Operation monitoring method and system for deviation early warning on thermal power monitoring disc

Abstract

The invention discloses an operation monitoring method and system for deviation early warning on a thermal power monitoring disc, and relates to the technical field of thermal power operation abnormity monitoring and deviation calculation, and the method comprises the steps: collecting and processing the historical operation data of a thermal power generating unit from an SIS system and a DCS system, and extracting stable working condition data according to a load stability rule; and constructing a prediction algorithm based on a Gaussian mixture model, setting a model training and verification standard, and generating a model operator for deviation identification. And performing deviation monitoring on the real-time measurement point value based on a model operator, and generating early warning information according to a threshold rule. According to the method, by constructing the prediction algorithm based on the Gaussian mixture model, label-free deviation recognition of the key measuring points of the thermal power monitoring disc is achieved, the accuracy, the interpretability and the practicability of anomaly detection are effectively improved by combining steady-state data screening, feature correlation modeling and a multi-condition triggering early warning mechanism, the false alarm rate is reduced, and the accuracy, the interpretability and the practicability of the thermal power monitoring disc are improved. And the method has good engineering deployability.

Description

Technical Field

[0001] The present invention relates to the technical field of thermal power operation abnormality monitoring and deviation calculation, and in particular to an operation monitoring method and system for deviation early warning on a thermal power monitoring panel. Background Art

[0002] With the continuous improvement of the automation level of thermal power plants, distributed control systems (DCS) and safety instrument systems (SIS) have been fully deployed in the operation of power plants, forming a centralized monitoring system based on parameter collection, data feedback, and operation interlocking. In this context, in order to cope with the challenges of increasing equipment complexity and dynamic fluctuations in operating loads, more and more power plants have begun to introduce data-driven operating status identification methods, including measurement point anomaly detection methods based on time series analysis, statistical learning, and artificial intelligence models. Especially in the field of early identification of abnormal states, the coupling behavior between high-dimensional monitoring data and nonlinear characteristics has become a research focus, and deviation detection through model prediction residuals has become the current mainstream idea. However, this type of method still faces the trade-off between real-time performance, accuracy, and engineering interpretability in practical applications, making it difficult to directly adapt to multi-operating thermal power scenarios.

[0003] Existing deviation detection methods are primarily categorized into two types: static warning schemes based on empirical thresholds and intelligent prediction methods based on black-box models. The former employs manually set upper and lower limits to generate real-time alarms, but this strategy struggles to adapt to fluctuations in measurement points caused by operational state changes, and its fixed nature leads to numerous false alarms and missed alarms. While methods such as neural networks and support vector machines offer some predictive capabilities, they lack physical mechanism connections, resulting in uninterpretable model results and a reliance on large numbers of labeled samples for training, limiting their applicability in engineering deployments. More critically, existing methods generally lack an end-to-end closed-loop process encompassing steady-state segment extraction, feature selection, model training, dynamic deviation monitoring, and state verification. This leads to a disconnect between model training and warning execution, hindering effective comprehensive assessment of complex deviation events such as sudden anomalies, slope shifts, and sustained instability in thermal power systems. Furthermore, existing technologies often lack standardized mechanisms for model parameter selection, training strategies, and the establishment of warning logic rules. These methods fail to incorporate the physical response patterns of equipment measurement points, provide a clear basis for interpretation, and provide a reliable model credibility assessment, thus reducing the controllability and verification capabilities of the warning system. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problems solved by the present invention are: the existing deviation identification and early warning methods of thermal power monitoring systems rely on static rules, have a high false alarm rate, lack a dynamic working condition modeling mechanism, and the early warning judgment process is separated from model training. In addition, the present invention also solves the problem of how to achieve highly interpretable real-time early warning of key measurement point deviations under label-free conditions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solutions: a method for monitoring the operation of a thermal power plant with deviation warning on a monitoring panel, comprising collecting and processing historical operating data of thermal power units from the SIS system and the DCS system, and extracting stable operating condition data according to load stability rules. A prediction algorithm based on a Gaussian mixture model is constructed, model training and verification standards are set, and a model operator for deviation identification is generated. Deviation monitoring is performed on real-time measurement point values ​​based on the model operator, and warning information is generated according to threshold rules. Constructing a prediction algorithm based on a Gaussian mixture model includes constructing a prediction algorithm based on a Gaussian mixture model for each target measurement point, constructing a feature vector, performing fitting training on the input feature vector, setting the initial number of clusters, and determining the final number of clusters.

[0007] As a preferred embodiment of the method for monitoring the operation of a thermal power plant monitoring panel using deviation warnings described herein, the acquisition and processing of historical thermal power unit operation data from the SIS and DCS systems includes extracting historical operating point data through the power plant's DCS and SIS interfaces, recording data containing timestamps, equipment identifiers, measurement point types, and original values, and performing timeline padding on the imported data. Field standardization is performed on all measurement point fields, unifying the unit format and establishing a corresponding relationship with the equipment structure tree. Time periods with data anomalies during continuous operation are eliminated, including equipment maintenance periods, null values, and failure identification periods.

[0008] As a preferred solution of the operation monitoring method of the deviation warning on the thermal power monitoring panel described in the present invention, the extraction of stable operating condition data according to the load stability rule includes evaluating the load change per unit time for each time period, setting a load stability standard, and each data must simultaneously meet the load stability standard for all key measurement point values, as a training data set.

[0009] As a preferred embodiment of the method for monitoring the operation of the deviation warning on the thermal power monitoring panel described in the present invention, the method comprises: constructing a prediction algorithm based on the Gaussian mixture model for each target measuring point, selecting the top five input parameter sets according to the ranking of the Pearson correlation coefficient with the target variable to construct a feature vector. The input feature vector is fitted and trained using the expectation maximization algorithm, the initial number of clusters is set, and the final number of clusters is determined using the Bayesian Information Criterion and the Akaike Information Criterion. The model covariance matrix type is selected as a diagonal type.

[0010] As a preferred embodiment of the method for monitoring the operation of a thermal power plant monitoring panel using deviation warnings described in the present invention, the prediction algorithm based on the Gaussian mixture model includes performing Z-score normalization on all input features, setting the initial number of clusters k to a range of 4 to 8, and calculating the Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC) indicators for each clustering result. The two indicators comprehensively evaluate convergence stability and generalization ability, and ultimately select the k value corresponding to the minimum comprehensive score as the final number of clusters. The Gaussian distribution covariance structure in the model is fixed as a diagonal matrix, the training termination condition is set to a maximum number of iterations of 300, and the maximum deviation rate in each training round is set to 2%.

[0011] As a preferred solution of the operation monitoring method of the deviation warning on the thermal power monitoring panel described in the present invention, the generation of the model operator for deviation identification includes using five-fold cross-validation for each training result and using the determination coefficient as the model validity criterion. If the determination coefficient score is greater than 0.95, it is a valid model operator.

[0012] As a preferred solution of the operation monitoring method of the deviation warning on the thermal power monitoring panel described in the present invention, the deviation monitoring of the real-time measurement point value based on the model operator includes: accessing the real-time operation data stream in the thermal power monitoring panel platform, and performing a difference calculation between the real-time value of the target measurement point and the predicted value calculated by the corresponding model operator, and the difference is used for deviation trigger judgment. The deviation trigger judgment includes the following rules: judging whether the difference exceeds the upper and lower limit interval boundaries of the historical deviation, which is the 95% confidence interval of the difference distribution during the model training period. Judging whether the difference over-limit state lasts for more than 30 seconds. Judging whether the change rate of the current target measurement point exceeds the dynamic slope threshold. When all judgments are met, it is confirmed as a valid deviation trigger and an early warning is generated.

[0013] As a preferred solution of the operation monitoring method of the deviation warning on the thermal power monitoring panel described in the present invention, the deviation trigger judgment includes introducing a steady-state data filtering mechanism before executing the difference threshold judgment, and performing a linkage check on the sliding mean and sliding variance of the target measuring point within the past 60 seconds. If the difference between the mean and the predicted value of the target measuring point within 60 seconds is continuously less than 80% of the preset threshold, and the variance change rate is less than 5%, the deviation judgment stage is entered.

[0014] As a preferred solution of the operation monitoring method of the deviation warning on the thermal power monitoring panel described in the present invention, the generation of warning information according to the threshold rule includes setting the state tracking logic after the deviation judgment of each measuring point, which is divided into three state nodes: normal, abnormal confirmation, and warned. When the duration of any deviation reaches the threshold, it jumps from normal to abnormal confirmation, and enters the warned state if it has not recovered within 30 seconds. The timestamp, measuring point ID, predicted value, actual value and deviation value are recorded for each state change. The generated warning information includes the unique alarm number, time, measuring point name, parameter value and trigger condition, which are synchronously displayed on the homepage of the monitoring platform.

[0015] Another object of the present invention is to provide an operation monitoring system for deviation warning on a thermal power monitoring panel, which can construct a prediction algorithm module for constructing a prediction algorithm based on a Gaussian mixture model, set model training and verification standards, and generate a model operator for deviation identification, thereby solving the problems of existing thermal power monitoring system deviation identification and warning methods that rely on static rules, have a high false alarm rate, lack a dynamic working condition modeling mechanism, and are separated from the warning judgment process and model training, as well as how to achieve highly interpretable real-time warning of key measurement point deviations under unlabeled conditions.

[0016] As a preferred solution of the operation monitoring system of the deviation early warning on the thermal power monitoring panel described in the present invention, it includes: a data collection and processing module, a prediction algorithm construction module, and a deviation monitoring module.

[0017] The data acquisition and processing module is used to acquire and process historical operating data of thermal power units from the SIS system and the DCS system, and extract stable operating condition data according to load stability rules.

[0018] The prediction algorithm construction module is used to construct a prediction algorithm based on a Gaussian mixture model, set model training and verification standards, and generate a model operator for deviation identification.

[0019] The deviation monitoring module is used to monitor the deviation of the real-time measurement point values ​​based on the model operator and generate early warning information according to the threshold rule.

[0020] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a step of an operation monitoring method for deviation early warning on a thermal power monitoring panel.

[0021] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for monitoring the operation of a deviation warning on a thermal power monitoring panel.

[0022] The beneficial effects of the present invention are as follows: the operation monitoring method for deviation warning on the thermal power monitoring panel provided by the present invention collects and processes historical data, realizes the construction of high-quality stable operating condition data, ensures the integrity, consistency and representativeness of the modeling input data, avoids the misjudgment or overfitting problems caused by dirty data or unstable operating conditions in model training, significantly improves the accuracy and generalization ability of subsequent models, and solves the problem that the existing methods cannot automatically filter abnormal samples and operating condition drift.

[0023] The GMM-based prediction model achieves high adaptability and interpretability in deviation identification. By constructing a highly targeted input set through statistical correlation, the model structure is tightly coupled with the characteristics of the measurement points, improving prediction accuracy. The diagonal covariance structure reduces training computational complexity and improves model stability. The dual-metric evaluation mechanism avoids the risk of overfitting, significantly outperforming traditional neural network-based black-box algorithms in interpretability and deployment-friendliness in power industry scenarios. This addresses the challenges of complex models, difficulty in debugging, and implementation in existing technologies.

[0024] The construction of multi-condition deviation trigger logic, achieving false alarm rate control and status visualization early warning mechanisms, reduces false alarm and misjudgment rates and improves the ability to identify sudden or trending anomalies. The multi-factor fusion judgment logic fills the gap in the existing system's lack of a sensitive judgment mechanism for the combined changes of time, intensity, and rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0026] Figure 1 This is an overall flow chart of a method for monitoring the operation of a thermal power plant monitoring panel with deviation warning provided in the first embodiment of the present invention.

[0027] Figure 2 This is a trend analysis diagram of an operation monitoring method for deviation early warning on a thermal power monitoring panel provided by the second embodiment of the present invention.

[0028] Figure 3 This is an overall schematic diagram of an operation monitoring system for deviation warning on a thermal power monitoring panel provided by a third embodiment of the present invention. DETAILED DESCRIPTION

[0029] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0030] Example 1, reference Figure 1 , as one embodiment of the present invention, provides a method for monitoring the operation of a thermal power plant monitoring panel with a deviation warning, comprising:

[0031] S1: Collect and process historical operating data of thermal power units from the SIS system and DCS system, and extract stable operating condition data based on load stability rules.

[0032] Extract historical records of operating measurement point data through the power plant's DCS and SIS interfaces, including timestamps, equipment identifiers, measurement point types, and original values. Timeline padding is then performed on the imported data. Standardize all measurement point fields, unify unit formats, and establish correspondence with the equipment structure tree. Periods of continuous operation with data anomalies are eliminated, including periods of equipment maintenance, null values, and failure identification.

[0033] It should be noted that the time axis of the imported data is padded with a minimum sampling period of 30 seconds, that is, if the interval between two adjacent occurrences of a certain measuring point is greater than 30 seconds, the system automatically interpolates to fill in the gaps, and the interpolation method is linear interpolation. If the interval between adjacent sampling points is less than 30 seconds, the first value is taken as the standard value. Field standardization operations are performed on all measuring points, including unified unit conversion (such as pressure is unified as MPa, temperature is unified as ℃), and each measuring point is associated and bound with its corresponding device node according to the pre-set device structure tree to form a device-measuring point mapping relationship table. In order to eliminate abnormal data, the system will identify all records marked as equipment maintenance period, measuring point failure (such as sensor loss of connection), value Null or exceeding the upper and lower technical limits of the measuring point as abnormal segments and clear them.

[0034] For each time period, the load change per unit time is evaluated and the load stability standard is set. Each data set must simultaneously meet the load stability standard for all key measurement point values ​​and serve as a training data set.

[0035] A preferred solution of the load stability standard specifically includes defining the load stability judgment conditions as follows:

[0036] The selected load segment must be one of the set operating points. The set typical operating points include 150MW, 170MW, 200MW, 220MW, 270MW, 300MW, and 320MW.

[0037] During the time period to be evaluated, the load variation range of the main steam generator set (i.e., the real-time grid power output) must meet the ±1% rated load condition, where the rated load is defined as the system-calibrated 300MW.

[0038] The stability assessment window is 5 consecutive minutes, and the fluctuation of the main load measurement point value within this time window shall not exceed ±3MW.

[0039] At the same time, all operating parameters defined as key measuring points (main steam temperature, main steam pressure, reheat steam temperature, furnace negative pressure, boiler water level) must be maintained within their respective set historical normal fluctuation ranges during this period. The fluctuation range is the 90% confidence interval of the measuring point under the corresponding operating conditions.

[0040] Only time periods that meet both the aforementioned load fluctuation limits and the critical measurement point interval determination can be considered valid stable operating condition data. The system automatically extracts these stable data segments, constructs a model training dataset, and stores them by operating point, providing a comprehensive input foundation with reliable data quality for subsequent Gaussian mixture model training.

[0041] It should be noted that the design concept of S1 is to structure and screen the original historical operating data with high quality through unified sampling period, field structure, anomaly elimination and load stability identification rules, to ensure that the data used for model training has integrity, consistency and engineering stability.

[0042] S2: Build a prediction algorithm based on the Gaussian mixture model, set model training and verification standards, and generate model operators for deviation identification.

[0043] A prediction algorithm based on the Gaussian mixture model is constructed for each target measuring point, and the top five input parameter sets are selected according to the Pearson correlation coefficient ranking with the target variable to construct the feature vector.

[0044] For each target measuring point Y (t) , extract relevant input features from the constructed stable working condition training data set. By calculating the target variable Y (t) With other collectible variables Pearson correlation coefficient between Filter out the top five feature variables with the highest correlation with the target variable:

[0045]

[0046] Among them, Y (t) Indicates the value of the target measuring point at time t. Represents the value of the i-th characteristic variable at time t. σ Y is the standard deviation of Y, For Xi Cov(·) represents the covariance function. N represents the total number of eigenvectors.

[0047] The selected 5 feature variables constitute the input vector

[0048] Z-score normalization is performed on all input features, and the initial number of clusters k is set to traverse the range of 4 to 8.

[0049] Performing Z-score normalization on all input features is expressed as:

[0050]

[0051] in, Represents the features after normalization. Represents the jth original input feature. μ j Represents the mean of the jth feature. σ j represents the standard deviation of the j-th feature.

[0052] The normalized input data is fed into the GMM model to construct the joint probability density function. The GMM structure is as follows:

[0053]

[0054] in, Represents the GMM modeling structure. M represents the number of Gaussian components, and the initial traversal is set to M∈{4,5,6,7,8}. Represents the input features after normalization. m Represents the weight coefficient of the mth Gaussian component. μ m Represents the mean vector of the mth Gaussian component. ∑ m Represents the covariance matrix of the mth Gaussian component. m represents the mth Gaussian component.

[0055] Furthermore, μ m Expressed as:

[0056]

[0057] ∑ m Expressed as:

[0058]

[0059] Among them, each component corresponds to a typical operating condition feature combination. s Indicates the total number of stable working condition samples used to train the model. The sample data is extracted from historical operating data after load fluctuation judgment and measurement point interval screening, ensuring data integrity and stability.n,m represents the posterior probability that the nth historical sample is generated by the mth Gaussian component during the current modeling process. diag(·) extracts each dimension of the sample variance vector as a diagonal element, forming a simplified form of the covariance matrix. This method retains only the variance of each measurement point's features and ignores any coordinated fluctuations between different measurement points (for example, the coordinated fluctuations between boiler water level and reheat steam temperature are not modeled). This improves model training efficiency and convergence stability.

[0060] Model training uses an iterative optimization algorithm called Expectation-Maximization (EM). The number of clusters, M, is randomly set, and the mean vector, covariance matrix, and mixing weights of each Gaussian component are initialized. The probability that each training sample belongs to each Gaussian component is calculated as the posterior probability. This probability represents the likelihood that a sample was generated by that component. Using the calculated responsibility, the parameters of each component are updated, including its mean vector, covariance matrix, and mixing weights. The update rule is to maximize the log-likelihood under the current responsibility. This algorithm is repeated until the model reaches 300 iterations. The final output is the optimal parameter estimate.

[0061] By calculating the Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC) corresponding to each group of clustering results, the two indicators are used to comprehensively evaluate the convergence stability and generalization ability, and finally the k value corresponding to the minimum comprehensive score is selected as the final number of clusters.

[0062] For each traversed Gaussian component number M, calculate the BIC (Bayesian Information Criterion) and AIC (Akaike Information Criterion) corresponding to the model:

[0063] BIC M =-2·logL M +k M ·logN'

[0064] AIC M =-2·logL M +2k M

[0065] Among them, BIC M The Bayesian Information Criterion (AIC) represents the number of Gaussian components. M Akaike Information Criterion representing the number of the Mth Gaussian component. L M Indicates the maximum likelihood function value when the number of clusters is M. k M Indicates the number of free parameters of the current model. N' indicates the total number of training samples.

[0066] Finally, BIC M and AIC M The 1:1 weighted comprehensive score is used as the evaluation index, and the M corresponding to the minimum comprehensive score is taken. *As the final number of clusters. * This strategy strikes a balance between structural complexity control and fitting ability, improving the model's generalization ability on non-training samples.

[0067] Furthermore, the Gaussian distribution covariance structure in the model is fixed to a diagonal matrix, the training termination condition is set to a maximum number of iterations of 300, and the maximum deviation rate in each round of training is set to 2%.

[0068] A five-fold cross-validation was used for each training result, and the coefficient of determination was used as the model validity criterion. If the coefficient of determination score was greater than 0.95, the model operator was considered valid:

[0069] After the model training is completed, the prediction ability is evaluated using five-fold cross validation. The coefficient of determination R is defined as 2 As a validation indicator:

[0070]

[0071] in, Represents the model prediction value. represents the mean value of the target variable in the sample. T represents the number of samples.

[0072] It should be noted that the core of S2's design is to build a deviation prediction model with strong interpretability and stable structure for thermal power monitoring points. The multi-dimensional joint distribution of historical stable operating condition data is modeled through the Gaussian mixture model (GMM), and the input variables are screened in combination with the Pearson coefficient to ensure that each prediction model is targeted and robust. Compared with traditional black box models based on neural networks or support vector machines, S2 uses diagonal covariance structure, BIC / AIC dual-criteria clustering selection, Z-score standardization and R 2 The verification mechanism realizes the unity of structural optimization selection and result controllability under data drive, which is particularly suitable for operation data modeling scenarios with high coupling, nonlinearity and strong noise characteristics in thermal power systems. It is significantly superior to existing technologies in terms of model transparency, stability and actual deployability.

[0073] S3: Deviation monitoring of real-time measurement point values ​​is performed based on model operators, and early warning information is generated according to threshold rules.

[0074] The real-time data stream is connected to the thermal power monitoring platform. The difference between the real-time value of the target measurement point and the predicted value calculated by the corresponding model operator is calculated. The difference is used to trigger the deviation judgment. Before executing the difference threshold judgment, a steady-state data filtering mechanism is introduced. The sliding mean and sliding variance of the target measurement point over the past 60 seconds are linked and verified. If the difference between the mean and the predicted value of the target measurement point is continuously less than 80% of the preset threshold within 60 seconds, and the variance change rate is less than 5%, the deviation judgment phase is entered.

[0075] Deviation trigger judgment includes the following rules:

[0076] Determine whether the difference exceeds the upper and lower limits of the historical deviation interval, which is the 95% confidence interval of the difference distribution during the model training period.

[0077] Determine whether the difference exceeds the limit state for more than 30 seconds.

[0078] Determine whether the change rate of the current target measuring point exceeds the dynamic slope threshold.

[0079] When all judgments are met, it is confirmed as a valid deviation trigger and an early warning is generated.

[0080] Furthermore, a preferred embodiment of the dynamic slope threshold specifically includes: In the present invention, the dynamic slope threshold is used as a criterion for determining whether a target measurement point has a sudden change in behavior. It is defined as the upper limit of the rate of change of the measurement point value per unit time. To adapt to the actual characteristics of different measurement points, the dynamic slope threshold adopts a fixed reference value based on historical data statistics. The setting principle is as follows:

[0081] For pressure measuring points (such as main steam pressure), the dynamic slope threshold is set to 0.1 MPa / min.

[0082] For temperature measurement points (such as reheat steam temperature), the dynamic slope threshold is set to 3.0℃ / min.

[0083] For flow measurement points (such as water supply flow), the threshold is set to 5.0t / h / min.

[0084] If it is a nonlinear response type measuring point, the dynamic threshold is set based on 80% of the maximum change rate of its historical stable operating range.

[0085] After determining deviations at each measurement point, status tracking logic is implemented, categorizing the status into three nodes: normal, confirmed abnormality, and warned. When any deviation persists beyond a threshold, the system transitions from normal to confirmed abnormality. If the system persists for 30 seconds and still does not recover, it enters the warned state. Each state change records the timestamp, measurement point ID, predicted value, actual value, and deviation. Generated warning information includes a unique alarm number, time, measurement point name, parameter value, and trigger condition, and is displayed synchronously on the monitoring platform homepage.

[0086] It should be noted that the core of the S3 step design is to build a real-time deviation trigger mechanism that integrates multiple conditions. By introducing the triple logic of sliding window steady-state filtering, difference threshold judgment, duration judgment, and dynamic slope judgment, high-precision identification of abnormal states of thermal power monitoring points can be achieved. Different from the traditional strategy that relies only on static upper and lower limits or a single alarm value, S3 combines the historical deviation distribution confidence interval and change rate characteristics for the first time to establish a dynamic multi-factor judgment system, which significantly reduces the false alarm rate and missed alarm rate. At the same time, a state machine tracking model is used to achieve explicit modeling of the alarm state evolution process, enhancing the system's interpretability and visualization capabilities. This strategy is particularly suitable for monitoring key measurement points in complex multi-variable systems of thermal power plants, breaking through the limitation of existing technologies that cannot integrate time series volatility and structural offset behavior.

[0087] Example 2, as Figure 2 As shown, an embodiment of the present invention provides an operation monitoring method for deviation warning on a thermal power monitoring panel. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0088] After commissioning the deviation warning module in a coal-fired power plant's monitoring system, it detected 504 large deviation warnings within a month. These included a variety of abnormalities, such as large differential pressures at the inlet and outlet of the coal machine, large deviations in secondary air pressure at the inlet of the air preheater, and large deviations in flue gas pressure at the outlet of fan A, all common sudden faults. Of these, 467 abnormal deviations were confirmed as valid by on-site operators, achieving an accuracy rate of over 92%. Furthermore, the multi-parameter trained Gaussian mixture model detected problems faster than DCS threshold alarms, significantly assisting operators in monitoring the system.

[0089] At 18:04 on May 20, 2024, the monitoring system's deviation warning module detected a significant pressure deviation at the inlet and outlet of Grind C, issuing an early warning. After confirmation by the operator, the defect was registered in the defect management module for closed-loop resolution.

[0090] Around 5:50 PM on May 20, 2024, the coal flow at feeder 1C slowly increased, and the differential pressure at mill C's inlet and outlet increased accordingly. The differential pressure at mill 1C remained around 3.34, while the predicted differential pressure fluctuated around 3.16. For several minutes thereafter, the difference between the predicted differential pressure at mill C's inlet and outlet remained greater than 0.1, triggering an early warning. The coal flow at mill C was reduced on-site, and the differential pressure subsequently decreased, eliminating the early warning.

[0091] When modeling the deviation warning model for the large deviation warning of the differential pressure at the inlet and outlet of the C mill, 29,494 data records of 25 relevant measurement points such as the C mill current, the coal feed rate feedback of the coal feeder C, the powder pipe pressure, the wind speed, the differential pressure, etc. under various stable load conditions were imported. After data merging and value range cleaning, the data was imported into the Gaussian mixture model with a test set ratio of 0.2 and then trained. After training, the AI ​​operator's score prediction for the C mill current had a root mean square error of 0.2532 and an R2 score of 0.9789. The training results were deemed qualified and published in the C coal machine deviation warning model. In the warning model, the warning value is set according to the difference between the predicted value and the actual value of the differential pressure at the inlet and outlet of the C mill. Here, the upper and lower limits of the deviation are set to 0.1. The model modeling is completed through the above mechanism rules. This model has verified the correctness of the alarm many times and discovered anomalies 2-3 minutes earlier than manual detection.

[0092] Example 3, reference Figure 3 , which is an embodiment of the present invention, provides an operation monitoring system for deviation warning on a thermal power monitoring panel, including a data collection and processing module 100, a prediction algorithm construction module 200, and a deviation monitoring module 300.

[0093] S4: The data collection and processing module 100 is used to collect and process the historical operating data of the thermal power units from the SIS system and the DCS system, and extract stable operating condition data according to the load stability rule.

[0094] The data collection and processing module 100 includes a data extraction and normalization submodule 101 and a steady-state data screening submodule 102 .

[0095] Furthermore, the data extraction and standardization submodule 101 is used to batch extract historical measurement point data from thermal power units through the DCS and SIS interfaces, including timestamps, device identifiers, measurement point types, and original values. It then performs field standardization on the collected data, unifying the unit format for each measurement point. It also generates a device-to-measurement point mapping table based on the device structure tree, enabling structured and unified management of measurement point information. Simultaneously, data is padded at 30-second intervals on the timeline, linear interpolation is used to address missing segments, and null values, failure markers, and sampling anomaly records are eliminated.

[0096] The steady-state data screening submodule 102 is used to extract stable operating condition data segments based on load stability rules. The load fluctuation range is set to no more than ±1% of the rated load. The module then assesses whether key measurement points (such as main steam temperature and main steam pressure) within a 5-minute sliding window are within their historical 90% confidence intervals. Time periods that meet all constraints are marked as stable segments and used as input for subsequent modeling.

[0097] It should be noted that the data extraction and standardization submodule 101 is the entry point for the entire system operation, and its structured output directly serves as the input basis for the steady-state data screening submodule 102. The two work together to form the data support platform for the model training stage.

[0098] S5: Constructing a prediction algorithm module 200 is used to construct a prediction algorithm based on a Gaussian mixture model, set model training and verification standards, and generate a model operator for deviation identification.

[0099] The prediction algorithm module 200 includes a feature construction and standardization submodule 201 and a model training and selection submodule 202 .

[0100] Furthermore, the feature construction and normalization submodule 201 is used to select the top five most correlated features based on the Pearson correlation between the target measurement point and other available variables to form an input vector. Z-score normalization is performed on all feature data to ensure that the data in each dimension has a uniform dimension and distribution, adapting to the input requirements of Gaussian mixture modeling.

[0101] The model training and selection submodule 202 is used to traverse different initial cluster numbers and perform the expectation maximization (EM) algorithm to fit the Gaussian mixture model in each training set. The AIC and BIC indicators are calculated respectively and the cluster number corresponding to the minimum comprehensive score is taken as the optimal structure. The covariance matrix is ​​set to diagonal type and the maximum number of iterations is 300. Finally, the five-fold cross validation and the determination coefficient (R 2 )Evaluate the model effectiveness and output high-precision prediction model operators.

[0102] It should be noted that the feature construction and standardization submodule 201 provides structural input vectors for model training, and the model training and selection submodule 202 determines the effective model operators that can ultimately be used for deviation judgment. The two together complete the structural generation and performance constraints of the algorithm construction process.

[0103] S6: The deviation monitoring module 300 is used to monitor the deviation of the real-time measurement point values ​​based on the model operator and generate early warning information according to the threshold rule.

[0104] The deviation monitoring module 300 includes a real-time deviation judgment submodule 301 and a status tracking and early warning submodule 302 .

[0105] Furthermore, the real-time deviation judgment submodule 301 is used to calculate the difference between the real-time measurement point value in the monitoring platform and the corresponding model operator prediction value. By introducing a 60-second sliding window, it first determines whether the target measurement point is in the steady-state range (the mean difference is continuously below the 80% confidence threshold, and the variance change is less than 5%). Then, in the steady-state state, a triple deviation judgment is performed: whether it exceeds the 95% confidence limit, whether the limit is exceeded for 30 seconds, and whether the rate of change exceeds the dynamic slope threshold. If all three conditions are met, it is considered a valid deviation trigger.

[0106] The status tracking and warning submodule 302 establishes a three-state model (normal, confirmed abnormality, and warned). It records each deviation judgment result, along with its trigger time, measurement point ID, predicted value, real-time value, and deviation value. If an abnormality persists beyond a certain time limit, the system enters the "warned" state and generates a uniquely numbered warning message, which is pushed to the main interface of the monitoring platform to support subsequent traceback and confirmation.

[0107] It should be noted that the real-time deviation judgment submodule 301 and the state tracking and warning submodule 302 jointly construct an interpretable and controllable warning execution chain, realizing multi-layer judgment and real-time response to the model output, and ensuring the stability and traceability of the system operation.

Claims

1. A method for monitoring the operation of a thermal power plant monitoring panel with a deviation warning, characterized in that: include: Collect and process historical operating data of thermal power units from SIS and DCS systems, and extract stable operating condition data based on load stability rules; Build a prediction algorithm based on a Gaussian mixture model, set model training and verification standards, and generate model operators for deviation identification; Monitor deviations of real-time measurement point values ​​based on model operators and generate early warning information based on threshold rules; Constructing a prediction algorithm based on a Gaussian mixture model includes constructing a prediction algorithm based on a Gaussian mixture model for each target measuring point, constructing a feature vector, performing fitting training on the input feature vector, setting an initial number of clusters, and determining a final number of clusters.

2. The method for monitoring the operation of a thermal power plant monitoring panel with a deviation warning according to claim 1 is characterized in that: The acquisition and processing of historical operating data of thermal power units from the SIS system and the DCS system includes: Extract historical records of operating measurement point data through the DCS and SIS interfaces of thermal power plants. The records include timestamps, equipment identifiers, measurement point types, and original values, and perform timeline completion on the imported data. Perform field standardization operations on all measurement point fields, unify the unit format and establish a corresponding relationship with the equipment structure tree. At the same time, eliminate time periods with data anomalies during continuous operation, including equipment maintenance periods, null values, and failure identification segments.

3. The method for monitoring the operation of a thermal power plant monitoring panel with a deviation warning according to claim 1 or 2, characterized in that: The extracting of stable working condition data according to load stability rules includes: For each time period, the load change per unit time is evaluated and the load stability standard is set. Each data set must simultaneously meet the load stability standard for all key measurement point values ​​and serve as a training data set.

4. The method for monitoring the operation of a thermal power plant monitoring panel with a deviation warning according to claim 3 is characterized in that: The prediction algorithm based on the Gaussian mixture model is constructed as follows: A prediction algorithm based on a Gaussian mixture model is constructed for each target measurement point. The top five input parameter sets are selected according to the Pearson correlation coefficient ranking with the target variable to construct a feature vector. The input feature vector is fitted and trained using the expectation maximization algorithm, the initial number of clusters is set, and the final number of clusters is determined by the Bayesian Information Criterion and the Akaike Information Criterion; The model covariance matrix type is selected as diagonal.

5. The method for monitoring the operation of a thermal power plant monitoring panel with a deviation warning according to claim 4 is characterized in that: The prediction algorithm based on Gaussian mixture model includes: All input features are Z-score normalized, and the initial number of clusters k is set to 4 to 8. The Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC) corresponding to each clustering result are calculated separately. The convergence stability and generalization ability are comprehensively evaluated by the two indicators. Finally, the k value corresponding to the minimum comprehensive score is selected as the final number of clusters. The Gaussian distribution covariance structure in the model is fixed to a diagonal matrix, the training termination condition is set to a maximum number of iterations of 300, and the maximum deviation rate in each round of training is set to 2%.

6. The method for monitoring the operation of a thermal power plant monitoring panel with a deviation warning according to claim 1, 2 or 4, characterized in that: The generating of the model operator for deviation identification includes: A five-fold cross validation was used for each training result, and the coefficient of determination was used as the model validity criterion. If the coefficient of determination score was greater than 0.95, it was considered a valid model operator.

7. The method for monitoring the operation of a thermal power plant monitoring panel with a deviation warning according to claim 6, characterized in that: The deviation monitoring of the real-time measurement point value based on the model operator includes: The real-time operation data stream is connected to the thermal power monitoring platform, and the difference between the real-time value of the target measurement point and the predicted value calculated by the corresponding model operator is calculated. The difference is used for deviation trigger judgment; Deviation trigger judgment includes the following rules: Determine whether the difference exceeds the upper and lower limits of the historical deviation interval, which is the 95% confidence interval of the difference distribution during the model training period; Determine whether the difference exceeds the limit state for more than 30 seconds; Determine whether the change rate of the current target measuring point exceeds the dynamic slope threshold; When all judgments are met, it is confirmed as a valid deviation trigger and an early warning is generated.

8. The method for monitoring the operation of a thermal power plant monitoring panel with a deviation warning according to claim 7, characterized in that: The deviation triggering judgment includes: Before executing the difference threshold judgment, a steady-state data filtering mechanism is introduced to perform a linkage check on the sliding mean and sliding variance of the target measuring point in the past 60 seconds. If the difference between the mean and the predicted value of the target measuring point is continuously less than 80% of the preset threshold within 60 seconds, and the variance change rate is less than 5%, the deviation judgment stage is entered.

9. The method for monitoring the operation of a thermal power plant monitoring panel using a deviation warning as claimed in claim 1, 2, 4 or 7, characterized in that: The generating of warning information according to the threshold rule includes: After the deviation judgment of each measuring point, the status tracking logic is set up, which is divided into three status nodes: normal, abnormal confirmation, and warned; When the duration of any deviation reaches the threshold, the system jumps from normal to abnormal confirmation. If it does not recover within 30 seconds, it enters the warning state. Each state change records the timestamp, measurement point ID, predicted value, actual value and deviation value; The generated warning information includes the unique alarm number, time, measurement point name, parameter value and trigger condition, and is displayed synchronously on the monitoring platform homepage.

10. An operation monitoring system with deviation warning on a thermal power monitoring panel, characterized by: It includes a data collection and processing module (100), a prediction algorithm construction module (200), and a deviation monitoring module (300); The data acquisition and processing module (100) is used to acquire and process historical operating data of thermal power units from the SIS system and the DCS system, and extract stable operating condition data according to load stability rules; The prediction algorithm construction module (200) is used to construct a prediction algorithm based on a Gaussian mixture model, set model training and verification standards, and generate a model operator for deviation identification; The deviation monitoring module (300) is used to monitor deviations of real-time measurement point values ​​based on a model operator and generate early warning information according to a threshold value rule.

11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the operation monitoring method of the deviation warning on the thermal power monitoring panel according to any one of claims 1 to 9 are implemented.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the operation monitoring method of the deviation warning on the thermal power monitoring panel according to any one of claims 1 to 9 are implemented.

Images