A method and system for monitoring abnormal behavior data of a genset

Abstract

The present application belongs to the field of electric power data monitoring, and relates to a method and system for monitoring abnormal behavior data of a generator set, comprising: obtaining historical data of a group of units; constructing index data, and based on the index data, correcting abnormal values of historical monitoring data to obtain corrected historical monitoring data; decomposing and reconstructing the corrected historical data to obtain decomposed historical data of multiple frequencies, respectively predicting data based on the decomposed historical data to obtain decomposed monitoring data prediction values of multiple frequencies, and integrating the decomposed monitoring data prediction values to obtain monitoring data prediction results; based on the monitoring data prediction results and a preset deviation range of monitoring data, setting a monitoring data range, and determining whether the monitoring data of the unit is abnormal based on the monitoring data range; to ensure the accuracy of the data monitoring of the generator set and the safety of the operation of the power system, and to determine the threshold value of the pre-estimate index based on the decomposed and integrated prediction method.

Description

Technical Field

[0001] This invention relates to the field of power generation data monitoring, and specifically discloses a method and system for monitoring abnormal behavior data of generator sets. Background Technology

[0002] As the core unit of power system energy production and output, generator sets generate multi-dimensional data during their operation, which is crucial for reflecting the unit's operating status and ensuring the balance and stability of power system supply and demand. The stable output and reasonable fluctuations of this data are directly related to the load matching accuracy, power transmission efficiency, and overall operational safety of the power grid. If abnormal fluctuations occur and are not identified and addressed in a timely manner, it can easily lead to power supply and demand imbalances, deviations of grid operating parameters from safe ranges, and even affect the stability of power transmission, threatening the reliable operation of the power system. Currently, anomaly monitoring of the aforementioned multi-dimensional data from generator sets mainly relies on two technical approaches: one is monitoring indicators designed based on the operating characteristics of the power system, including relevant parameters used in traditional power systems to assess operating status; the other is anomaly identification methods combined with intelligent algorithms, such as using fuzzy judgment theory to comprehensively evaluate the abnormal characteristics of multiple data dimensions, or building indicator screening models based on machine learning to achieve automated identification of data anomalies. In terms of specific research directions, on the one hand, there is a focus on proposing new monitoring criteria, such as using sensitivity analysis of optimal power flow to correlate the coupling relationship between power generation, grid-connected power, and power flow, thereby assisting in the identification of data anomalies; on the other hand, there is an emphasis on optimizing the intelligence level of identification methods to improve the accuracy of anomaly detection. However, existing anomaly monitoring methods for multi-dimensional data of generator units still have significant technical shortcomings and are difficult to meet the needs of accurate monitoring under complex operating conditions.

[0003] In view of this, the present invention proposes a method and system for monitoring abnormal behavior data of generator sets, which takes into account both operating condition adaptability and cross-dimensional collaborative identification capabilities, and has self-optimization characteristics. This abnormal monitoring method solves the problems of difficulty in determining thresholds, historical anomaly interference, and lack of self-learning ability in the prior art, ensuring the accuracy of generator set data monitoring and the safety of power system operation, and determines the threshold of pre-evaluation indicators based on the decomposition and integration prediction method. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for monitoring abnormal behavior data of generator sets, addressing the problems of difficulty in determining thresholds, interference from historical anomalies, and lack of self-learning ability in the detection of abnormal power generation data in existing technologies; the specific solution is as follows:

[0005] A method for monitoring abnormal behavior data of generator sets includes: acquiring historical data of a generator group; the historical data includes historical monitoring data and historical feature data; the historical feature data includes historical meteorological data, historical load data, and historical fuel data; constructing index data based on the historical data, and correcting outliers in the historical monitoring data based on the index data to obtain corrected historical monitoring data; the index data includes structural indicators and behavioral indicators; decomposing and reconstructing the corrected historical data to obtain decomposed historical data at multiple frequencies, performing data prediction based on the decomposed historical data to obtain predicted values ​​for decomposed monitoring data at multiple frequencies, and integrating the predicted values ​​for decomposed monitoring data to obtain a monitoring data prediction result; the corrected historical data includes corrected historical monitoring data and historical feature data; setting a monitoring data range based on the monitoring data prediction result and a preset deviation range of the monitoring data, and determining whether the monitoring data of the generator set is abnormal based on the monitoring data range.

[0006] Furthermore, based on the historical data, structural indicators are constructed for the unit group, including: determining the share index of the unit group based on its share ratio among multiple unit groups; using the sum of the share ratios of a preset number of unit groups with the highest share ratios as the dominance index of the unit group; and identifying abnormal unit groups based on the share index and the dominance index.

[0007] Furthermore, based on the historical data, behavioral indicators are constructed for each unit in the abnormal unit group, including: performing time-period alignment processing on historical monitoring data from different days to obtain time-period aligned monitoring data; filtering historical feature data to determine similar days to the test day; analyzing the time-period aligned monitoring data of similar days using a box plot to obtain the normal data range of the monitoring data; identifying monitoring data of the test day that is not within the normal data range as abnormal monitoring data; identifying units with abnormal monitoring data as abnormal units, and identifying the abnormal monitoring data as outliers.

[0008] Furthermore, the step of filtering historical feature data to determine similar days for the test day includes: constructing a feature vector for the test day based on the feature data of the test day; constructing a feature vector for historical operating days based on historical feature data, and combining multiple historical operating day feature vectors to obtain a historical operating feature matrix; constructing a monitoring vector for historical operating days based on historical monitoring data, and combining multiple historical operating day monitoring vectors to obtain a historical monitoring feature matrix; extracting features from multiple historical operating days in the historical operating feature matrix to obtain multiple historical operating feature vectors corresponding to multiple features; calculating the relationship strength between the historical operating feature vector of each feature and the historical monitoring feature matrix using the maximum mutual information coefficient to obtain a feature weight vector; calculating the feature correlation coefficient for each feature based on the similarity between the test day and multiple historical operating days on each feature; performing a weighted average summation of the feature correlation coefficients based on the feature weight vector to obtain the historical day correlation coefficient for each historical operating day; and filtering historical operating days as similar days according to the high and low historical day correlation coefficients.

[0009] Furthermore, the feature weight vector is:

[0010] ;

[0011] in, This represents the weight of the k-th feature; k represents the feature variable, and K represents the total number of features; This represents the historical running feature vector of the k-th feature. The maximum mutual information coefficient between the feature matrix P and the historical monitoring feature matrix; P represents the historical operational feature vector; P represents the historical monitoring feature matrix.

[0012] The formula for calculating the maximum mutual information coefficient is as follows:

[0013] ;

[0014] in, Indicates that when the condition is met The maximum value under the given conditions; This indicates that the total number of cells in the historical operation feature vector and historical monitoring feature matrix under a specific grid division must be less than [a certain value]. ; The expression represents a function of the sample size; S represents the sample size. This represents the historical feature vector of the k-th feature. Mutual information with the historical monitoring feature matrix P; This represents a logarithmic function with a base of 2; min indicates taking the minimum value. This refers to the number of grids after data gridding when calculating mutual information;

[0015] The feature correlation coefficient is:

[0016] ;

[0017] in, The eigenvalue represents the eigencorrelation coefficient between the k-th feature of the historical operating day s and the day to be measured; s represents the historical operating day variable. This represents the minimum absolute difference among all characteristics across all historical operating days. This represents the k-th feature value of the day to be measured; This represents the k-th feature value of the s-th historical running day; Represents the resolution coefficient; This represents the maximum absolute difference among all characteristics across all historical running days.

[0018] The historical daily correlation coefficient is:

[0019]

[0020] in, This represents the historical day correlation coefficient for the s-th historical operating day.

[0021] Furthermore, the normal data range is as follows:

[0022] [Q1-1.5(Q3-Q1),Q3+1.5(Q3-Q1)];

[0023] Where Q1 is the first quantile parameter; Q3 is the third quantile parameter;

[0024] The monitoring data for the day to be measured are as follows:

[0025] ;

[0026] ;

[0027] in, This indicates the offset rate of unit i on abnormal historical operating days; This represents the weighted average of the monitoring data for unit i; The value represents the weighted average of monitoring data from units of the same type as unit i during normal operation; h represents the time period variable. This represents the monitoring data of unit i during each time period h; This indicates the amount of monitoring data consumed by unit i during each time period h on abnormal historical operating days.

[0028] Furthermore, the predicted results of the unit monitoring data are obtained, including: performing multivariate empirical mode decomposition on the corrected historical data to obtain multiple intrinsic mode functions and trend terms; reconstructing the multiple intrinsic mode functions into high-frequency terms and low-frequency terms; and using a dynamic model to process the trend terms, the high-frequency terms and the low-frequency terms to obtain the predicted values ​​of the unit monitoring data for each time period.

[0029] Furthermore, constructing the dynamic model includes: using the augmented Dickey-Fuller test to perform unit root tests on the low-frequency and trend terms to determine whether the sequence is stationary, obtaining a stationarity test result; based on the stationarity test result, establishing an autoregressive integral moving average model, including: if the sequence passes the stationarity test, establishing a first autoregressive integral moving average model; if the sequence fails the stationarity test, establishing a second autoregressive integral moving average model after differencing the sequence; the sequence refers to the low-frequency term and / or trend term; based on the autoregressive integral moving average model, according to the self-phase... The parameters of the autoregressive integral moving average model are determined by using the correlation function and partial autocorrelation function to obtain the final autoregressive integral moving average model. The autoregressive conditional heteroscedasticity (ACH) effect of the residuals of the final autoregressive integral moving average model is examined, and the dynamic model is determined based on whether the residuals are correlated. This includes: if the residuals are correlated, a generalized autoregressive conditional heteroscedasticity (GCH) model is constructed for the residual sequence, and the GCH model is used as the dynamic model; if the residuals are not correlated, the residuals of the final autoregressive integral moving average model are used as the dynamic model.

[0030] Furthermore, the autoregressive integral moving average model is as follows:

[0031] ;

[0032] in, This represents the value of the monitoring data sequence for each time period on day s. This indicates that the monitoring data sequence for each time period is in s- The value of the day; Indicates the first constant term; Let L represent the number of autoregressive variables; Represents the coefficient of the autoregressive term; represents the residual term on day s; e represents the moving average term variable; and E represents the total number of moving average terms. The coefficient representing the moving average term; This represents the residual term on day se; k represents the coefficient variable, and K represents the total number of coefficient variables; Represents eigenvalues The coefficient; This represents the k-th feature value of the s-th historical running day;

[0033] The generalized autoregressive conditional heteroscedasticity model is as follows:

[0034] ;

[0035] ;

[0036] in, This represents the conditional standard deviation on day s. Represents the standardized residual; Represents conditional variance. denoted as the second constant term; r represents the autoregressive conditional heteroscedasticity coefficient variable, and R represents the total number of autoregressive conditional heteroscedasticity coefficient variables; The coefficients of the autoregressive conditional heteroscedasticity effect variable term; This represents the residual term for day ti introduced from the autoregressive integral moving average model; Let G represent the coefficient variable of the generalized autoregressive conditional heteroscedasticity term, and let G represent the total number of generalized autoregressive conditional heteroscedasticity terms. This represents the coefficient of the generalized autoregressive conditional heteroscedasticity term; This indicates the number of characteristic variables introduced into the variance equation. This represents the total number of features introduced into the variance equation; Indicates the first The coefficients of each feature; Indicates the first The observed values ​​of each feature on day s.

[0037] This invention also provides a system for monitoring abnormal behavior data of generator sets using the aforementioned method, comprising a data collection and preprocessing module, a historical monitoring data anomaly identification and correction module, a decomposition and integration hybrid prediction module, and an early warning module for anomalies; the data collection and preprocessing module is used to acquire historical data of the generator set group; the historical data includes historical monitoring data and historical feature data; the historical feature data includes historical meteorological data, historical load data, and historical fuel data; the historical monitoring data anomaly identification and correction module is used to construct indicator data based on the historical data, and to correct outliers in the historical monitoring data based on the indicator data. The system obtains corrected historical monitoring data; the indicator data includes structural indicators and behavioral indicators; the decomposition and integration hybrid prediction module is used to decompose and reconstruct the corrected historical data to obtain decomposed historical data of multiple frequencies, perform data prediction based on the decomposed historical data to obtain predicted values ​​of decomposed monitoring data of multiple frequencies, and integrate the predicted values ​​of decomposed monitoring data to obtain the monitoring data prediction result; the corrected historical data includes corrected historical monitoring data and historical feature data; the pre-anomaly warning module is used to set the monitoring data range based on the monitoring data prediction result and the preset deviation range of the monitoring data, and determine whether the monitoring data of the unit is abnormal based on the monitoring data range.

[0038] The present invention has the following advantages and beneficial effects:

[0039] This invention designs a two-stage anomaly identification and correction system. The first stage uses structural indicators (adapted to correlate the supply-demand matching degree of power generation and grid connection, and the rationality of the distribution of clearing prices and clearing volumes) to pre-identify generating units with potential anomalies, quickly narrowing down the key monitoring scope and effectively reducing the overall difficulty and computational complexity of anomaly identification. The second stage uses abnormal behavior data indicators to further analyze the collaborative anomaly characteristics of multi-dimensional data within the pre-identified key units. For example, it combines historical data to design two types of judgment criteria: pre-event (based on the normal fluctuation range of power generation / grid connection under similar operating conditions) and post-event (the magnitude of the deviation of the clearing price from the normal range, and the deviation rate of power generation from the planned value). Following the logic of first identifying the type of data anomaly and then analyzing its impact on the system, it gradually locks down specific abnormal behavior data items, ultimately achieving comprehensive, hierarchical, and accurate identification of multi-dimensional data anomalies in generating units, avoiding anomaly omissions or misjudgments caused by monitoring a single indicator.

[0040] The historical data anomaly identification and correction method designed in this invention can identify and correct potential undetected latent anomalies in historical monitoring data of generator units, generating a clean historical dataset free from anomaly interference. This correction process effectively alleviates the monitoring bias caused by traditional methods that directly use historical data containing anomalies. It avoids misjudgments due to biased historical data when determining current anomalies and provides a reliable data foundation for subsequent model training based on historical data, thereby improving the accuracy of anomaly monitoring and correlation prediction from the source.

[0041] This invention deeply correlates historical feature data with historical monitoring data within the generator set, constructing a decomposition-integration hybrid prediction model that can fully adapt to the actual operating scenarios of the generator set. On the one hand, the model can accurately predict the normal fluctuation range of multi-dimensional monitoring data of the generator set, providing reasonable and dynamic indicator thresholds for anomaly detection (e.g., differences in power generation anomaly thresholds under different meteorological conditions), avoiding the inapplicability of fixed thresholds under complex operating conditions. On the other hand, the model has higher prediction accuracy for segmented data (e.g., clearing volume at different times or power generation in different load intervals), providing a more accurate reference benchmark for anomaly monitoring and further improving the reliability of overall anomaly identification. Attached Figure Description

[0042] Figure 1 This is an exemplary schematic diagram of a system for monitoring abnormal behavior data of generator sets proposed in this invention;

[0043] Figure 2 This is an exemplary process for applying the method for monitoring abnormal behavior data of generator sets to the monitoring of abnormal market forces.

[0044] Figure 3 This is an exemplary flowchart illustrating the application of the method for correcting outliers in historical generator monitoring data to market power anomaly monitoring. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed in various different configurations. The method of the present invention can be used in various anomaly monitoring scenarios, and is not limited to power generation scenarios. When the present invention is applied to a power generation scenario, the monitored data objects may include, but are not limited to, anomaly detection of grid-connected electricity, power generation, price quotations, market power, unit operating time and / or operating status, etc.

[0046] This invention provides a method for monitoring abnormal behavior data of generator sets, comprising:

[0047] Step 1: Obtain historical data for the generating unit group. This historical data includes historical monitoring data and historical characteristic data. Historical characteristic data includes historical meteorological data, historical load data, and historical fuel data. Historical data refers to multi-dimensional data related to the historical operation of the generating units. Historical data can include historical monitoring data, historical characteristic data, and unit structure data. Monitoring data refers to data reflecting the internal operation of the generating units obtained by monitoring them. For example, when monitoring abnormal power generation behavior data, the monitoring data can refer to the unit's power generation data, which can include power generation and grid connection power. Similarly, when monitoring abnormal market behavior data, the monitoring data can refer to the unit's transaction data, which can include unit bid data, clearing price, and clearing volume. Characteristic data refers to external characteristic data that affects the historical monitoring data of the generating units. For example, characteristic data can include meteorological data, load demand data, and fuel supply data. Unit structure data mainly refers to the basic parameters of the units; these parameters mainly include the unit type and unit name. Each power plant's generating units can be considered a unit group, and multiple unit groups can be obtained for multiple power plants. Meteorological data mainly refers to data closely related to electricity, such as temperature, humidity, wind speed, and irradiance. In some embodiments, unit structure data can be obtained by connecting with the power grid dispatch system, unit transaction data can be obtained by connecting with the power trading platform, unit power generation data can be obtained by the power generation data acquisition system, and connections can be established with global reanalysis datasets and regional high-resolution meteorological data. Finally, fuel price-related data can be obtained. Fuel price-related data includes the coal purchase price index and the comprehensive natural gas purchase price.

[0048] In some embodiments, the method further includes filling in missing values ​​for the acquired historical data, primarily addressing potential gaps in meteorological and monitoring data acquired during the acquisition process. Kriging interpolation is used to repair meteorological data, and linear interpolation is used to fill in the missing values ​​for the time-correlation characteristics of the one-dimensional monitoring data from each unit.

[0049] Step two: Based on the historical data, construct indicator data, and based on the indicator data, correct outliers in the historical monitoring data to obtain corrected historical monitoring data. The indicator data includes structural indicators and behavioral indicators; specifically, structural indicators and abnormal behavioral data indicators. Structural indicators are used to identify potentially abnormal generating units. For example, identifying generating units with potential market manipulation. Another example is identifying generating units with abnormal grid connection electricity. Behavioral indicators are used to identify potentially abnormal generating units. For example, identifying anomalies in the historical data of potential market manipulation units' declarations and exits. Another example is identifying anomalies in the grid connection electricity of generating units.

[0050] In some embodiments, based on the historical data, structural indicators are constructed for the unit group, including:

[0051] The share index of a generating unit group is determined based on its proportion within multiple generating unit groups. When monitoring abnormal market power behavior, the share proportion can be considered as market share. When monitoring abnormal power generation behavior data, the share proportion can be considered as the proportion of the unit's total power generation. The expression for the share index is:

[0052] ;

[0053] Where HHI represents the share index; n represents the unit group variable; and N represents the total number of unit groups. This represents the amount of monitoring data consumed by unit group n in the total unit group. For example, when monitoring abnormal power generation behavior data of the units, it can be the on-grid power consumption of the unit group; Q represents the amount of monitoring data consumed by the total unit group.

[0054] The sum of the share percentages of a predetermined number of generating unit groups with the highest share percentages is used as the dominance index of the generating unit groups. For example, the sum of the share percentages of the top four generating unit groups in terms of on-grid electricity is used as the dominance index. Based on the share index and the dominance index, abnormal generating unit groups are identified. Generating unit groups whose share index and dominance index are both greater than their corresponding thresholds are identified as abnormal generating unit groups.

[0055] In some embodiments, based on the historical data, constructing behavioral indicators for each unit in the abnormal unit group includes: performing time-period alignment processing on historical monitoring data from different days to obtain time-period aligned monitoring data; filtering historical feature data to determine similar days to the day to be measured; and analyzing the time-period aligned monitoring data of similar days using a box plot to obtain the normal data range of the monitoring data; wherein the normal data range is:

[0056] [Q1-1.5(Q3-Q1),Q3+1.5(Q3-Q1)];

[0057] Where Q1 is the first quantile parameter; Q3 is the third quantile parameter.

[0058] Monitoring data for the test day that is outside the normal data range will be considered abnormal monitoring data; units with abnormal monitoring data will be considered abnormal units, and the abnormal monitoring data will be considered outliers. The monitoring data for the test day is as follows:

[0059] ;

[0060] ;

[0061] in, This indicates the offset rate of unit i on abnormal historical operating days; This represents the weighted average of the monitoring data for unit i, for example, the weighted average of the electricity generated. The value represents the weighted average of monitoring data from units of the same type as unit i during normal operation; h represents the time period variable. This represents the monitoring data of unit i during each time period h; This indicates the amount of monitoring data consumed by unit i during each time period h on abnormal historical operating days.

[0062] In some embodiments, the step of filtering historical feature data to determine similar days to the test day includes: constructing a feature vector for the test day based on the feature data of the test day; constructing a feature vector for historical operating days based on historical feature data, and combining multiple historical operating day feature vectors to obtain a historical operating feature matrix; constructing a monitoring vector for historical operating days based on historical monitoring data, and combining multiple historical operating day monitoring vectors to obtain a historical monitoring feature matrix; extracting features from multiple historical operating days in the historical operating feature matrix to obtain multiple historical operating feature vectors corresponding to multiple features; calculating the relationship strength between the historical operating feature vector of each feature and the historical monitoring feature matrix using the maximum mutual information coefficient to obtain a feature weight vector; the feature weight vector is:

[0063] ;

[0064] in, This represents the weight of the k-th feature; k represents the feature variable, and K represents the total number of features; This represents the historical feature vector of the k-th feature. The maximum mutual information coefficient between the feature matrix P and the historical monitoring feature matrix; represents the historical operational feature vector; P represents the historical monitoring feature matrix.

[0065] The formula for calculating the maximum mutual information coefficient is as follows:

[0066] ;

[0067] in, Indicates that when the condition is met The maximum value under the given conditions; This indicates that the total number of cells in the historical operation feature vector and historical monitoring feature matrix under a specific grid division must be less than [a certain value]. ; The expression represents a function of the sample size; S represents the sample size. This represents the historical feature vector of the k-th feature. The mutual information with the historical monitoring feature matrix P measures how much the uncertainty of one variable is reduced after knowing the value of another variable. This represents a logarithmic function with a base of 2; min indicates taking the minimum value. This indicates taking the absolute value.

[0068] Based on the similarity between the test day and multiple historical operation days for each feature, the feature correlation coefficient for each feature is calculated; the feature correlation coefficient is:

[0069] ;

[0070] in, The feature correlation coefficient represents the k-th feature of the historical running day s and the day to be measured. It measures the shape similarity between the day to be measured and the historical day s on a single feature k. The closer the value is to 1, the more similar the two are in shape. s represents the historical running day variable. This represents the minimum absolute difference among all characteristics across all historical operating days. This represents the k-th feature value of the day to be measured; This represents the k-th feature value of the s-th historical running day; This represents the resolution coefficient, typically set to 0.5, and is used to adjust the resolution of the calculated results. A smaller value results in a higher resolution. This represents the maximum absolute difference among all characteristics across all historical operating days.

[0071] The historical daily correlation coefficient is obtained by weighting and summing the feature correlation coefficients based on the feature weight vector; the historical daily correlation coefficient is:

[0072]

[0073] in, This represents the historical day correlation coefficient for the s-th historical operating day.

[0074] Historical running days are selected as similar days based on their correlation coefficients.

[0075] Example 1

[0076] Taking the monitoring of abnormal market behavior data as an example, such as Figure 2 and Figure 3 As shown, step two will be further explained:

[0077] Structural indicators can include share index and share percentage. The share index can be the sum of the squares of the percentages of all cleared bids from each power generation company's units relative to the total cleared market volume. The parameter Q in the share index can be the total cleared volume in the electricity market. This can be viewed as the clearing volume of all units within the nth enterprise; the larger the HHI value, the higher the market concentration and the higher the degree of monopoly. If the value is 1 / N, it indicates that the market share of all enterprises in the industry is the same; if the value is 1, it indicates that the industry is monopolized by a single enterprise. The expression for market share can be:

[0078] ;

[0079] Here, V represents the market share of the top M power generation companies in the market, i.e., the clearing volume ratio, which is the core indicator for measuring market concentration. m represents the unit group variable M, which is a set threshold for the number of suppliers. It is usually taken as 4, that is, the total market share of the four largest power suppliers in the market. This indicator judges the market structure by setting a specific threshold (such as Top-4 ≥ 65%): exceeding the threshold indicates that the market has oligopolistic characteristics, and the higher the value, the stronger the concentration. Based on the above structural primary indicator values, potential market-controlling units are identified, and key observation targets are determined.

[0080] Behavioral indicators can be categorized into historical bid normal range and clearing price deviation rate based on pre- and post-event considerations and the transmission relationship between indicators. The historical bid normal range is a pre-event indicator, while the clearing price deviation rate is a post-event indicator. The former is used to identify abnormal generating units bidding high prices, and then the clearing prices of these abnormal units are further anomaly identified. Since the number of segmented bids for the same generating unit varies across different days, granular alignment of the unit's segmented bid data is necessary before calculating the secondary indicators for bid anomalies. The specific steps for granular alignment are as follows:

[0081] Assume that Unit 1 submits its segmented bid on day s. Segmentation start segment and segment termination segment They are respectively:

[0082]

[0083]

[0084]

[0085] Assume that unit 1 is in the The segmented pricing for the day is The starting segment is The segment termination segment is First, the application segments for generating units were standardized, and the granularity of the standardized pricing segments was then determined as follows: Tianhe The union of the daily price segments, i.e. After union processing, all segment points are sorted, resulting in the new price segment. For Unit 1 on day s and day s The daily price is re-segmented according to the new price segment. The segments do not have any maximum filler values ​​for the price data, such as 99999. These values ​​are ignored in the indicator calculation and are not included in the calculation.

[0086] The historical price range is defined as follows: similar days are determined based on historical meteorological data, and then box plots are used to analyze the price data of similar days to determine the normal range and outliers of the prices.

[0087] The expression for determining similar days is: MIC-GRA similar day selection method: Assuming there are S historical operating days, each historical day consists of K features. It consists of a target quote sequence, forming a shape of The feature matrix, where each row represents a historical day and each column represents a feature, is denoted as:

[0088] ;

[0089] in, Let be the k-th feature value of the s-th historical day. The day to be measured also has a shape as follows: eigenvectors.

[0090] Assume the target price sequence is P, and its shape is... Where H represents the number of segments in the unit's segmented pricing, and each row represents the complete pricing curve corresponding to a historical day:

[0091] ;

[0092] in, Let P be the historical monitoring feature matrix of the unit in the h-th segment for the s-th historical day.

[0093] Constructing feature weights: Using the Maximum Information Coefficient (MIC), calculate the strength of the relationship between each feature and the price curve, and use this strength as the weight of that feature. Let P be the vector composed of the k-th feature of all historical days, i.e., the k-th column of the feature matrix. Let P be the set of all historical day quote sequences.

[0094] After calculating the MIC value for each feature column, normalization is performed to obtain the feature weight vector. The similarity between the test date and each historical date across all features is calculated. Each feature is standardized by dividing each feature value by the mean of that feature for both historical and test dates, and the feature correlation coefficient is calculated. The historical date correlation coefficients are then calculated. These historical date correlation coefficients are sorted from largest to smallest; the larger the coefficient, the greater the similarity. The dates ranked highest are selected as the most similar days.

[0095] The method of identifying outliers using box plots is primarily based on the interquartile range (IQR) method. The principle of IQR is as follows: Determining the outlier range: The segmented bid sequences of historical similar days and the day to be tested are sorted according to their values. After sorting, the bid value of each segment is divided into four equal parts, denoted as Q1, Q2, and Q3, representing the 1st, 2nd, and 3rd quartiles, respectively. The IQR expression is: IQR = Q3 - Q1. The range from Q1 - 1.5IQR to Q3 + 1.5IQR is set as the non-outlier range; data outside this range is considered abnormal behavior data.

[0096] Units with excessively high bids were identified by analyzing historical bids within the normal range, and then post-hoc performance metrics were calculated. During the unified bid granularity processing, bids that were filled with 99999 because no bid was submitted for that segment were not recorded as abnormal bids.

[0097] The monitoring data on the test date can be used as the clearing price deviation rate, which can then be used as a post-hoc indicator to determine the impact of bidding high on the clearing price. For the purpose of reporting a high price, unit i should do so on the day of reporting the high price (an abnormal day). The clearing price deviation rate; The weighted average of the clearance prices for high-priced generating units; For the same type of unit normal day The weighted average of clearing prices. The clearing price for abnormally bidding units (e.g., units with excessively high bids) in each time period h. This refers to the winning bid volume of the high-bid units during each time period (h) on the day of the high bid. A deviation rate of the clearing price exceeding the threshold indicates that the units not only bid high, but their high-bid behavior also affected the clearing price.

[0098] After identifying abnormal price data based on the above indicators, the outliers are corrected. The outliers in each segment are corrected to the average of the normal price segments for the same type of generator set on the same day. If no normal price is available for the same type of generator set on the same day, the correction is made to the average of the normal price segments for the same type of generator set on similar historical days. This yields the corrected historical generator set trading data.

[0099] Step 3: Decompose and reconstruct the corrected historical data to obtain decomposed historical data at multiple frequencies. Based on the decomposed historical data, perform data prediction to obtain predicted values ​​of decomposed monitoring data at multiple frequencies. Integrate the predicted values ​​of the decomposed monitoring data to obtain the monitoring data prediction result. The corrected historical data includes corrected historical monitoring data and historical feature data.

[0100] In some embodiments, obtaining the predicted results of unit monitoring data includes: performing multivariate empirical mode decomposition on the corrected historical data to obtain multiple intrinsic mode functions and trend terms; reconstructing the multiple intrinsic mode functions into high-frequency terms and low-frequency terms; and processing the trend terms, the high-frequency terms and the low-frequency terms using a dynamic model to obtain the predicted values ​​of the unit's monitoring data for each time period.

[0101] In some embodiments, the multiple intrinsic mode functions (IMFs) after decomposition are reconstructed into high-frequency and low-frequency terms based on the Lempel-Ziv algorithm. The specific steps of the Lempel-Ziv algorithm are as follows:

[0102] Calculate the lz value for each IMF. The sum of the IMFs whose cumulative lz values ​​exceed a threshold is the high-frequency term. The threshold is typically 0.8. Specifically:

[0103]

[0104] Where z represents the z-th IMF after sorting by frequency; This represents the number of IMFs that satisfy the above condition after being sorted by frequency from highest to lowest. Let represent the complexity of the z-th IMF; This represents a pre-set threshold, typically 0.8. If the above formula is satisfied, then... to The sum of is high frequency, and the sum of the remaining IMFs is low frequency. The trend term represents the trend direction of the sequence.

[0105] In some embodiments, constructing the dynamic model includes:

[0106] The augmented Dickey-Fuller test (ADF) is used to perform unit root tests on the low-frequency and trend terms to determine whether the series is stationary, yielding the stationarity test results. For example, the ADF test can be used to perform unit root tests on the low-frequency and trend terms to determine whether the series is stationary. If the ADF test shows that the series does not contain a unit root, i.e., it passes the stationarity test, and an ARIMA(L,0,E) model can be directly established. If the ADF test shows that the series contains a unit root, i.e., it fails the stationarity test, the series needs to be differencing to establish an ARIMA(L,d,E) model.

[0107] Based on the stationarity test results, an autoregressive integral moving average model is established, including: if the series passes the stationarity test, a first autoregressive integral moving average model is established; if the series fails the stationarity test, a second autoregressive integral moving average model is established after differencing the series; the series refers to the low-frequency term and / or the trend term; the autoregressive integral moving average model is:

[0108] ;

[0109] in, This represents the value of the monitoring data sequence for each time period on day s. This indicates that the monitoring data sequence for each time period is in s- The value of the day; Indicates the first constant term; Let L represent the number of autoregressive variables; Represents the coefficient of the autoregressive term; represents the residual term on day s; e represents the moving average term variable; and E represents the total number of moving average terms. The coefficient representing the moving average term; Represents the residual on day se; k represents the coefficient variable, and K represents the total number of coefficient variables; Represents eigenvalues The coefficient; This represents the k-th feature value of the s-th historical running day.

[0110] Based on the autoregressive integral moving average model, the parameters of the autoregressive integral moving average model are determined according to the autocorrelation function and the partial autocorrelation function, thus obtaining the final autoregressive integral moving average model.

[0111] The autoregressive conditional heteroscedasticity (ACH) effect of the final autoregressive integral moving average model residuals is examined, and the dynamic model is determined based on whether the residuals are correlated. This includes: if the residuals are correlated, a generalized autoregressive conditional heteroscedasticity (GCH) model is constructed for the residual sequence, and this GCH model is used as the dynamic model; if the residuals are not correlated, the final autoregressive integral moving average model residuals are used as the dynamic model. The generalized autoregressive conditional heteroscedasticity (GCH) model is as follows:

[0112] ;

[0113] ;

[0114] in, This represents the conditional standard deviation on day s. Represents the standardized residual; Represents conditional variance. denoted as the second constant term; r represents the autoregressive conditional heteroscedasticity coefficient variable, and R represents the total number of autoregressive conditional heteroscedasticity coefficient variables; The coefficients of the autoregressive conditional heteroscedasticity effect variable term; This represents the residual term for day ti introduced from the autoregressive integral moving average model; Let G represent the coefficient variable of the generalized autoregressive conditional heteroscedasticity term, and let G represent the total number of generalized autoregressive conditional heteroscedasticity terms. This represents the coefficient of the generalized autoregressive conditional heteroscedasticity term; This represents the number of features introduced into the variance equation; these features that introduce variance are a subset of all features. This represents the total number of features introduced into the variance equation; Indicates the first The coefficients of each feature; Indicates the first The observed value of a feature on day s.

[0115] Example 2

[0116] Taking the monitoring of abnormal market behavior data as an example, step three will be further explained:

[0117] Based on the decomposition-integration prediction framework, the corrected historical unit transaction data is decomposed and reconstructed into modules of different frequencies. Similar days are determined based on features to obtain the transaction data prediction value of the high-frequency module. The mid- and low-frequency modules are modeled to obtain the prediction value of the mid- and low-frequency modules. Finally, the prediction values ​​of each decomposed module are integrated to obtain the final price data prediction result.

[0118] To reduce computational costs, this step only forecasts bids for units identified using the two-stage indicators; that is, it forecasts bids for units with the potential to abuse market power, determining the threshold for abnormal bids. The units mentioned in Step Three are all those with the potential to abuse market power.

[0119] The decomposition and integration prediction framework mainly includes:

[0120] a) The Multi-component Empirical Mode Decomposition (MEMD) method is used to decompose the unit bidding data and unit reporting data into Intrinsic Mode Functions (IMFs) and trend terms at different frequencies. The MEMD process is as follows: after two-stage index correction, historical data of unit segmented bidding is obtained (corrected historical unit monitoring data). Each segmented bidding sequence is then merged with the feature sequence to form a multivariate sequence. ,in For the historical bidding sequence of the h-th segment of unit i, this multivariate sequence is input into the MEMD method, which outputs multiple IMFs: and trend items The decomposed multivariate sequence has the same number of rows and columns as the original sequence.

[0121] b) The high-frequency, low-frequency, and trend terms after multivariate decomposition are in the following forms: ,in These represent the high-frequency, low-frequency, and trend components of the price quote sequence, respectively. These represent the high-frequency, low-frequency, and trend terms of the characteristic sequence, respectively. Using meteorological characteristics, fuel price characteristics, and load demand characteristics from the high-frequency terms, historical similar days are identified for the day to be predicted. Based on the MIC-GRA similar day screening method described above, after determining the set of similar days, the mean of the high-frequency data of the corrected segmented price quotations for the same-named units on similar days is used as the predicted value of the high-frequency term for the day to be predicted.

[0122] c) Construct pricing ARIMA-GARCH models for low-frequency modules and trend terms respectively:

[0123] Determine the ARIMA model parameters based on the ACF (Autocorrelation Function) and PCAF (Partial Autocorrelation Function). and Among them, the autoregressive integral moving average model This represents the value of each segment's price sequence on day s.

[0124] Test the residuals of the ARIMA model for pricing. ARCH effect: Analyze the correlation of residuals. If there is no correlation, the modeling ends and proceeds to step 5; if there is a correlation, a GARCH model for pricing is established. Use the model to make predictions and obtain the predicted values ​​of low-frequency terms and the prices of each segment of low-frequency terms.

[0125] d) Integrate the predicted prices of each segment of high-frequency, low-frequency, and trend items to obtain the final predicted price of each unit in each segment.

[0126] Step four: Based on the predicted results of the unit monitoring data and the preset deviation range of the unit monitoring data, set the unit monitoring data range, and determine whether the current unit monitoring data is abnormal based on the unit monitoring data range. For example, set the predicted value of the power monitoring data as the standard value of the unit segmented monitoring data, and determine the reasonable monitoring data range of each unit based on the preset deviation range to achieve early warning of unit anomalies.

[0127] like Figure 1 As shown, the present invention also provides a system for monitoring abnormal behavior data of generator sets, including a data collection and preprocessing module, a historical monitoring data anomaly identification and correction module, a decomposition-integration hybrid prediction module, and an anomaly identification module. The data collection and preprocessing module is used to acquire historical data of the generator set group. The historical data includes historical monitoring data and historical feature data. The historical feature data includes historical meteorological data, historical load data, and historical fuel data. The historical monitoring data anomaly identification and correction module is used to construct indicator data based on the historical data, and to correct the abnormal values ​​of the historical monitoring data based on the indicator data to obtain corrected historical monitoring data. The indicator data includes structural indicators and behavioral indicators. The decomposition-integration hybrid prediction module is used to decompose and reconstruct the corrected historical data to obtain decomposed historical data of multiple frequencies, perform data prediction based on the decomposed historical data to obtain predicted values ​​of decomposed monitoring data of multiple frequencies, and integrate the predicted values ​​of the decomposed monitoring data to obtain a monitoring data prediction result. The corrected historical data includes corrected historical monitoring data and historical feature data. The anomaly identification module is used to set a monitoring data range based on the monitoring data prediction result and a preset deviation range of the monitoring data, and to determine whether the monitoring data of the generator set is abnormal based on the monitoring data range.

[0128] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of monitoring abnormal behavior data of a genset, characterized by, The method comprises the following steps: acquiring historical data of a unit group; the historical data comprises historical monitoring data and historical characteristic data; the historical characteristic data comprises historical meteorological data, historical load data and historical fuel data; based on the historical data, constructing index data, and based on the index data, correcting outliers of the historical monitoring data to obtain corrected historical monitoring data; the index data comprises structure indexes and behavior indexes; decomposing and reconstructing the corrected historical data to obtain decomposed historical data of multiple frequencies, respectively predicting data based on the decomposed historical data to obtain decomposed monitoring data prediction values of multiple frequencies, and integrating the decomposed monitoring data prediction values to obtain monitoring data prediction results; the corrected historical data comprises corrected historical monitoring data and historical characteristic data; obtaining monitoring data prediction results comprises: performing multivariate empirical mode decomposition on the corrected historical data to obtain multiple intrinsic mode functions and a trend item; reconstructing the multiple intrinsic mode functions into high-frequency items and low-frequency items; processing the trend item, the high-frequency items and the low-frequency items by using a dynamic model to obtain monitoring data prediction values of the unit at each time period; the dynamic model is constructed by: performing unit root test on the low-frequency items and the trend item by using an augmented Dickey-Fuller test method to determine whether a sequence is stationary to obtain a stationary test result; based on the stationary test result, establishing an autoregressive integrated moving average model, comprising: if the sequence passes the stationary test, establishing a first autoregressive integrated moving average model; if the sequence fails to pass the stationary test, establishing a second autoregressive integrated moving average model after differencing the sequence; the sequence refers to the low-frequency items and / or the trend item; based on the autoregressive integrated moving average model, determining parameters of the autoregressive integrated moving average model according to autocorrelation functions and partial autocorrelation functions to obtain a final autoregressive integrated moving average model; testing autoregressive conditional heteroscedasticity effects of residuals of the final autoregressive integrated moving average model, and determining the dynamic model based on whether the residuals have correlation, comprising: if the residuals have correlation, constructing a generalized autoregressive conditional heteroscedasticity model for the residual sequence, and taking the generalized autoregressive conditional heteroscedasticity model as the dynamic model; if the residuals have no correlation, taking the residuals of the final autoregressive integrated moving average model as the dynamic model; based on the monitoring data prediction results and a preset deviation range of monitoring data, setting a monitoring data range, and determining whether the monitoring data of the unit is abnormal based on the monitoring data range.

2. The method of monitoring generator set abnormal behavior data of claim 1, wherein, based on the historical data, constructing structure indexes for the unit group, comprising: determining a share index of the unit group based on a share proportion of the unit group in multiple unit groups; adding share proportions of a preset number of unit groups with higher share proportions as a dominant index of the unit group; determining an abnormal unit group based on the share index and the dominant index.

3. The method of monitoring generator set abnormal behavior data of claim 2, wherein, based on the historical data, constructing behavior indexes for each unit in the abnormal unit group, comprising: performing time period alignment processing on historical monitoring data of different days to obtain time period aligned monitoring data; Filtering historical characteristic data to determine a similar day of a to-be-tested day; Analyzing the monitoring data of the similar day after time alignment by using a box plot to obtain a normal data range of the monitoring data; Taking the monitoring data of the to-be-tested day that is not within the normal data range as abnormal monitoring data; Taking a unit with abnormal monitoring data as an abnormal unit and taking the abnormal monitoring data as an abnormal value.

4. The method of monitoring generator set abnormal behavior data of claim 3, wherein, The filtering of the historical characteristic data to determine the similar day of the to-be-tested day comprises: constructing a to-be-tested day feature vector based on to-be-tested day feature data; constructing historical operation day feature vectors based on historical feature data, and combining a plurality of historical operation day feature vectors to obtain a historical operation feature matrix; constructing historical operation day monitoring vectors based on historical monitoring data, and combining a plurality of historical operation day monitoring vectors to obtain a historical monitoring feature matrix; extracting features of a plurality of historical operation days in the historical operation feature matrix to obtain a plurality of historical operation feature vectors corresponding to a plurality of features; calculating the relationship strength between the historical operation feature vector of each feature and the historical monitoring feature matrix by using a maximum mutual information coefficient to obtain a feature weight vector; calculating a feature correlation coefficient of each feature based on the similarity of the to-be-tested day and a plurality of historical operation days on each feature; performing weighted average summation on the feature correlation coefficients based on the feature weight vector to obtain a historical day correlation coefficient of each historical operation day; filtering historical operation days as similar days according to the high and low of the historical day correlation coefficients.

5. The method of monitoring generator set abnormal behavior data of claim 4, wherein, The feature weight vector is: ； wherein, represents the weight of the kth feature; k represents a feature variable, and K represents the total number of features; represents the historical running feature vector of the kth feature represents the maximum mutual information coefficient of the historical monitoring feature matrix P; represents the historical running feature vector; P represents the historical monitoring feature matrix; The calculation formula of the maximum mutual information coefficient is: ； wherein, represents the maximum value under the condition that ; represents that the total number of grids of the historical operation feature vector and the historical monitoring feature matrix under a certain grid division is less than ; represents a function about sample size; S represents sample size; represents the mutual information of the historical operation feature vector of the kth feature and the historical monitoring feature matrix P; represents a logarithmic function with base 2; min represents taking the minimum value; represents that the number of grids after the data is gridded is used when calculating the mutual information; The feature correlation coefficient is: ； wherein, represents the feature correlation coefficient of the kth feature of the historical operation day s and the day to be tested; s represents the historical operation day variable; represents the minimum absolute difference in all historical operation days and all features; represents the kth feature value of the day to be tested; represents the kth feature value of the s-th historical operation day; represents the resolution coefficient; represents the maximum absolute difference in all historical operation days and all features; The historical day correlation coefficient is: wherein, Hsdenotes the historical day correlation for the s-th historical run day.

6. The method of monitoring generator set abnormal behavior data of claim 3, wherein, The normal data range is: [Q1-1.5(Q3-Q1), Q3+1.5(Q3-Q1)]; wherein Q1 is a first quantile parameter; and Q3 is a third quantile parameter. The monitoring data of the to-be-tested day is: ； ； wherein, represents the deviation rate of unit i on the abnormal historical operation day; represents the weighted average of the monitoring data of unit i; represents the weighted average of the monitoring data of the unit of the same type as unit i when it is normally operated; h represents a time period variable, represents the monitoring data of unit i in each time period h; represents the monitoring data consumption of unit i in each time period h on the abnormal historical operation day.

7. The method of monitoring generator set abnormal behavior data of claim 1, wherein, The autoregressive integrated moving average model is: ； wherein, represents the value of the each-period monitoring data sequence on day s, represents the value of the each-period monitoring data sequence on day s- ; represents a first constant term; represents an autoregressive term variable, L represents the total number of autoregressive terms; represents the coefficient of the autoregressive term; represents a residual term on day s; e represents a moving average term variable, E represents the total number of moving average terms; represents the coefficient of the moving average term; represents a residual term on day s-e; k represents a coefficient variable, K represents the total number of coefficient variables; represents the coefficient of the characteristic value ; represents the kth characteristic value of the s th historical operation day; The generalized autoregressive conditional heteroscedasticity model is: ； ； in, This represents the conditional standard deviation on day s. Represents the standardized residual; Represents conditional variance. denoted as the second constant term; r represents the autoregressive conditional heteroscedasticity coefficient variable; R represents the total number of autoregressive conditional heteroscedasticity coefficient variables. The coefficients of the autoregressive conditional heteroscedasticity effect variable term; This represents the residual term for day ti introduced from the autoregressive integral moving average model; Let G represent the coefficient variable of the generalized autoregressive conditional heteroscedasticity term, and let G represent the total number of generalized autoregressive conditional heteroscedasticity terms. This represents the coefficient of the generalized autoregressive conditional heteroscedasticity term; This indicates the number of characteristic variables introduced into the variance equation. This represents the total number of features introduced into the variance equation; Indicates the first The coefficients of each feature; Indicates the first The observed values ​​of each feature on day s.

8. A system for monitoring abnormal behavior data of a generator set using the method of any one of claims 1-7, wherein, The method comprises a data collection and preprocessing module, a historical monitoring data anomaly identification and correction module, a decomposition and integrated hybrid prediction module, and an anomaly identification module. The data collection and preprocessing module is used to obtain historical data of a unit group; the historical data comprises historical monitoring data and historical characteristic data; the historical characteristic data comprises historical meteorological data, historical load data, and historical fuel data; The historical monitoring data anomaly identification and correction module is used to construct index data based on the historical data, and correct abnormal values of historical monitoring data based on the index data to obtain corrected historical monitoring data; the index data comprises structure indexes and behavior indexes; The decomposition and integrated hybrid prediction module is used to decompose and reconstruct corrected historical data to obtain decomposition historical data of a plurality of frequencies, perform data prediction based on the decomposition historical data to obtain decomposition monitoring data prediction values of a plurality of frequencies, and integrate the decomposition monitoring data prediction values to obtain a monitoring data prediction result; The corrected historical data comprises corrected historical monitoring data and historical characteristic data; The abnormality identification module is configured to set a monitoring data range based on the preset deviation range of the monitoring data prediction result and the monitoring data, and determine whether the monitoring data of the unit is abnormal based on the monitoring data range.

Images