A method and system for constructing event-based knowledge model for power grid monitoring business based on deep learning

Abstract

The present invention discloses a method and system for constructing an event-based knowledge model of power grid monitoring services based on deep learning. The method comprises: collecting alarm information data during the historical operation of the power grid, preprocessing the power grid monitoring alarm text, and constructing a power grid monitoring alarm information event-based rule set; using the event-based rule set to process the historical alarm information data to obtain a data set of various labeled monitoring alarm events; performing unsupervised pre-training on the pre-processed monitoring alarm text through a word2vec model to obtain a vectorized expression of the monitoring alarm text; offline training and online application; the system comprises a preprocessing module, a pretraining module, a testing module, and an identification module. The present invention provides timeliness guarantee for power grid fault diagnosis and accident analysis; provides support for event definition and autonomous generation of event rules, and can realize automatic expansion of the monitoring service event knowledge model.

Description

Technical Field

[0001] The present invention relates to the field of power grid monitoring and operation, and in particular to a method and system for constructing an event-based knowledge model for power grid monitoring business based on deep learning. Background Art

[0002] With the continuous improvement of the intelligence level of the power grid and the gradual advancement of the integration of power grid operations, the number of substations, power grid equipment and alarm information connected to centralized monitoring by control agencies continues to increase.

[0003] These alarm information is a summary of the status response and rule experience of power grid equipment. If the information value can be deeply mined, it can not only improve the operation and management level of the power grid, but also provide better power services to the society. Faced with a large amount of power grid operation alarm information, the current smart grid dispatching and control system has designed and implemented decision support functions such as centralized monitoring alarm application and comprehensive intelligent alarm, providing technical support for power grid dispatching fault judgment and disposal and equipment monitoring and operation. However, the existing technology still has the following problems:

[0004] The real-time processing method of monitoring alarm signals is inefficient. With the expansion of the scale of power grid equipment and the improvement of intelligent monitoring level, the number of equipment failure alarm information has shown a geometric growth trend. Since the alarm information has no distinction in importance, the existing power grid monitoring alarm relies on the monitoring method of responding to alarm information one by one. It is necessary to judge, analyze and feedback each piece of information one by one. Due to the large number of monitoring equipment involved, the large amount of monitoring information, and the complexity of abnormal fault conditions, it is easy to have equipment failure or abnormal misjudgment or omission with traditional technical means such as scrolling alarm information windows and light-emitting sign lists, which has caused great pressure on the monitoring screen for monitoring personnel and can no longer meet the higher requirements of power grid monitoring alarms for coordination and technical support means under the current situation.

[0005] The active real-time perception of power grid monitoring alarm events needs to be improved. With the continuous increase in the number and types of equipment in the power grid, the system architecture has become complex and changeable. When the power grid is in an abnormal state, a series of fault alarm information will be uploaded to the monitoring background. If the monitoring personnel do not handle it in time, it will cause serious economic losses. Therefore, it is necessary to build a power grid monitoring business event knowledge model based on deep learning to solve these problems. Summary of the invention

[0006] Purpose of the invention: The purpose of the present invention is to provide a method and system for constructing an event-based knowledge model for power grid monitoring services based on deep learning.

[0007] Technical solution: The method for constructing an event-based knowledge model for power grid monitoring services based on deep learning of the present invention is characterized in that the method comprises the following steps:

[0008] (1) Collect alarm information data during the historical operation of the power grid, pre-process the power grid monitoring alarm text, and build a set of event-based rules for power grid monitoring alarm information;

[0009] (2) Using the event-based rule set to process the historical alarm information data in step (1), obtaining a data set of various labeled monitoring alarm events;

[0010] (3) Perform unsupervised pre-training on the pre-processed monitoring alarm text through the word2vec model to obtain a vectorized expression of the monitoring alarm text;

[0011] (4) Offline training: Establish a knowledge model of power grid monitoring business events based on convolutional neural networks, and use the power grid monitoring alarm text dataset to train and test the model and learn the alarm text features;

[0012] (5) Online application: Use the trained model to identify monitoring alarm information, and through the knowledge model self-learning technology, realize the automatic expansion of the monitoring business event knowledge model based on the incremental information of power grid operation, thereby improving the robustness of the knowledge model.

[0013] Furthermore, the step (1) of preprocessing the power grid monitoring alarm text includes:

[0014] (1.1) Data cleaning, removing duplicate warning texts: avoid the same information being processed multiple times, reduce data redundancy, and handle missing values: if some key information in the warning text is missing, it needs to be supplemented or deleted according to the specific situation;

[0015] (1.2) Noise removal, filtering out special characters, garbled characters or meaningless symbols;

[0016] (1.3) Remove stop words, remove common words that are not very meaningful for text analysis;

[0017] (1.4) Word segmentation: Use the Jieba library to segment the warning text into words or phrases.

[0018] Furthermore, the step (2) includes: processing the historical alarm text based on the event-based rule set, constructing the event-based rule set according to multiple dimensions of voltage level, event type, equipment, and signal type, and mainly dividing the alarm text into accident level, abnormal level, displacement level, and notification level.

[0019] Furthermore, the step (3) includes: using the Skip-gram model of word2vec to perform word vector training, inputting the preprocessed monitoring alarm information features into the Word2vec model, taking the monitoring alarm text triggered by the fault as the unit of the output feature vector, performing different correlation processing on the feature vector in the text, and obtaining the vectorized expression of the monitoring alarm text, including:

[0020] (3.1) Predict the context words based on the central word. Given a central word ω t , maximizes the following probability:

[0021] P(w t-j , w t-j+1 , L, w t+j-1 , w t+j |w t )

[0022] (3.2) For each context word ω t+i , the loss function of the Skip-gram model is expressed as:

[0023]

[0024] Where T is the total number of words in the text, c is the context window size;

[0025] (3.3) In the Skip-gram model, p(w o |w i ), predict word ω0 based on word ω i The calculation uses the softmax function:

[0026]

[0027] where v i is the input word ω i The vector representation of u j is the output word ω j is a vector representation of , and V is the vocabulary size.

[0028] Furthermore, the main structure of the convolutional neural network in step (4) includes: an input layer, a convolutional layer, a pooling layer and a fully connected layer.

[0029] Furthermore, the training process of the convolutional neural network in step (4) includes:

[0030] (4.1) The input layer is the warning text vector matrix generated by word2vec;

[0031] (4.2) The convolution layer is composed of several convolution kernel modules. Each convolution kernel performs convolution calculation on the text vector representation matrix T to preliminarily extract the local features of the text. The local feature formula obtained by the i-th neuron is:

[0032] c i =RELU(wgx i：i+r-1 +b)

[0033] (4.3) The pooling layer reduces the dimension of the features extracted by the convolutional layer and captures the optimal solution in the local features. The calculation formula is as follows:

[0034] z r =max(c1,c2,c3,·····,c n-r+1 )

[0035] (4.4) The connection layer connects the maximum pooling results obtained by the convolution kernel to obtain the final feature formula of the text as follows:

[0036]

[0037] (4.5) The classification layer outputs the warning text classification results.

[0038] Furthermore, the step (5) comprises:

[0039] (5.1) Continuously collect alarm information during the operation of the power grid. When there is new monitoring alarm information and incremental information, input this information into the trained model for prediction. If the prediction result of the model does not match the actual situation, or a new alarm type appears that cannot be accurately identified, the self-learning mechanism is triggered;

[0040] (5.2) Update the model based on the new information, use the mini-batch gradient descent optimization algorithm, and use the new data to adjust the model weights;

[0041] (5.3) In the process of updating the model, an incremental learning method is used to avoid completely retraining the model.

[0042] The system for constructing a power grid monitoring service event-based knowledge model based on deep learning of the present invention comprises:

[0043] The preprocessing module is used to collect alarm information data during the historical operation of the power grid, preprocess the power grid monitoring alarm text, and build a set of event-based rules for power grid monitoring alarm information;

[0044] The pre-training module is used to perform unsupervised pre-training on the pre-processed monitoring alarm text through the word2vec model to obtain a vectorized expression of the monitoring alarm text;

[0045] The testing module is used to establish a knowledge model of power grid monitoring business events based on convolutional neural networks, and use the power grid monitoring alarm text dataset to train and test the model and learn the alarm text features;

[0046] The recognition module is used to use the trained model to identify the monitoring alarm information, and through the knowledge model self-learning technology, it realizes the automatic expansion of the monitoring business event knowledge model for the incremental information of the power grid operation, thereby improving the robustness of the knowledge model.

[0047] When the processor of the present invention executes the program, a method for constructing an event-based knowledge model for power grid monitoring services based on deep learning is implemented.

[0048] The computer program of the present invention is designed to implement a method for constructing an event-based knowledge model for power grid monitoring services based on deep learning during runtime.

[0049] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: the present invention transforms the monitoring and alarm mode from the traditional passive monitoring mode to the active real-time perception mode. When a fault or abnormal event occurs in the power grid, the intelligent alarm system needs to be able to actively, quickly and accurately infer the type of power grid monitoring alarm event, providing timeliness guarantee for power grid fault diagnosis and accident analysis; based on the small batch gradient descent optimization algorithm and incremental learning method, the event recognition rules are improved to provide technical support for the automatic improvement of event rules; the correlation between event information and event rules is mined, and event features are autonomously generated to provide support for event definition and autonomous generation of event rules, which can realize the automatic expansion of the monitoring business event knowledge model. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the present invention;

[0051] Figure 2 This is the structural diagram of the convolutional neural network. DETAILED DESCRIPTION

[0052] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0053] like Figure 1 As shown, the present invention comprises the following steps:

[0054] (1) Collect alarm information data during the historical operation of the power grid, preprocess the power grid monitoring alarm text, and build a set of event-based rules for power grid monitoring alarm information; the preprocessing process includes: data cleaning, removing duplicate alarm texts, noise removal, filtering out special characters, garbled or meaningless symbols, removing stop words, removing common words that are not very meaningful for text analysis, and word segmentation, using the Jieba library to segment the alarm text into words or phrases.

[0055] (2) The historical alarm texts are processed based on the event-based rule set. The event-based rule set is constructed based on multiple dimensions such as voltage level, event type, equipment, and signal type. The alarm texts are mainly divided into the following four levels: accident level, abnormal level, displacement level, and notification level.

[0056] (3) The pre-processed monitoring alarm text is unsupervisedly pre-trained through the word2vec model to obtain the vectorized expression of the monitoring alarm text; the word2vec Skip-gram model is mainly used for word vector training, and the pre-processed monitoring alarm information features are input into the Word2vec model. The output feature vector is taken as the unit of the fault-triggered monitoring alarm text, and different correlation processing is performed on the feature vector in the text to obtain the vectorized expression of the monitoring alarm text. The main steps are as follows:

[0057] (3.1) The goal is to predict the context words based on the central word. Given a central word ω t , which tries to maximize the following probability: P(w t-j , w t-j+1 , L, w t+j-1 , w t+j |w t );

[0058] (3.2) For each context word ω t+i , the loss function of the Skip-gram model can be expressed as: Where T is the total number of words in the text, c is the context window size;

[0059] (3.3) In the Skip-gram model, p(w o |w i )(Predict word ω0 based on word ω i ) is usually calculated using the softmax function: where v i is the input word ω i The vector representation of u j is the output word ω j is a vector representation of , and V is the vocabulary size.

[0060] (4) Offline training: Establish a knowledge model of power grid monitoring business events based on convolutional neural networks, and use the power grid monitoring alarm text dataset to train and test the model and learn the alarm text features; the convolutional neural network structure is shown in the figure. Figure 2 As shown, the basic calculation process is as follows:

[0061] (4.1) The input layer is the warning text vector matrix generated by word2vec;

[0062] (4.2) The convolution layer is composed of several convolution kernel modules. Each convolution kernel performs convolution calculation on the text vector representation matrix T to preliminarily extract the local features of the text. The local feature formula obtained by the i-th neuron is: i =RELU(wgx i：i+r-1 + b);

[0063] (4.3) The pooling layer reduces the dimension of the features extracted by the convolutional layer and captures the optimal solution in the local features. The calculation formula is as follows: r =max(c1,c2,c3,·····,c n-r+1 );

[0064] (4.4) The fully connected layer connects the maximum pooling results obtained by the convolution kernel to obtain the final feature formula of the text as follows:

[0065] (4.5) The classification layer outputs the warning text classification results.

[0066] (5) Online application: Use the trained model to identify monitoring alarm information, and use the knowledge model self-learning technology to automatically expand the monitoring business event knowledge model based on the incremental information of power grid operation, thereby improving the robustness of the knowledge model:

[0067] (5.1) Continuously collect alarm information during the operation of the power grid. When there is new monitoring alarm information and incremental information, input this information into the trained model for prediction; if the prediction result of the model does not match the actual situation, or a new alarm type appears that cannot be accurately identified, the self-learning mechanism is triggered;

[0068] (5.2) Update the model according to the new information, use the small batch gradient descent optimization algorithm, and use the new data to adjust the model weights; the small batch gradient descent optimization algorithm randomly selects a small batch of data from the new monitoring alarm information and incremental information, rather than using all the data for calculation. After obtaining the gradient, the model weights are adjusted according to the pre-set learning rate. Through continuous repetition, each time a small batch of data is processed, the value of the loss function is gradually reduced, so that the model's prediction results are closer to the actual situation;

[0069] (5.3) In the process of updating the model, the incremental learning method is used to avoid completely retraining the model and improve learning efficiency. The incremental learning method mainly uses the incremental principal component analysis (IPCA) method to calculate the product of the covariance matrix of the new data and the existing principal component matrix, update the principal component matrix and eigenvalues ​​by solving the eigenvalue problem, use the updated principal component matrix to reduce the dimension of the new data, extract the main features, input the reduced-dimensional data into the model, and update and adjust the model.

[0070] The system for constructing a power grid monitoring service event-based knowledge model based on deep learning of the present invention comprises:

[0071] The preprocessing module is used to collect alarm information data during the historical operation of the power grid, preprocess the power grid monitoring alarm text, and build a set of event-based rules for power grid monitoring alarm information;

[0072] The pre-training module is used to perform unsupervised pre-training on the pre-processed monitoring alarm text through the word2vec model to obtain a vectorized expression of the monitoring alarm text;

[0073] The testing module is used to establish a knowledge model of power grid monitoring business events based on convolutional neural networks, and use the power grid monitoring alarm text dataset to train and test the model and learn the alarm text features;

[0074] The recognition module is used to use the trained model to identify the monitoring alarm information, and through the knowledge model self-learning technology, it realizes the automatic expansion of the monitoring business event knowledge model for the incremental information of the power grid operation, thereby improving the robustness of the knowledge model.

Claims

1. A method for constructing an event-based knowledge model for power grid monitoring services based on deep learning, characterized in that: The steps include: (1) Collect alarm information data during the historical operation of the power grid, pre-process the power grid monitoring alarm text, and build a set of event-based rules for power grid monitoring alarm information; (2) Using the event-based rule set to process the historical alarm information data in step (1), obtaining a data set of various labeled monitoring alarm events; (3) Perform unsupervised pre-training on the pre-processed monitoring alarm text through the word2vec model to obtain a vectorized expression of the monitoring alarm text; (4) Offline training: Establish a knowledge model of power grid monitoring business events based on convolutional neural networks, and use the power grid monitoring alarm text dataset to train and test the model and learn the alarm text features; (5) Online application: Use the trained model to identify monitoring alarm information and use the knowledge model self-learning technology to automatically expand the monitoring business event knowledge model based on the incremental information of power grid operation.

2. The method for constructing an event-based knowledge model for power grid monitoring services based on deep learning according to claim 1, characterized in that: The step (1) of preprocessing the power grid monitoring alarm text comprises: (1.1) Data cleaning, removing duplicate warning texts: avoid the same information being processed multiple times, reduce data redundancy, and handle missing values: if some key information in the warning text is missing, it needs to be supplemented or deleted according to the specific situation; (1.2) Noise removal, filtering out special characters, garbled characters or meaningless symbols; (1.3) Remove stop words, remove common words that are not very meaningful for text analysis; (1.4) Word segmentation: Use the Jieba library to segment the warning text into words or phrases.

3. The method for constructing an event-based knowledge model for power grid monitoring services based on deep learning according to claim 1, characterized in that: The step (2) includes: processing the historical alarm text based on the event-based rule set, constructing the event-based rule set according to multiple dimensions of voltage level, event type, equipment, and signal type, and mainly dividing the alarm text into accident level, abnormal level, displacement level, and notification level.

4. The method for constructing an event-based knowledge model for power grid monitoring services based on deep learning according to claim 1, characterized in that: The step (3) includes: using the Skip-gram model of word2vec to perform word vector training, inputting the preprocessed monitoring alarm information features into the Word2vec model, taking the monitoring alarm text triggered by the fault as the unit of the output feature vector, performing different correlation processing on the feature vector in the text, and obtaining the vectorized expression of the monitoring alarm text, including: (3.1) Predict the context words based on the central word. Given a central word ω t , maximizes the following probability: P(w t-j ,w t-j+1 ,L,w t+j-1 ,w t+j |w t ) (3.2) For each context word ω t+i , the loss function of the Skip-gram model is expressed as: Where T is the total number of words in the text, c is the context window size; (3.3) In the Skip-gram model, p(w o |w i ), predict word ω0 based on word ω i The calculation uses the softmax function: where v i is the input word ω i The vector representation of u j is the output word ω j is a vector representation of , and V is the vocabulary size.

5. The method for constructing an event-based knowledge model for power grid monitoring services based on deep learning according to claim 1, characterized in that: The main structure of the convolutional neural network in step (4) includes: input layer, convolution layer, pooling layer and fully connected layer.

6. The method for constructing an event-based knowledge model for power grid monitoring services based on deep learning according to claim 1, characterized in that: The training process of the convolutional neural network in step (4) includes: (4.1) The input layer is the warning text vector matrix generated by word2vec; (4.2) The convolution layer is composed of several convolution kernel modules. Each convolution kernel performs convolution calculation on the text vector representation matrix T to preliminarily extract the local features of the text. The local feature formula obtained by the i-th neuron is: c i =RELU(wgx i：i+r-1 +b) (4.3) The pooling layer reduces the dimension of the features extracted by the convolutional layer and captures the optimal solution in the local features. The calculation formula is as follows: With r =max(c1,c2,c3,……,c n-r+1 ) (4.4) The connection layer connects the maximum pooling results obtained by the convolution kernel to obtain the final feature formula of the text as follows: (4.5) The classification layer outputs the warning text classification results.

7. The method for constructing an event-based knowledge model for power grid monitoring services based on deep learning according to claim 1, characterized in that: The step (5) comprises: (5.1) Continuously collect alarm information during the operation of the power grid. When there is new monitoring alarm information and incremental information, input this information into the trained model for prediction. If the prediction result of the model does not match the actual situation, or a new alarm type appears that cannot be accurately identified, the self-learning mechanism is triggered; (5.2) Update the model based on the new information, use the mini-batch gradient descent optimization algorithm, and use the new data to adjust the model weights; (5.3) In the process of updating the model, an incremental learning method is used to avoid completely retraining the model.

8. A system for constructing event-based knowledge model of power grid monitoring business based on deep learning, characterized in that: include: The preprocessing module is used to collect alarm information data during the historical operation of the power grid, preprocess the power grid monitoring alarm text, and build a set of event-based rules for power grid monitoring alarm information; The pre-training module is used to perform unsupervised pre-training on the pre-processed monitoring alarm text through the word2vec model to obtain a vectorized expression of the monitoring alarm text; The testing module is used to establish a knowledge model of power grid monitoring business events based on convolutional neural networks, and use the power grid monitoring alarm text dataset to train and test the model and learn the alarm text features; The recognition module is used to use the trained model to identify the monitoring alarm information, and through the knowledge model self-learning technology, it realizes the automatic expansion of the monitoring business event knowledge model for the incremental information of the power grid operation, thereby improving the robustness of the knowledge model.

9. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the method for constructing an event-based knowledge model for power grid monitoring services based on deep learning according to any one of claims 1 to 7.

10. A storage medium storing a computer program, characterized in that: The computer program is designed to implement the method for constructing an event-based knowledge model for power grid monitoring services based on deep learning according to any one of claims 1 to 7 when running.

Images