A method and system for predicting abnormal conditions of power equipment

Abstract

The application relates to a power equipment abnormal condition prediction method and system, and belongs to the technical field of data prediction, which solves the problem of inaccurate prediction of the existing power equipment abnormal condition prediction. The method comprises the following steps: obtaining monitoring data of power equipment in a power dispatching system, and obtaining an initial data set after preprocessing; extracting multi-dimensional data features from the initial data set, and constructing a sample set; constructing a feature fusion model and a result fusion model according to a clustering model and a classification model, and training the feature fusion model and the result fusion model by using the sample set; obtaining a to-be-predicted sample from the monitoring data according to the multi-dimensional data features, identifying whether the to-be-predicted sample has concept drift, if the to-be-predicted sample has concept drift, inputting the to-be-predicted sample into the trained feature fusion model, otherwise, inputting the to-be-predicted sample into the trained result fusion model; and obtaining each prediction result according to the output of the feature fusion model or the result fusion model. The method realizes accurate prediction of the power equipment abnormal condition.

Description

Technical Field

[0001] This invention relates to the field of data prediction technology, and in particular to a method and system for predicting abnormal conditions of power equipment. Background Technology

[0002] As power equipment enters the era of intelligence, data related to the operating status of power equipment is constantly being monitored, measured, and stored, resulting in a massive accumulation of data in the power grid. From the perspective of terminal power equipment, the amount of equipment status information recorded by individual devices is gradually accumulating, and given the wide variety of power equipment, the aggregated data becomes a massive dataset. How to monitor and predict abnormal conditions of power equipment based on this massive dataset without increasing computational overhead is an increasingly important issue.

[0003] Existing methods for anomaly detection mainly focus on detecting sample points that are inconsistent with the distribution of data within the majority of normal ranges and are outside the data clusters. These methods include those based on statistics, classification, and clustering.

[0004] Statistical methods, which use the similarity between the sample points to be tested and the normal data distribution as the evaluation standard, are relatively simple and easy to implement, but they are not suitable for massive amounts of data. Furthermore, when power equipment data experiences concept drift due to oxidation of internal components or irregular load changes, classification and clustering methods alone cannot accurately detect anomalies. Concept drift refers to the unpredictable changes in data distribution over time, significantly impacting the accuracy of power equipment condition predictions. Summary of the Invention

[0005] Based on the above analysis, the present invention aims to provide a method and system for predicting abnormal conditions of power equipment, in order to solve the problem of inaccurate prediction of abnormal conditions of existing power equipment.

[0006] On one hand, embodiments of the present invention provide a method for predicting abnormal conditions of power equipment, including the following steps:

[0007] Acquire monitoring data of power equipment in the power dispatching system, preprocess it to obtain an initial dataset; extract multidimensional data features from the initial dataset to construct a sample set;

[0008] Based on clustering and classification models, feature fusion and result fusion models are constructed, and the feature fusion and result fusion models are trained using sample sets respectively;

[0009] Based on the multidimensional data features, the sample to be predicted is obtained from the monitoring data. It is then identified whether the sample to be predicted has experienced concept drift. If so, the sample to be predicted is passed to the trained feature fusion model; otherwise, it is passed to the trained result fusion model. The prediction results are obtained based on the output of the feature fusion model or the result fusion model.

[0010] Further improvements to the above method involve extracting multidimensional data features from the initial dataset, including: obtaining basic data features from the power equipment data and their associated line data in the initial dataset; deriving domain data features based on the relationships between the basic data features; and deriving error data features based on the errors between the basic data features and the theoretical data.

[0011] Based on further improvements to the above methods, the feature fusion model is a model that extracts new data features from the results of the clustering model, fuses them with the data features of the samples, and then uses the classification model to obtain the prediction results; the result fusion model is a model that weights and fuses the results of the classification model and the clustering model to obtain the prediction results.

[0012] Further improvements to the above method involve training the feature fusion model and the result fusion model using the sample set separately. This is achieved by dividing the sample set into a sample set without concept drift and a sample set with concept drift. The result fusion model is trained using the sample set without concept drift, and the feature fusion model is trained using the sample set with concept drift.

[0013] Based on further improvements to the above method, the resulting fusion model includes a first classification module, a first clustering module, and a logistic regression module. The first classification module trains a classification model using a sample set without concept drift and obtains the first probability of a sample being classified as an anomaly based on the trained model. The first clustering module uses the number of power equipment types as the number of clusters, divides the sample set without concept drift into multiple clusters using the clustering model, obtains the distance from each sample to the center of its cluster and normalizes it to the range [0,1] as the second probability of each sample being classified as an anomaly. The logistic regression module uses the first and second probabilities of each sample as input to the logistic regression model, the sample label as output, calculates the weights of the two inputs using maximum likelihood estimation, and then weights the probabilities of the two anomalies according to the weights to obtain the fused anomaly probability value.

[0014] Based on further improvements to the above method, the feature fusion model includes a second clustering module, a feature fusion module, and a second classification module. The second clustering module uses the number of power equipment types as the number of clusters, and uses the clustering model to divide the sample set experiencing concept drift into multiple clusters, obtaining the cluster number and distance to the center of each cluster for each sample. The feature fusion module combines the cluster number and distance to the center of each sample output by the second clustering module with the original data of the corresponding sample to obtain the sample set to be classified. The second classification module uses the sample set to be classified to train a classification model, obtaining a classification model for actual prediction.

[0015] Based on further improvements to the above method, the basic data characteristics include: sampling time, equipment type, voltage level, measurement status, upper limit of line current, line power limit, short-term current carrying capacity of line, line current value, line voltage value, and current and voltage values ​​of power equipment in phases A, B, and C of a three-phase line.

[0016] Based on further improvements to the above method, the domain data features include: dividing the line current by the upper limit of the line current to obtain the current load factor; dividing the useful power of the power equipment by the line power limit to obtain the power load factor; dividing the B-phase current value by the C-phase current value to obtain the BC-phase current ratio; dividing the A-phase current value by the B-phase current value to obtain the AB-phase current ratio; dividing the A-phase voltage value by the B-phase voltage value to obtain the AB-phase voltage ratio; dividing the A-phase voltage value by the C-phase voltage value to obtain the AC-phase voltage ratio; dividing the BC-phase voltage value by the B-phase voltage value to obtain the BC-phase first interphase voltage ratio; dividing the BC-phase voltage value by the C-phase voltage value to obtain the BC-phase second interphase voltage ratio; and dividing the BC-phase voltage value by the CA-phase voltage value to obtain the BC-phase third interphase voltage ratio.

[0017] Based on further improvements to the above method, the error data characteristics are obtained by comparing the power factor, active power, and reactive power of the power equipment with the calculated power factor, active power, and reactive power, respectively, to obtain the power factor error ratio, active power error ratio, and reactive power error ratio.

[0018] On the other hand, embodiments of the present invention provide a predictive system for abnormal conditions of power equipment, comprising:

[0019] The sample set construction module is used to acquire monitoring data of power equipment in the power dispatching system, preprocess it to obtain an initial dataset, and extract multi-dimensional data features from the initial dataset to construct a sample set.

[0020] The prediction model training module is used to construct feature fusion models and result fusion models based on clustering and classification models, and to train the feature fusion models and result fusion models respectively using sample sets;

[0021] The power equipment prediction module is used to obtain the sample to be predicted from the monitoring data based on multi-dimensional data features, identify whether the sample to be predicted has undergone concept drift, and if so, input the sample to be predicted into the trained feature fusion model; otherwise, input it into the trained result fusion model; and obtain each prediction result based on the output of the feature fusion model or the result fusion model.

[0022] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0023] 1. By combining knowledge from the power industry to extract multi-dimensional data features, capturing data trends and interactions, improving the dimensions of data features, and increasing prediction accuracy.

[0024] 2. Considering the impact of concept drift, model the data using different methods from both classification and clustering perspectives to improve the model's predictive and generalization capabilities; for data that has not experienced concept drift, perform weighted fusion of the two types of results to improve the accuracy of the results; for data that has experienced concept drift, extract new data features from the clustering results to overcome the impact of concept drift.

[0025] 3. Applicable to real-time monitoring of abnormal conditions of power equipment in domestic and international power dispatching systems, providing guidance for scientific decision-making by dispatching and operation personnel.

[0026] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0027] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0028] Figure 1 This is a flowchart of a method for predicting abnormal conditions of power equipment in Embodiment 1 of the present invention;

[0029] Figure 2 This is a schematic diagram of the structure of a power equipment abnormality prediction system according to Embodiment 2 of the present invention. Detailed Implementation

[0030] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0031] Example 1

[0032] A specific embodiment of the present invention discloses a method for predicting abnormal conditions of power equipment, such as... Figure 1 As shown, it includes the following steps:

[0033] S1. Obtain monitoring data of power equipment in the power dispatching system, preprocess it to obtain an initial dataset; extract multi-dimensional data features from the initial dataset to construct a sample set.

[0034] It should be noted that the power dispatching system monitors all power equipment, periodically sampling and storing the monitoring data of each power device and its associated lines in a database. Power equipment includes: meters, transformers, generators, etc. Since each power device is associated with a line, the system collects power equipment data while simultaneously associating it with the corresponding line data to distinguish lines of different voltage levels. Power equipment data includes: equipment type, current, voltage, power factor, active power, and reactive power. Line data includes: voltage level, upper current limit, power limit, short-term current carrying capacity, line length, line current, and line voltage. Voltage levels include: 110kV, 220kV, and 500kV; the upper current limit is the rated current, the power limit is the rated power, and the short-term current carrying capacity is the maximum current value within a short period, used to reflect the line's resilience to sudden events.

[0035] Because the amount of stored monitoring data is relatively large, ETL tools are used to obtain streaming data of multiple power equipment and their lines from the database. The streaming data consists of key-value pairs composed of line IDs and data records. Through data transformation, the attribute data of each data record is extracted to obtain the original dataset.

[0036] The initial dataset is obtained by preprocessing the original dataset, which involves removing outlier data according to the exemption rules. Specifically, the exemption rules include: the power line to which the power equipment belongs is empty; the voltage level, reference voltage, or substation to which the equipment associated with the same line is inconsistent; and the regional information associated with the power equipment is missing.

[0037] It should be noted that this embodiment, in addition to considering individual data features, also explores the interactions between data features and the errors between multiple features and theoretical data to improve the model's generalization ability. Multidimensional data features include basic data features, domain data features, and error data features. Basic data features are obtained from the power equipment data and its associated line data in the initial dataset, and are actual measured values, including: sampling time, equipment type, voltage level, measurement status, upper limit of line current, line power limit, short-term line current carrying capacity, line current value, line voltage value, and the current and voltage values ​​of the power equipment in phases A, B, and C of a three-phase line.

[0038] Domain data features are derived from electrical engineering theory based on the relationships between basic data features, including:

[0039] Divide the line current by the upper limit of the line current to obtain the current load factor; divide the useful power of the power equipment by the line power limit to obtain the power load factor; divide the B-phase current value by the C-phase current value to obtain the BC-phase current ratio; divide the A-phase current value by the B-phase current value to obtain the AB-phase current ratio; divide the A-phase voltage value by the B-phase voltage value to obtain the AB-phase voltage ratio; divide the A-phase voltage value by the C-phase voltage value to obtain the AC-phase voltage ratio; divide the BC-phase voltage value by the B-phase voltage value to obtain the first phase-to-phase voltage ratio of BC; divide the BC-phase voltage value by the C-phase voltage value to obtain the second phase-to-phase voltage ratio of BC; divide the BC-phase voltage value by the CA-phase voltage value to obtain the third phase-to-phase voltage ratio of BC.

[0040] Error data characteristics are derived from the error between basic data characteristics and theoretical data, and include:

[0041] The power factor error ratio is obtained by dividing the power factor of the power equipment by the calculated power factor; the reactive power error ratio is obtained by dividing the reactive power of the power equipment by the calculated reactive power; and the active power error ratio is obtained by dividing the active power of the power equipment by the calculated active power.

[0042] It should be noted that the calculated power factor is obtained by calculating the ratio of active power to apparent power, the calculated reactive power is obtained by calculating the sine value of apparent power and power factor angle, and the calculated active power is obtained by calculating the cosine value of apparent power and power factor angle.

[0043] The basic data features, domain data features, and error data features are extracted from each initial data point to form each sample data point. Based on the historical anomaly information in the monitoring data and the specific power equipment and its line data, the sample data is labeled as abnormal, and the sample data and sample labels are put into the sample set together.

[0044] In one embodiment, feature importance analysis is performed on the multidimensional data features obtained in step S1. For example, the importance scores of the output features are used by random forest to calculate the correlation coefficient between each dimension feature and the label, or the chi-square test is used to calculate the chi-square statistics of each feature to select multiple features that contribute significantly to the prediction of electronic device status. The data of these features are then used to construct a sample set for training the model in step S2, thereby reducing the complexity of the model and saving computational overhead.

[0045] S2. Construct a feature fusion model and a result fusion model based on the clustering model and the classification model, and train the feature fusion model and the result fusion model respectively using the sample set.

[0046] It should be noted that this embodiment takes into account the possibility of concept drift in power equipment data. When concept drift occurs, it is difficult for the model trained using data without concept drift to obtain accurate prediction results. Therefore, two models are constructed for prediction of data with and without concept drift, respectively, and the corresponding model is automatically selected for prediction in real time according to the data situation to improve prediction accuracy.

[0047] Specifically, the result fusion model is used to predict anomalies for samples that have not experienced concept drift. It is a model that weights and fuses the results of classification and clustering models to obtain the prediction result. The feature fusion model is used to predict anomalies for samples that have experienced concept drift. It is a model that extracts new data features from the results of the clustering model, fuses them with the data features of the samples, and then uses the classification model to obtain the prediction result.

[0048] The sample set is divided into a sample set that has not experienced concept drift and a sample set that has experienced concept drift through the following steps:

[0049] The sample set obtained in step S2 is sorted according to the sampling time, and the multiple samples corresponding to each sampling time are used as a time window sample. One time window sample is selected as the baseline time window and placed into the dataset where no concept drift has occurred. The samples of other time windows are compared with the samples of the baseline time window in each dimension. If the number of dimensions exceeding the statistical difference threshold is greater than the quantity threshold, it indicates that the samples of the current time window have undergone concept drift and are placed into the dataset of samples with concept drift; otherwise, they are placed into the dataset of samples where no concept drift has occurred.

[0050] It should be noted that this embodiment uses statistical values ​​to compare whether there are significant differences between samples from two time windows. The statistical values ​​include the mean and variance. Each dimension has a corresponding mean difference threshold and variance difference threshold. When both difference values ​​exceed the corresponding difference threshold, the number of difference dimensions increases by 1. When the final number of difference dimensions exceeds the threshold, it is considered that the difference between the samples in the current time window and the samples in the baseline time window is significant, and concept drift has occurred.

[0051] It should be noted that there are many methods for identifying concept drift. Depending on the actual situation, t-tests and KS tests can also be used to compare whether there are significant differences in the distribution of data. If there are significant differences, it indicates that concept drift has occurred.

[0052] During training, each sample set is further divided into a training set and a test set. The training set is used to train the model, while the test set is used to test the model's performance. Performance metrics such as accuracy, recall, and F-score are used to determine whether the model training has met the requirements, in order to prevent overfitting or undertraining of the model to the training set. These are standard practices and will not be elaborated on separately.

[0053] The training processes for the result fusion model and the feature fusion model will be explained separately below.

[0054] (1) Result Fusion Model

[0055] It should be noted that the result fusion model includes a first classification module, a first clustering module, and a logistic regression module.

[0056] Specifically, the first classification module is used to train a classification model using a sample set that has not experienced concept drift, and to obtain the first probability of a sample being classified as an anomaly based on the trained classification model. The method or model structure used in the classification model is not limited. The model is trained using a sample set that has not experienced concept drift. After training, each sample data point is fed into the trained classification model, and the first probability of each sample being classified as an anomaly is output.

[0057] Understandably, the classification model outputs the probability of each label, that is, the probability of classifying as normal and the probability of classifying as abnormal. In existing technologies, the maximum probability value is usually taken as the final classification. In this embodiment, only the first probability of classifying as abnormal is taken to calculate the subsequent weights.

[0058] The first clustering module uses the number of power equipment types as the number of clusters. It then uses a clustering model to divide the sample set that has not experienced concept drift into multiple clusters. The distance from each sample to its cluster center is obtained and normalized to the range [0,1] as the second probability that each sample is classified as an anomaly. For example, the clustering model uses the K-Means clustering model, and the distance from each sample to its cluster center is calculated using Euclidean distance.

[0059] The logistic regression module takes the first and second probabilities of each sample as input to the logistic regression model, and the sample label as output. The output for anomaly labels is 1, and the output for normal labels is 0. Maximum likelihood estimation is used to calculate the weights of the two inputs. During actual prediction, the probabilities of the two anomalies are weighted according to these weights to obtain a fused anomaly probability value. If this value exceeds a threshold, the sample is considered an anomaly.

[0060] Specifically, the logistic regression model is shown below:

[0061]

[0062] Where x represents the first probability / second probability that the input sample is an anomaly, y represents the label of the sample, ω represents the weight corresponding to the first probability / second probability of anomaly, and b represents the bias term.

[0063] (2) Feature fusion model

[0064] It should be noted that the feature fusion model includes a second clustering module, a feature fusion module, and a second classification module.

[0065] Specifically, the second clustering module uses the number of types of power equipment as the number of clusters, and uses a clustering model to divide the sample set that has experienced concept drift into multiple clusters, obtaining the cluster number to which each sample belongs and the distance to the center of its cluster.

[0066] The feature fusion module is used to merge the cluster number and distance to the center of each sample output by the second clustering module with the original data of the corresponding sample to obtain the sample set to be classified.

[0067] The second classification module is used to train a classification model using the sample set to be classified, and obtain a classification model for actual prediction.

[0068] It should be noted that when dividing the sample set, the baseline time window can be selected multiple times. Each selection results in a sample set without concept drift and a sample set with concept drift. Multiple training sessions are conducted, and the model with the best performance index is selected as the final model.

[0069] Compared with existing technologies, this embodiment considers the impact of concept drift in data and models the data in different ways from both classification and clustering perspectives to improve the predictive and generalization capabilities of the model.

[0070] S3. Obtain the sample to be predicted from the monitoring data based on the multidimensional data features, identify whether the sample to be predicted has undergone concept drift, if so, input the sample to be predicted into the trained feature fusion model, otherwise, input it into the trained result fusion model; obtain each prediction result based on the output of the feature fusion model or the result fusion model.

[0071] It should be noted that, in order to accurately identify whether concept drift has occurred in the data, the actual prediction is to obtain monitoring data with the same sampling time from the monitoring data, and then extract multi-dimensional data features according to step S1 to form the sample to be predicted.

[0072] Identifying whether a sample to be predicted has experienced concept drift involves detecting concept drift between the sample to be predicted and the sample from the previous sampling time.

[0073] If concept drift occurs, the sample to be predicted is fed into the trained feature fusion model, which includes: first, passing through the second clustering module to obtain the cluster number to which each sample to be predicted belongs and the distance to the center of the cluster; then, passing through the feature fusion module to merge the output of the second clustering module with the data of the original sample to be predicted and inputting it into the second classification module; the category corresponding to the highest probability output is the prediction result.

[0074] If no concept drift occurs, the sample to be predicted is passed to the trained result fusion model, including: passing the sample to be predicted to the first classification module and the first clustering module respectively, obtaining the first probability and the second probability of being classified as abnormal from their respective outputs, and then passing it to the logistic regression module, obtaining the label value of the sample according to formula (1), and comparing it with the threshold. If it is greater than the threshold, the prediction result is abnormal; otherwise, it is normal.

[0075] Real-time forecasting of monitoring data allows for the detection of anomalies in power equipment, alerting dispatch and operation personnel to promptly repair the equipment and prevent major problems. Statistical analysis of the forecast results from regional, temporal, and equipment type perspectives facilitates informed decision-making.

[0076] For example, when conducting statistical analysis by region, the analysis considers the rational use of power lines in different regions and plans adjustments to the power transmission pattern based on the lines to which the power equipment belongs. When conducting statistical analysis by time, the analysis combines different time periods each day, different dates each week and month, and different months and seasons each year with regional statistical analysis to rationally schedule power transmission to high-consumption areas during off-peak hours, reducing cost losses caused by line anomalies due to high-power transmission. When conducting statistical analysis by equipment type, the stable operating status, usage time, usage environment, and anomaly occurrence of different types of equipment are correlated to statistically analyze the anomalies caused by different equipment, thereby using the most suitable type of equipment under reasonable conditions.

[0077] Compared with existing technologies, this embodiment provides a method for predicting abnormal conditions of power equipment. It combines knowledge from the power industry to extract multi-dimensional data features, capture data trends and interactions, improve data feature dimensions, and enhance prediction accuracy. Considering the impact of concept drift, it employs different modeling methods from both classification and clustering perspectives to improve the model's predictive and generalization capabilities. For data without concept drift, it performs weighted fusion of the two types of results to improve accuracy. For data with concept drift, it extracts new data features from the clustering results to overcome the impact of concept drift. This method is applicable to real-time monitoring of abnormal conditions of power equipment in domestic and international power dispatching systems, providing guidance for scientific decision-making by dispatching personnel.

[0078] Example 2

[0079] Another embodiment of the present invention discloses a predictive system for abnormal conditions of power equipment, thereby implementing the predictive method for abnormal conditions of power equipment in Embodiment 1. The specific implementation of each module is described in the corresponding description in Embodiment 1. Figure 2 As shown, the system includes:

[0080] The sample set construction module 101 is used to acquire monitoring data of power equipment in the power dispatching system, preprocess it to obtain an initial dataset, and extract multi-dimensional data features from the initial dataset to construct a sample set.

[0081] The prediction model training module 102 is used to construct a feature fusion model and a result fusion model based on the clustering model and the classification model, and to train the feature fusion model and the result fusion model respectively using the sample set;

[0082] The power equipment prediction module 103 is used to obtain the sample to be predicted from the monitoring data based on the multidimensional data features, identify whether the sample to be predicted has undergone concept drift, and if so, input the sample to be predicted into the trained feature fusion model; otherwise, input it into the trained result fusion model; and obtain each prediction result based on the output of the feature fusion model or the result fusion model.

[0083] Since the predictive system for abnormal conditions of power equipment in this embodiment and the aforementioned predictive method for abnormal conditions of power equipment are related and can be mutually referenced, this description is redundant and will not be repeated here. Because this system embodiment shares the same principle as the aforementioned method embodiment, it also possesses the corresponding technical effects of the aforementioned method embodiment.

[0084] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0085] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting abnormal conditions of power equipment, characterized in that, Includes the following steps: Acquire monitoring data of power equipment in the power dispatching system, and obtain an initial dataset after preprocessing; Extract multidimensional data features from the initial dataset and construct a sample set, including: obtaining basic data features from the power equipment data and their associated line data in the initial dataset, deriving domain data features based on the relationship between the basic data features, and deriving error data features based on the error between the basic data features and the theoretical data. Based on clustering and classification models, a feature fusion model and a result fusion model are constructed. The feature fusion model extracts new data features from the clustering model results and fuses them with the data features of the samples before using the classification model to obtain the prediction result. The result fusion model is a model that weights and fuses the results of the classification model and the clustering model to obtain the prediction result. The feature fusion model and the result fusion model are trained using sample sets. Specifically, the sample set is divided into a sample set without concept drift and a sample set with concept drift. The result fusion model is trained using the sample set without concept drift, and the feature fusion model is trained using the sample set with concept drift. Based on the multidimensional data features, the sample to be predicted is obtained from the monitoring data. It is then identified whether the sample to be predicted has experienced concept drift. If so, the sample to be predicted is passed to the trained feature fusion model; otherwise, it is passed to the trained result fusion model. The prediction results are obtained based on the output of the feature fusion model or the result fusion model.

2. The method for predicting abnormal conditions of power equipment according to claim 1, characterized in that, The result fusion model includes a first classification module, a first clustering module, and a logistic regression module. The first classification module trains a classification model using a sample set without concept drift and obtains the first probability that a sample is classified as an anomaly based on the trained model. The first clustering module uses the number of power equipment types as the number of clusters, divides the sample set without concept drift into multiple clusters using the clustering model, obtains the distance from each sample to the center of its cluster and normalizes it to the range [0,1] as the second probability that each sample is classified as an anomaly. The logistic regression module uses the first and second probabilities of each sample as input to the logistic regression model, the sample label as output, calculates the weights of the two inputs using maximum likelihood estimation, and then weights the probabilities of the two anomalies according to the weights to obtain the fused anomaly probability value.

3. The method for predicting abnormal conditions of power equipment according to claim 1, characterized in that, The feature fusion model includes a second clustering module, a feature fusion module, and a second classification module. The second clustering module uses the number of power equipment types as the number of clusters, and uses the clustering model to divide the sample set experiencing concept drift into multiple clusters, obtaining the cluster number and distance to the center of each cluster for each sample. The feature fusion module combines the cluster number and distance to the center of each sample output by the second clustering module with the original data of the corresponding sample to obtain the sample set to be classified. The second classification module uses the sample set to be classified to train a classification model, obtaining a classification model for actual prediction.

4. The method for predicting abnormal conditions of power equipment according to claim 1, characterized in that, The basic data characteristics include: sampling time, equipment type, voltage level, measurement status, upper limit of line current, line power limit, short-term current carrying capacity of line, line current value, line voltage value, and current and voltage values ​​of power equipment in phases A, B, and C of a three-phase line.

5. The method for predicting abnormal conditions of power equipment according to claim 1, characterized in that, The data characteristics of the field include: dividing the line current by the upper limit of the line current to obtain the current load rate; dividing the useful power of the power equipment by the line power limit to obtain the power load rate; dividing the B-phase current value by the C-phase current value to obtain the BC-phase current ratio; dividing the A-phase current value by the B-phase current value to obtain the AB-phase current ratio; dividing the A-phase voltage value by the B-phase voltage value to obtain the AB-phase voltage ratio; dividing the A-phase voltage value by the C-phase voltage value to obtain the AC-phase voltage ratio; dividing the BC voltage value by the B-phase voltage value to obtain the BC first phase-to-phase voltage ratio; dividing the BC voltage value by the C-phase voltage value to obtain the BC second phase-to-phase voltage ratio; and dividing the BC phase voltage value by the CA-phase voltage value to obtain the BC third phase-to-phase voltage ratio.

6. The method for predicting abnormal conditions of power equipment according to claim 1, characterized in that, The error data characteristics are obtained by comparing the power factor, active power, and reactive power of the power equipment with the calculated power factor, active power, and reactive power, respectively, to obtain the power factor error ratio, active power error ratio, and reactive power error ratio.

7. A predictive system for abnormal conditions of power equipment, characterized in that, include: The sample set construction module is used to acquire monitoring data of power equipment in the power dispatching system and obtain the initial dataset after preprocessing. Extract multidimensional data features from the initial dataset and construct a sample set, including: obtaining basic data features from the power equipment data and their associated line data in the initial dataset, deriving domain data features based on the relationship between the basic data features, and deriving error data features based on the error between the basic data features and the theoretical data. The prediction model training module is used to construct a feature fusion model and a result fusion model based on a clustering model and a classification model. The feature fusion model is a model that extracts new data features from the results of the clustering model, fuses them with the data features of the samples, and then uses the classification model to obtain the prediction result. The result fusion model is a model that weights and fuses the results of the classification model and the clustering model to obtain the prediction result. The module trains the feature fusion model and the result fusion model using sample sets. Specifically, the sample set is divided into a sample set without concept drift and a sample set with concept drift. The result fusion model is trained using the sample set without concept drift, and the feature fusion model is trained using the sample set with concept drift. The power equipment prediction module is used to obtain the sample to be predicted from the monitoring data based on the multidimensional data features, identify whether the sample to be predicted has undergone concept drift, and if so, input the sample to be predicted into the trained feature fusion model; otherwise, input it into the trained result fusion model; and obtain each prediction result based on the output of the feature fusion model or the result fusion model.

Images