.

[0171] Furthermore, this method exhibits excellent scale-adaptive properties, adapting to both short-term perturbations and rapidly increasing the system's response strength in the event of sudden, severe disturbances. By incorporating this sensitivity coefficient, subsequent models can dynamically adjust weights based on the b value when performing integral predictions, risk assessments, or parameter corrections, thereby improving the overall prediction system's adaptability and accuracy to disturbance patterns.

[0172] By adopting the disturbance sensitivity calculation mechanism proposed in this embodiment, the system can more finely quantify the dynamic deviation between the current operating state and the historical benchmark state, thereby enhancing the context-awareness capability in the risk modeling process. This sensitivity factor can not only serve as a key parameter for the dynamic weighting of the prediction model, but also play a regulator role in the model output credibility assessment, risk threshold fine-tuning, and feedback control response. The logarithmic transformation structure ensures that even in scenarios with drastic disturbance changes, the system can stably output sensitivity factor values ​​within a limited range, effectively preventing excessive oscillations in the model response. At the same time, this method can reduce the amplification effect of abnormal data, enhance the overall noise resistance and stability of the model, and improve the adaptability of the dynamic optimization control system of the initial pressure of thermal power units to complex operating environments.

[0173] In time Abnormal concentration The specific calculation formula is:

[0174] ;

[0175] in, Indicates time Abnormal concentration, Indicates time The actual disturbance signal value, Represents the median of the disturbance values ​​within the past time window.

[0176] This implementation method improves the sensitivity and time adaptability of the disturbance behavior identification of thermal power units by constructing an abnormal concentration measurement function under time variables. The system defines a function for calculating the degree of abnormal aggregation of the disturbance signal at any time point. The function characterizes whether the signal is in an abnormal state based on the deviation amplitude between the current disturbance signal value and its historical median. The specific method is as follows: the system obtains the current disturbance signal value at any time, and at the same time calls the disturbance signal set within a period of time in the forward sliding window at that moment, and calculates the median of the set as a statistical reference benchmark. Subsequently, the system calculates the relative deviation value between the current value and the median, and takes its absolute value as the degree of deviation, and then performs logarithmic compression processing to finally obtain the abnormal concentration at that moment.

[0177] This approach places particular emphasis on the continuous perceptibility of temporal anomalies during modeling, avoiding the overly rigid boundaries of judgment created by simple threshold methods. Through dynamic comparison of medians and nonlinear compression of logarithmic functions, the model effectively suppresses the interference of extreme mutation values ​​on the results while maintaining sensitivity to small changes, thus better conforming to the safety logic characteristic of "gradual anomalies are greater than sudden anomalies" in power systems. Furthermore, this anomaly concentration value will be referenced by subsequent integral prediction models as an exponential decay factor to dynamically attenuate the contribution weight of historical disturbances, enabling the model to strengthen forward warnings during periods of intensive anomalies and enhance robustness during stable phases.

[0178] This implementation significantly enhances the system's ability to identify abnormal trends in the operating state early by introducing an abnormal concentration calculation mechanism based on the deviation of the disturbance signal from the median. Compared with the traditional abnormality identification method using the mean or maximum value, the use of the median as a reference can better resist the interference of outliers and improve the robustness of the judgment. At the same time, by performing a logarithmic transformation on the degree of deviation, the system avoids the sensitivity explosion problem caused by excessive amplitude of outliers, so that the model has better response linear control capabilities. Ultimately, this concentration measure not only directly participates in the trend integral calculation of the operating state, but also provides a reliable basis for risk warning level judgment, control parameter adjustment frequency setting, etc., thereby realizing a safer, smarter and more forward-looking thermal power unit operation control system.

[0179] The standardization method in the data preprocessing module is the Z-Score standardization method.

[0180] In this embodiment, to ensure uniformity and comparability of multi-source heterogeneous data during the modeling process, the data preprocessing module uses the Z-Score normalization method to uniformly transform all raw input parameters. Z-Score normalization, also known as zero-mean normalization, is a statistical processing method that converts data of varying physical quantities or dimensions into a standard normal distribution. Its implementation logic is as follows: for any raw data point, the mean of the sample set to which it belongs is first subtracted, and then divided by the standard deviation of that sample set, thereby converting the data point into a relative value centered at zero and scaled by the unit standard deviation. After this method is applied, all parameters have the same numerical scale and reflect the degree of deviation relative to their sample distribution. This method is particularly suitable for processing typical thermal power unit operating data such as temperature, pressure, flow, and vibration, where the raw dimensions, units, and fluctuation ranges vary significantly. This helps subsequent models accurately identify the relative relationships between different variables, avoids the phenomenon of "value dominance" in the model training process caused by the excessively large numerical range of a single variable, and thus ensures the fairness and scientific nature of the modeling results.

[0181] By introducing the Z-Score standardization method, the system can effectively address issues such as inconsistent dimensionality and large differences in numerical ranges in thermal power unit operating parameter data, improving the comparability of data input and the convergence efficiency of the modeling algorithm. While maintaining the original distribution of the data, this method significantly enhances the model's ability to handle variables of different scales, avoiding over-reliance on certain feature dimensions with large numerical ranges during model training, thereby improving overall modeling accuracy and generalization performance. In addition, the standardized data facilitates rapid alignment with historical standard models, expert rule bases, and the like, facilitating unified standards and rapid deployment in subsequent processing processes such as data fusion, model migration, and anomaly detection, further enhancing the system's compatibility and engineering practicality under complex operating conditions.

[0182] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic optimization control system for initial pressure of thermal power units based on big data, characterized in that: The system comprises: The data preprocessing module is used to obtain multi-dimensional parameter records from the historical operation database of industrial units, and uses standardized methods to unify the dimensions of the original data. At the same time, wavelet transform denoising technology is used to reduce data noise interference. If data is missing, the missing points are supplemented by linear interpolation to obtain a complete historical data set after processing; The risk modeling and processing module is used to obtain an updated risk prediction model based on the processed complete historical data set. For the updated risk prediction model, the latest operating parameter data is obtained and the parameter change rate and fluctuation range are calculated using sliding window technology. If the fluctuation range of the vibration or pressure parameters exceeds the preset safety range, the abnormal pattern screening technology is used again to identify whether the current operating state belongs to the risk mode and determine the current risk level. The dynamic control module is used to dynamically adjust the unit operating parameters based on the current risk level using a real-time control system. The adjustment priority is determined based on the parameter weight distribution results. If the risk probability value continues to drop below the preset safety threshold, the current control parameters are maintained to achieve a stable operating state. The operation optimization recording module is used to continuously record the detailed information of each parameter adjustment for a stable operation state, store the adjustment data through update log recording technology, and analyze the changing trend of the operation mode distribution map to determine the long-term operation mode optimization direction; The risk modeling and processing module obtains an updated risk prediction model based on the processed complete historical data set, including: Based on the processed complete historical data set, a clustering algorithm is used to classify and identify operating modes under different working conditions. Specifically, the characteristic values ​​of multi-dimensional parameters such as temperature, pressure, and vibration are separated from the data through working condition feature extraction technology. The differences between the operating modes are calculated using Euclidean distance as the distance metric. At the same time, the maximum number of iterations and the center point offset threshold are set as iterative convergence conditions to generate a set of operating modes with pattern classification labels. For the generated set of operating modes, the parameter weight distribution method is used to determine the degree of influence of each parameter on mode differentiation. In combination with the abnormal mode screening technology, abnormal operating modes with a small sample size are identified. An operating mode distribution map is generated as a result visualization output to determine the characteristic parameter combination of the risky operating mode. Based on the determined risky operating mode characteristic parameter combination, a time series-based risk prediction model is constructed. Specifically, the historical data backtracking method is used to extract operating data from the past period as the training basis. If the similarity between the current operating parameter combination and the characteristic value of the high-risk mode exceeds the preset threshold, a risk warning is triggered, and the probability of risk occurrence and the expected time window of occurrence are obtained; For the triggered risk warning and the expected occurrence time window, a real-time error calculation method is used to compare the deviation between the predicted risk value and the actual risk value. The time interval for updating the model parameters is determined through the feedback cycle setting technology. At the same time, the matching percentage between the model prediction result and the actual result is calculated by the prediction accuracy evaluation method to obtain the performance index of the current risk prediction model. According to the performance indicators of the current risk prediction model, if the prediction accuracy is lower than the preset threshold, the model parameter weights are dynamically updated through the weight adjustment step technology. At the same time, the warning threshold interval is adjusted in combination with the threshold dynamic range method, and the model is optimized based on the adjustment convergence target to obtain the updated risk prediction model.

2. The big data-based dynamic optimization control system for initial pressure of a thermal power unit according to claim 1, characterized in that: The specific formula for analyzing the changing trend of the operation mode distribution diagram is: ； in, Indicates the predicted value of the dynamic change trend of the operating status parameters of the thermal power unit, Indicates the overall sensitivity of the model output to input disturbances, Indicates time The actual disturbance signal value, Indicates time Abnormal concentration, Indicates the current moment, represents the historical integral variable, Represents the base of natural logarithms.

3. The big data-based dynamic optimization control system for initial pressure of a thermal power unit according to claim 2, characterized in that: The overall sensitivity of the model output to input disturbances The calculation formula is: ; in, Indicates the overall sensitivity of the model output to input disturbances, Indicates the current actual disturbance signal value, Represents the mean of historical disturbance signals.

4. The big data-based dynamic optimization control system for initial pressure of a thermal power unit according to claim 2, characterized in that: Said in time Abnormal concentration The specific calculation formula is: ； in, Indicates time Abnormal concentration, Indicates time The actual disturbance signal value, Represents the median of the disturbance values ​​within the past time window.

5. The big data-based dynamic optimization control system for initial pressure of a thermal power unit according to claim 1, characterized in that: The standardization method in the data preprocessing module is the Z-Score standardization method.

Images