Method and system for intelligently capturing abnormal data of power system based on Caffe deep learning framework

Abstract

The invention discloses a power system abnormal data intelligent capturing method and system based on a Caffe deep learning framework, and the system comprises an abnormal data automatic capturing module and a data analysis module. According to the invention, the preprocessing module, the label coding module and the automatic protocol stack in the abnormal data automatic capture module are used for cleaning, label coding and the like of the power data to capture the abnormal data; and then the captured abnormal data are classified in a learning node in a data analysis module based on a Caffe deep learning framework and then submitted to a power grid detection host, and the method has greater flexibility when adapting to new requirements of abnormal data detection of the power system, and improves the intelligent level of the system; therefore, the system can better adapt to different application scenes and requirements, and the problems of high abnormal data detection complexity, low processing efficiency, easy over-fitting and poor adaptability in the power grid are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of power data processing, and more specifically, to a method and system for intelligently capturing abnormal data of a power system based on a Caffe deep learning framework. Background Art

[0002] The continuous advancement of modern power system technology and the rapid increase in network coverage have made the structure and operation mode of the power system unprecedentedly complex. The data monitored by the power system is increasing, so it is necessary to improve the accuracy of data monitoring to ensure the stability of the power system operation and avoid misjudgment and wrong judgment. The primary problem is how to timely discover and capture abnormal data in huge data. For example, due to many factors such as power system measurement errors, certain errors may occur when measuring certain data of the power system. The existence of such errors will interfere with the overall prediction and accurate analysis of the power system. Therefore, it is necessary to detect abnormal data in the power grid in a timely and effective manner.

[0003] The amount of monitoring data of the power system is large, the types of data are diverse, and the data itself has a lot of noise. Therefore, how to distinguish and capture abnormal data is the premise of ensuring the prediction and accuracy of the power system. Deep learning is a common algorithm for abnormal data detection. This method can predict and identify abnormal data through model training. However, due to the high complexity of the data, difficulty in classification, and low degree of data correlation, the data processing efficiency of the existing model is low and the prediction accuracy is not high.

[0004] CN117807544A discloses an intelligent detection method for abnormal data of power grid based on deep learning; the method includes obtaining original power grid time series data; performing dimensionality reduction processing on the original power grid time series data through a variational autoencoder VAE model to obtain the reduced-dimensional power grid time series data; performing self-attention processing on the reduced-dimensional power grid time series data through a Transformer model to obtain power grid time series data with context information; and performing prediction processing on the power grid time series data with context information through a long short-term memory neural network LSTM model to obtain the detection result of abnormal data of power grid. The patent captures the context information of power grid time series data based on the attention mechanism of the Transformer model, effectively captures the long-range dependency in the input sequence, increases the time series correlation between data, and thus accurately identifies abnormal results. However, due to the use of multiple deep learning models, the complexity of the system is increased, the efficiency of data processing is low, and it may also cause overfitting problems, affecting the accuracy of the prediction results. At the same time, complex models have high requirements on the degree of correlation of data, especially when the power grid operating conditions or load patterns change, it may be difficult to generalize to new and unseen data. Summary of the invention

[0005] The main technical problem to be solved by the present invention is to provide a method for intelligently capturing abnormal data in a power system based on the Caffe deep learning framework in order to address the shortcomings of the prior art in detecting abnormal data in a power grid, such as high complexity, low processing efficiency, easy overfitting and poor adaptability.

[0006] Another technical problem of the present invention is to provide an intelligent capture system for abnormal data of a power system based on the Caffe deep learning framework.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A method for intelligently capturing abnormal data of a power system based on a Caffe deep learning framework, comprising the following steps:

[0009] S1. Preprocess the power data to eliminate dirty data and erroneous data that may affect the intelligent capture of abnormal data in the power system;

[0010] S2. The abnormal data generated by the power system has obvious characteristic parameters. When identifying abnormal data, certain discrimination conditions are given to the abnormal data, and the detected abnormal data is labeled and encoded;

[0011] S3. Use the automated protocol stack to copy abnormal data and determine data capture mapping conditions;

[0012] S4. Use the Caffe deep learning framework to build a deep learning network, then collect abnormal data, and then convert the abnormal data into storage application parameters in a predetermined format through analysis and integration.

[0013] Furthermore, the preprocessing in S1 includes:

[0014] S11. Determine the missing range and calculate the missing value ratio of each field, and formulate a removal strategy or a completion strategy based on the missing value ratio and field importance;

[0015] S12. Delete unnecessary fields and erroneous data;

[0016] S13. Modify the logically incorrect data according to the difference in power data logic, the results of the same indicator calculation, and the results of different indicator calculations;

[0017] S14. For data with high indicator importance but high missing rate, relevant data are retrieved from the original data of the power system to complete the data.

[0018] Furthermore, the missing value ratio calculation formula of the field is:

[0019]

[0020] Where x is the number of missing fields; all is the total number of records or rows; y is the proportion of missing values.

[0021] Furthermore, the label encoding of abnormal data is expressed as:

[0022]

[0023] Among them, e0 represents the minimum detection label encoding condition of abnormal data in the power system, e n represents the maximum coding condition, P represents the power system anomaly detection label coding, β represents the automatic detection and processing authority of unconventional power system data, It represents the capture output mean value of abnormal data in the power system, Q represents the cleaning ability of abnormal data in the power system, and i is the established label encoding coefficient.

[0024] Furthermore, the copy value of the abnormal data in S3 is expressed as:

[0025]

[0026] Where, d represents the abnormal data perception parameter of the power system to be detected in a certain period of time, ΔT represents the unit time length of the time period corresponding to the abnormal data perception parameter, and τ n and τ1 represent the power system abnormal data copy information parameters of the nth and first input corresponding to the abnormal data perception parameter formation time period, respectively, and l represents the characteristic copy parameter of the power grid abnormal data.

[0027] Furthermore, the step of formulating the data capture mapping conditions in S3 includes:

[0028] S31. Data analysis and exploration: Understand the relationship between different data points in existing data, change trends, and possible anomalies;

[0029] S32. Feature Engineering: Extract and select features relevant to target data capture;

[0030] S33. Establishing mapping rules: establishing mapping rules based on thresholds, logical operations or outputs of machine learning models;

[0031] S34. Verification and adjustment: Verify and adjust the mapping rules.

[0032] Furthermore, the working nodes in the Caffe deep learning framework described in S4 include:

[0033] S41. Define the model: Build a linear neural network as a deep learning model, using the ReLU activation function;

[0034] S42. Training model: training the model using labeled data;

[0035] S43. Application model: Apply to real-time or offline data analysis to detect new abnormal data;

[0036] S44.Data screening: Screen out data marked as abnormal from the original data set;

[0037] S45. Feature extraction: Extract features from abnormal data.

[0038] Furthermore, the power grid monitoring host in the Caffe deep learning framework described in S4 includes: integrating the extracted features into a format suitable for deep learning model input, including data reshaping and data normalization.

[0039] Furthermore, the data normalization is expressed as:

[0040]

[0041] Where x is the original data, x min and x max are the minimum and maximum values ​​of the data respectively.

[0042] An intelligent capture system for abnormal data of a power system based on the Caffe deep learning framework, including an abnormal data automatic capture module and a data analysis module;

[0043] The abnormal data automatic capture module includes power data cleaning, abnormal detection label coding and automation protocol stack. The power data cleaning performs anti-infection processing on the massive data of the power system to eliminate dirty data or erroneous data in the data. The abnormal detection label coding then sets necessary discrimination conditions for abnormal data of the power system. The automation protocol stack provides necessary connection protocol information for abnormal data capture under the Caffe deep learning framework. The Caffe deep learning module acquires abnormal data by copying abnormal data and setting scientific abnormal data capture node conditions.

[0044] The data analysis module includes a learning node, a power grid monitoring host, and an abnormal data set. The abnormal data automatic capture module is connected to the learning node, the power grid monitoring host is connected to the learning node, and the abnormal data set is connected to the power grid monitoring host; the learning node obtains abnormal data of the power grid system from the abnormal data automatic capture module, and the power grid monitoring host mainly functions to physically connect the learning node and the abnormal data set, which can realize the learning node's transfer of captured abnormal data to the abnormal data set through the data transmission channel, and evenly distribute the abnormal data to the learning node according to the deep learning results.

[0045] Compared with the prior art, the beneficial effects are:

[0046] The present invention combines the Caffe deep learning framework to construct an intelligent capture system for abnormal data of a power system. By cleaning power data, setting abnormal detection label coding conditions, and using an automated protocol stack to effectively allocate the intelligent capture nodes of abnormal data of the power system and form data capture mapping conditions, the abnormal data appearing in the power system is integrated into an independent transmission subject, which facilitates the data analysis and transmission of data nodes in the data analysis module, and effectively prevents the abnormal data in the power system from interfering with the power grid monitoring host. Finally, the intelligent capture of abnormal data of the power system is realized through the data analysis module. The present invention utilizes the efficient operation capability of the Caffe framework on the GPU to achieve efficient processing of abnormal data of the power system. This enables the system to complete processing tasks such as cleaning of a large amount of data, abnormal detection label coding, etc. in a short time, thereby ensuring that the system can efficiently and accurately capture abnormal data and improve the overall processing efficiency of the system. By integrating the abnormal data appearing in the power system into an independent transmission subject, the system can effectively prevent the abnormal data from interfering with the power grid monitoring host. At the same time, it facilitates the data analysis and transmission of data nodes in the data analysis module, so that the system can better utilize abnormal data for in-depth analysis and mining. Combined with the automated protocol stack and intelligent dispatching mechanism of the Caffe framework, the system can realize the intelligent capture and processing of abnormal data in the power system.

[0047] The Caffe framework of the present invention is modularly designed from the beginning, allowing easy expansion to new data formats, network layers and loss functions, which makes the framework more flexible in adapting to new requirements for abnormal data detection in power systems, improves the intelligence level of the system, and enables the system to better adapt to different application scenarios and requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a framework diagram of the intelligent capture of abnormal data of an electric power system based on the intelligent capture method of abnormal data of an electric power system under the Caffe deep learning framework of the present invention.

[0049] Figure 2 It is a general flow chart of power data cleaning of the power system abnormal data intelligent capture method based on the Caffe deep learning framework of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be further explained and illustrated below in conjunction with the embodiments, but the specific embodiments do not limit the present invention in any form.

[0051] Example 1

[0052] This embodiment provides a power system abnormal data intelligent capture system based on the Caffe deep learning framework, including an abnormal data automatic capture module and a data analysis module.

[0053] The abnormal data automatic capture module includes power data cleaning, abnormal detection label coding and automation protocol stack. The power data cleaning refers to the anti-infection treatment of the data after the data capture module obtains the massive data of the power system to eliminate the dirty data or erroneous data in the data. The abnormal detection label coding refers to the necessary discrimination conditions set for the abnormal data of the power system. The setting of the abnormal detection label coding needs to rely on the Caffe deep learning module to assign a quantitative value to it after learning. The automation protocol stack is responsible for providing the necessary connection protocol information for abnormal data capture under the Caffe deep learning framework, and realizes the acquisition of abnormal data by the Caffe deep learning module by copying abnormal data and setting scientific abnormal data capture node conditions.

[0054] The data analysis module includes a learning node, a power grid monitoring host, and an abnormal data set. The abnormal data automatic capture module is connected to the learning node, the power grid monitoring host is connected to the learning node, and the abnormal data set is connected to the power grid monitoring host; the learning node obtains abnormal data of the power grid system from the abnormal data automatic capture module, and the power grid monitoring host mainly serves to physically connect the learning node and the abnormal data set, which can not only realize the transfer of the captured abnormal data to the abnormal data set through the data transmission channel, but also evenly distribute the abnormal data to the learning node according to the deep learning results. The abnormal data set collects the abnormal data obtained by the power grid monitoring host, and then converts the abnormal data into storage application parameters in a predetermined format through analysis and integration.

[0055] Example 2

[0056] A method for intelligently capturing abnormal data of a power system based on a Caffe deep learning framework, comprising the following steps:

[0057] S1. Preprocess the power data to eliminate dirty data and erroneous data that may affect the intelligent capture of abnormal data in the power system;

[0058] S11. Remove or fill in missing data: By clarifying the range of missing values, calculating the proportion of missing values ​​in each field, and formulating strategies to remove / fill missing data according to the proportion and field importance.

[0059] The missing value range refers to the specific location or number of rows where data is missing in each field in the data set. This step requires data exploration and analysis to determine which fields have missing values ​​and the distribution of missing values. Fields with a high missing value ratio may affect the overall data quality, especially in deep learning models, where a high proportion of missing data may lead to unstable model training or degraded performance. Therefore, fields with a high proportion may need to be given priority for completion or excluded in subsequent analysis. Determine the importance of each field to the task of capturing abnormal data based on the characteristics and business needs of the power system data. For example, power indicators that are closely related to anomaly detection may be more important, while secondary or irrelevant indicators can be handled more flexibly in data processing.

[0060] The missing value ratio formula for a field is:

[0061]

[0062] xThe number of missing fields; x all The total number of records or rows; the proportion of missing values ​​in y.

[0063] Removal strategy: For fields with low importance and high missing value ratio, these fields or corresponding records can be directly deleted to avoid adverse effects on subsequent analysis or model training.

[0064] Completion strategy: mean / median filling, applicable to numerical data, fills missing values ​​with the mean or median of the field.

[0065] S12. Remove unnecessary fields: After backing up the data in S11, directly delete the erroneous data.

[0066] S13. Modify the data with logical errors: fill in the data according to the difference in power data logic, fill in the data with the results calculated by the same indicator, and fill in the data with the results calculated by different indicators.

[0067] This method of filling the same indicator calculation result is applicable to the missing data or logical errors of the same indicator at different times or locations. The specific steps include: for a certain power indicator (such as current, voltage), if there is missing or erroneous data at a certain time point or device, it can be filled with the data of the indicator at adjacent time points or other devices of the same type.

[0068] This method involves using different but related indicator data to fill in missing or erroneous data. The specific operation includes: based on the physical characteristics or correlation of the power system, other indicators are selected for data filling. For example, if the current data of a transformer at a certain point in time is missing, the current data can be estimated by dividing the voltage data of the transformer at the same time by the resistance value.

[0069] S14. Re-obtain data: For indicators with high importance but high missing rate, relevant data are retrieved from the original data of the power system to complete the data.

[0070] S2. The abnormal data generated by the power system has obvious characteristic parameters. To identify abnormal data, we need to give the abnormal data certain discrimination conditions. Assume that e0 represents the minimum detection label encoding condition of abnormal data in the power system, e n represents the maximum encoding condition, P represents the power system anomaly detection label encoding, then P is expressed as:

[0071]

[0072] Among them, β represents the automatic detection and processing authority of unconventional power system data, which can be set by power system analysis managers according to their actual needs. It represents the captured output mean value of abnormal data in the power system. This value is usually determined by the difference in characteristic value parameters of abnormal data in different power systems. Q represents the cleaning ability of abnormal data in the power system, and i is the established label coding coefficient.

[0073] S3. Use the automated protocol stack to copy abnormal data and determine data capture mapping conditions;

[0074] S31. Preprocessing of abnormal data is achieved by consuming a large amount of server memory. Its advantage is that it can greatly shorten the time for the system to intelligently capture abnormal data and improve the execution rate of detecting abnormal data label encoding. Assuming W is the copy value of the abnormal data of the power system, the value can be expressed as:

[0075]

[0076] Where d represents the abnormal data perception parameter of the power system to be detected in a certain period of time, ΔT represents the unit time length of the time period corresponding to the abnormal data perception parameter, and τ n and τ1 represent the power system abnormal data copy information parameters of the nth and first inputs in the corresponding abnormal data perception parameter formation time period, respectively. Characteristic copy parameter representing abnormal power grid data.

[0077] S32. Formulate data capture mapping conditions: The formulation of data capture mapping conditions mainly revolves around the original node location information of abnormal data in the power system, etc. In the actual formulation of data capture mapping conditions, the mapping conditions of data capture are determined through data-driven mapping.

[0078] Data-driven mapping is the process of determining rules for data capture based on analysis and pattern recognition of existing data.

[0079] Data analysis and exploration: First, conduct in-depth analysis and exploration of existing data in the power system, including understanding the relationship between different data points, changing trends, and possible anomalies.

[0080] Feature engineering: During the data analysis phase, feature engineering is performed to extract and select features relevant to the target data capture. These features can be statistical characteristics of the original data, time series features, or derived features derived from domain knowledge.

[0081] Establish mapping rules: Based on the results of data analysis, establish mapping rules or conditions. These rules can be based on thresholds, logical operations, or the output of machine learning models. For example: if the value of a sensor exceeds a certain threshold within a certain period of time, data capture is triggered; if the value change trend of multiple sensors shows a specific pattern (such as increase or decrease), data capture is triggered.

[0082] Verification and adjustment: Verify and adjust the established mapping rules to ensure that the rules can accurately capture the target data without false triggering or missing data.

[0083] S4. Use the Caffe deep learning framework to build a deep learning network, then collect abnormal data, and then convert the abnormal data into storage application parameters in a predetermined format through analysis and integration.

[0084] A learning node is a server or computer running the Caffe deep learning framework. On this node, deep learning models are defined, trained, and applied.

[0085] Define the model: Use Caffe's prototxt file to define the network structure. Here, a linear neural network is built as a deep learning model, using the ReLU activation function.

[0086] Training the model: Use labeled data (usually a mixture of normal data and known anomalies) to train the model. This involves an iterative process of forward propagation (calculating the output) and backpropagation (updating the weights).

[0087] Applying the model: Once the model is trained, it can be used in real-time or offline data analysis to detect new abnormal data. The abnormal data set refers to the data set collected from the power grid monitoring host and marked as abnormal by the deep learning model.

[0088] Data screening: Filter out data marked as abnormal by the model from the original data set.

[0089] Feature extraction: Extract meaningful features from abnormal data, such as statistical features of time series, spectral features, etc. These features will be used for subsequent abnormal classification or analysis.

[0090] The power grid monitoring host is responsible for collecting real-time data from the power grid, including voltage, current, power factor, etc. These data are usually in the form of time series and may contain noise and outliers. The power grid data is collected through sensors, smart meters and other devices. The raw data is pre-processed by cleaning, filtering, scaling and other operations so that it can be input into the deep learning model.

[0091] Data integration: Integrate the extracted features into a format suitable for deep learning model input. This may involve operations such as reshaping and normalization of the data.

[0092] Data normalization: Scale the data to a uniform range, such as [0, 1] or [-1, 1]. This helps speed up model training and improve the generalization ability of the model. The formula is as follows:

[0093]

[0094] x is the original data, x min and x max are the minimum and maximum values ​​of the data respectively.

[0095] Example 3

[0096] This embodiment provides a verification experiment of the intelligent capture performance of abnormal data in the power system. The power system environment used simulates two generators with a capacity of 500MW and a voltage level of 11kV. Step-up transformer: raises the 11kV voltage to 110kV with a capacity of 600MVA. Secondary high-voltage substation: contains 5 110kV feeders, which are supplied to different low-voltage substations. Low-voltage substation: reduces the 110kV voltage to 10kV and supplies it to factories and general users. Two large factories, each with a daily load curve showing morning and evening peaks, with an average power demand of 20MW. General user loads are residential and commercial electricity, and the total power demand fluctuates over time. A 110kV feeder is simulated to have a short-circuit fault in the middle of the experiment, causing a sudden drop in the voltage of the low-voltage substation powered by the feeder. Abnormal data features include abnormal changes in voltage, current, and power factor. The power system abnormal data intelligent capture system built with the currently commonly used Map-Reduce framework and ISODATA clustering algorithm framework is used as comparison group A and comparison group B to carry out system performance comparative analysis.

[0097] The specific results are as follows:

[0098] Table 1

[0099]

[0100]

[0101] Table 2

[0102]

[0103] As shown in Table 1, in the abnormal data electronic transmission processing volume per unit time at 5, 10, 15, and 20 minutes, the experimental group of the present invention is higher than the comparison group A and the comparison group B, indicating that the power system abnormal data capture limit strength constructed by the present invention is higher than that of the other two groups. On the other hand, in the process of extending the experimental time from 5 minutes to 20 minutes, the abnormal data electronic transmission processing volume of the experimental group has always been stable at about 8.3×1014T / min, while the processing volume of the other two groups is obviously in a downward trend, indicating that the working capacity of the power system abnormal data intelligent capture system constructed by the present invention will not decrease with the extension of the processing time.

[0104] As shown in Table 2, when the data processing volume increased from 1Mb to 5Mb, the time consumed by the experimental group to process abnormal data remained stable at about 2.5s, while the time consumed by the comparison groups A and B to process the same amount of data gradually increased. This is mutually verified with the results reflected in Table 1, that is, the efficiency of the system constructed by the present invention in processing abnormal data of the power system is more stable, and will not decrease due to the increase in processing volume or the extension of processing time.

[0105] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for intelligently capturing abnormal data of a power system based on the Caffe deep learning framework, characterized in that the steps include: S1. Preprocess the power data to eliminate dirty data and erroneous data that may affect the intelligent capture of abnormal data in the power system; S2. The abnormal data generated by the power system has obvious characteristic parameters. When identifying abnormal data, certain discrimination conditions are given to the abnormal data, and the detected abnormal data is labeled and encoded; S3. Use the automated protocol stack to copy abnormal data and determine data capture mapping conditions; S4. Use the Caffe deep learning framework to build a deep learning network, then collect abnormal data, and then convert the abnormal data into storage application parameters in a predetermined format through analysis and integration.

2. According to the method for intelligently capturing abnormal data of a power system based on the Caffe deep learning framework according to claim 1, it is characterized in that: S1 The preprocessing includes: S11. Determine the missing range and calculate the missing value ratio of each field, and formulate a removal strategy or a completion strategy based on the missing value ratio and field importance; S12. Delete unnecessary fields and erroneous data; S13. Modify the logically incorrect data according to the difference in power data logic, the results of the same indicator calculation, and the results of different indicator calculations; S14. For indicators with high importance but high missing rate, relevant data are retrieved from the original data of the power system to complete the data.

3. According to the method for intelligently capturing abnormal data of a power system based on the Caffe deep learning framework according to claim 2, it is characterized in that: The missing value ratio of a field is calculated as: x is the number of missing fields; all is the total number of records or rows; y is the proportion of missing values.

4. According to the method for intelligently capturing abnormal data of a power system based on the Caffe deep learning framework according to claim 1, it is characterized in that: The label encoding of abnormal data is expressed as: Among them, e0 represents the minimum detection label encoding condition of abnormal data in the power system, e n represents the maximum encoding condition, P represents the power system anomaly detection label encoding, β represents the automatic detection and processing authority of unconventional power system data, σ represents the capture output mean value of power system anomaly data, Q represents the cleaning ability of power system anomaly data, and i is the established label encoding coefficient.

5. According to the method for intelligently capturing abnormal data of a power system based on the Caffe deep learning framework according to claim 1, it is characterized in that: The copy value of the abnormal data in S3 is expressed as: Where, d represents the abnormal data perception parameter of the power system to be detected in a certain period of time, ΔT represents the unit time length of the time period corresponding to the abnormal data perception parameter, and τ n and τ1 represent the power system abnormal data copy information parameters of the nth and first inputs in the corresponding abnormal data perception parameter formation time period, respectively. Characteristic copy parameter representing abnormal power grid data.

6. According to the method for intelligently capturing abnormal data of a power system based on the Caffe deep learning framework according to claim 1, it is characterized in that: The step of formulating the data capture mapping conditions described in S3 includes: S31. Data analysis and exploration: Understand the relationship between different data points in existing data, change trends, and possible anomalies; S32. Feature Engineering: Extract and select features relevant to target data capture; S33. Establishing mapping rules: establishing mapping rules based on thresholds, logical operations or outputs of machine learning models; S34. Verification and adjustment: Verify and adjust the mapping rules.

7. The method for intelligently capturing abnormal data of a power system based on the Caffe deep learning framework according to claim 1 is characterized in that: The working nodes in the Caffe deep learning framework described in S4 include: S41. Define the model: Build a linear neural network as a deep learning model, using the ReLU activation function; S42. Training model: training the model using labeled data; S43. Application model: Apply to real-time or offline data analysis to detect new abnormal data; S44.Data screening: Screen out data marked as abnormal from the original data set; S45. Feature extraction: Extract features from abnormal data.

8. The method for intelligently capturing abnormal data of a power system based on the Caffe deep learning framework according to claim 1, characterized in that: The power grid monitoring host in the Caffe deep learning framework described in S4 includes: integrating the extracted features into a format suitable for deep learning model input, including data reshaping and data normalization.

9. The method for intelligently capturing abnormal data of a power system based on the Caffe deep learning framework according to claim 1, characterized in that: The data is normalized as follows: Where x is the original data, x min and x max are the minimum and maximum values ​​of the data respectively.

10. An intelligent capture system for abnormal data of power system based on Caffe deep learning framework, characterized in that: Including abnormal data automatic capture module and data analysis module; The abnormal data automatic capture module includes power data cleaning, abnormal detection label coding and automation protocol stack. The power data cleaning performs anti-infection processing on the massive data of the power system to eliminate dirty data or erroneous data in the data. The abnormal detection label coding then sets necessary discrimination conditions for abnormal data of the power system. The automation protocol stack provides necessary connection protocol information for abnormal data capture under the Caffe deep learning framework. The Caffe deep learning module acquires abnormal data by copying abnormal data and setting scientific abnormal data capture node conditions. The data analysis module includes a learning node, a power grid monitoring host, and an abnormal data set. The abnormal data automatic capture module is connected to the learning node, the power grid monitoring host is connected to the learning node, and the abnormal data set is connected to the power grid monitoring host; the learning node obtains abnormal data of the power grid system from the abnormal data automatic capture module, and the power grid monitoring host mainly functions to connect the learning node and the abnormal data set in a physical form, which can realize the learning node's transmission of captured abnormal data and abnormal data sets through the data transmission channel, and evenly distribute the abnormal data to the learning node according to the deep learning results.

Images