CN111274395A - Power grid monitoring alarm event identification method based on convolution and long-short term memory network - Google Patents

Power grid monitoring alarm event identification method based on convolution and long-short term memory network Download PDF

Info

Publication number
CN111274395A
CN111274395A CN202010058261.4A CN202010058261A CN111274395A CN 111274395 A CN111274395 A CN 111274395A CN 202010058261 A CN202010058261 A CN 202010058261A CN 111274395 A CN111274395 A CN 111274395A
Authority
CN
China
Prior art keywords
monitoring alarm
alarm information
convolution
event
information
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202010058261.4A
Other languages
Chinese (zh)
Other versions
CN111274395B (en
Inventor
臧海祥
白子瑜
程礼临
孙国强
卫志农
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hohai University HHU
Original Assignee
Hohai University HHU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hohai University HHU filed Critical Hohai University HHU
Priority to CN202010058261.4A priority Critical patent/CN111274395B/en
Publication of CN111274395A publication Critical patent/CN111274395A/en
Application granted granted Critical
Publication of CN111274395B publication Critical patent/CN111274395B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/088—Non-supervised learning, e.g. competitive learning

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Alarm Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a power grid monitoring alarm event identification method based on convolution and long-short term memory network, which generates information vector by historical monitoring alarm information and time stamp in a power grid monitoring system, extracts event samples from the collected historical monitoring alarm information and constructs an alarm event sample library; secondly, establishing a deep learning identification model based on a combination of a long-short term memory network and a convolutional neural network, and training the model by utilizing an alarm event sample; and finally, recognizing the monitoring alarm information by using the trained deep learning model, and outputting the event type with the maximum probability as a recognition result. The method combines the excellent performance of the long-term and short-term memory network in processing the time sequence problem and the convolutional neural network in mining the local characteristics of the short text, establishes a combined model, can realize the quick identification of the power grid alarm event, effectively reduces the screen monitoring pressure of monitoring service personnel, and improves the working efficiency of daily monitoring and accident abnormity handling.

Description

Power grid monitoring alarm event identification method based on convolution and long-short term memory network
Technical Field
The invention belongs to the intelligent alarm control technology of an electric power system, and particularly relates to a power grid monitoring alarm event identification method based on convolution and long-short term memory networks.
Background
With the continuous expansion of the scale of the power grid, higher requirements are provided for quick response of power grid equipment faults and timely recovery of the power grid operation mode by regulators, so that the intelligent level of power grid equipment operation monitoring is improved, the automatic identification of power grid alarm events is realized, and the method has important significance for improving the working efficiency of daily monitoring and accident abnormity handling.
The power grid monitoring alarm information is used as Chinese text data and is an important data basis for monitoring the power grid operation state by regulation and control personnel. With the scale enlargement of power grid equipment and the improvement of intelligent monitoring level, the electric power data is explosively increased, the number of monitoring alarm information shows a geometric level increase trend, and the collected information is all displayed according to a time sequence without any reasoning judgment processing. The power grid regulation and control personnel need to distinguish, analyze and feed back each piece of information one by one, important alarm information is easily omitted, accurate identification cannot be made in a short time, and the condition of equipment failure or abnormal missing judgment and misjudgment occurs. The method cannot meet the higher requirements of the power grid monitoring service under the current situation. The invention applies a long-short term memory network and a convolution neural network in a deep learning algorithm to identify the power grid monitoring alarm event.
Traditional machine learning models such as Logistic regression, support vector machines, random forest algorithms and the like are suitable for processing scenes with small sample size and generally do not have feature processing capability. Therefore, when these algorithms are applied, feature extraction needs to be performed on the raw data, which increases the complexity of the modeling process. The deep learning algorithm processes data in a layer-by-layer training mode, can obtain high-level feature representation of an original input variable set, improves prediction and classification accuracy, and is widely applied to feature processing problems and big data scenes. In the face of the requirement of processing massive power grid operation warning information, deep learning can more fully learn and monitor the sample characteristics of the big data. The long-term and short-term memory network has a powerful function of processing sequences with time correlation, monitoring alarm information triggered by a power grid monitoring alarm event continuously occurs in a short time, information of the whole event is arranged according to the occurrence time in sequence, a time sequence relation is formed, and each piece of alarm information is used as a time step to extract time sequence characteristics of the whole event. Meanwhile, in terms of statement expression meaning, a plurality of adjacent alarm messages contain important characteristics of alarm events, and the CNN has the characteristics of local perception and excellent feature extraction performance and can mine the relevance features between adjacent monitoring alarm messages. And the convolutional neural network adopts a local perception and weight sharing mode, thereby greatly reducing the network parameter and relieving the overfitting problem of the model.
Disclosure of Invention
The purpose of the invention is as follows: the invention provides a power grid monitoring alarm event identification method based on convolution and long-short term memory networks, aiming at the problems that manual judgment in the existing power grid monitoring alarm is easy to miss judgment and misjudgment and the identification efficiency is not high.
The technical scheme is as follows: a power grid monitoring alarm event identification method based on convolution and long-short term memory network comprises the following steps:
(1) collecting historical monitoring alarm information and the time scale of each piece of alarm information in a power grid monitoring system, wherein all transformer substations and line names contained in the alarm information form a training data set required by a power grid monitoring alarm event identification model;
(2) carrying out data preprocessing on historical monitoring alarm information, carrying out unsupervised training on the monitoring alarm information through a word2vec model, and generating an information vector containing signal characteristics;
(3) extracting a monitoring alarm information set from the collected historical monitoring alarm information according to a sliding time window, determining the event type of the alarm information set and the sign word of the alarm event, obtaining various labeled monitoring alarm event samples, and constructing an alarm event sample library;
(4) establishing a deep learning identification model based on a combination of a long-term and short-term memory network and a convolutional neural network, setting various hyper-parameters, establishing a target function and selecting an optimization algorithm, performing iterative training on the model by using samples in a monitoring alarm event sample library, and gradually iteratively updating the parameters of the identification model by calculating the gradient of a loss function;
(5) and identifying the monitoring alarm information by using the trained deep learning model, and outputting an event type with the maximum probability as an alarm event identification result.
Further, the specific processes of data preprocessing and generation vector generation of the historical monitoring alarm information in the step (2) are as follows:
(21) word segmentation and stop word
(211) Updating the power word bank, and importing the transformer substation name and the line name derived from the historical monitoring alarm information into the power word bank as a power dictionary used for word segmentation;
(212) performing initial word segmentation by adopting an accurate mode of a Jieba word segmentation tool, and generating monitoring alarm information which is composed of a series of Chinese words and is sequenced according to time;
(213) establishing a stop word list, and removing the stopped electric power words in the monitoring alarm information to realize data cleaning;
(22) monitoring alarm information vectorization
(221) Carrying out unsupervised training on the monitoring alarm information by using a word2vec model, and calculating to generate distributed word vector representation of each word in the alarm information;
(222) calculating the vector of all words in the monitoring alarm information when the alarm event occurs, and averaging to obtain the vector representation of the monitoring alarm information with the same dimension as the word vector, wherein the calculation formula is as follows:
Figure BDA0002373532610000031
in the formula: d represents a piece of monitoring alarm information; word _ num represents the number of words in d; t represents a word in the monitoring alarm information; word2vec (t) represents a vector of t; word2vec _ sum (d) represents a distributed vector representation of a piece of monitoring alarm information.
The specific process of establishing the deep learning identification model based on the combination of the long-short term memory network and the convolutional neural network in the step (4) is as follows:
(41) the input to the LSTM layer is an alarm event sample, denoted X ═ X1,x2,…,xnIn which xiIs a distributed vector representation of the monitoring alarm information, i ═ 1,2, … n; n is the number of monitoring alarm information contained in the alarm event sample;
the information quantity calculation formula stored in the memory unit in the network input at the current moment in the input gate is as follows:
it＝σ(wxixt+whiht-1+bi)
in the formula: i.e. itIs the output of the input gate; x is the number oftAnd ht-1Current input and previous hidden layer output respectively; w is axiAnd whiAre respectively input xtAnd ht-1The weight of (c); biIs the offset of the input gate; σ denotes the sigmoid activation function.
The input gate outputs a temporary memory unit c'tThe calculation formula of (a) is as follows:
c′t＝tanh(wxcxt+whcht-1+bc)
in the formula: w is axcAnd whcAre respectively input xtAnd ht-1The weight of (c); bcIs a temporary memory unit c'tBias of (3);
the formula for calculating the information quantity of the memory unit reserved to the current moment by the memory unit at the previous moment in the forgetting gate is as follows:
ft＝σ(wxfxt+whfht-1+bf)
in the formula: f. oftIs the output of the forget gate; w is axfAnd whfAre respectively input xtAnd ht-1The weight of (c); bfIs the bias of the forgetting gate. The second part is obtained by the input of the current moment acting on the input gateA temporary memory unit;
the current memory cell calculation formula is as follows:
ct＝ft·ct-1+it·c′t
in the formula: c. Ct-1Is the output value of the memory cell at the previous moment;
the output expressions of the output gates and the hidden layer are as follows:
ot＝σ(wxoxt+whoht-1+bo)
ht＝ot·tanh(ct)
in the formula: otAnd htRespectively representing the output gate and the output of the current hidden layer; w is axo，whoAnd boAre each xtWeight of (a), ht-1Weight sum o oftIs used to control the bias of (1).
(42) The hidden layer output matrix H on each time step in the long-short term memory network belongs to Rn×kInputting the time sequence length of an alarm event sample into a convolutional layer to extract local characteristics of alarm information, wherein n is the time sequence length of the alarm event sample and represents the quantity of monitoring alarm information contained in an event, k is the vector dimension of an output value, a convolutional matrix W belonging to R and having the row number of H and the column number same as that of an input layer matrix H is adoptedh×kPerforming convolution operation, wherein the expression of the convolution result is as follows:
ri＝W·Hi:i+h-1
in the formula: hi:i+h-1Representing a sub-matrix consisting of the ith to (i + H-1) th rows of the matrix H; the operational symbol "·" is a dot product operation, which means that elements at the same position of two matrices are multiplied and then summed;
the result of each convolution after the nonlinear operation is:
ci＝ReLU(ri+bi)
in the formula: biIs a bias term; ReLU is an activation function, and the calculation formula is as follows:
ReLU＝(0,x)
all the results are sequentially arranged and stacked to obtain a convolution layerFeature vector c ∈ Rn-h+1N-h +1 is the number of convolution operations performed in common;
(43) the pooling layer performs dimensionality reduction on the feature vectors through a down-sampling rule, and takes the maximum value in each feature vector c obtained through convolution operation as a feature value through a maximum pooling method, and the feature value is expressed as follows:
cmax＝max{c}
all the characteristic values extracted by the pooling operation of different characteristic vectors are spliced to form a pooled layer output vector q ∈ RvWhere v is m · k, m is the number of classes of convolution windows, and k is the number of convolution windows of each class;
(44) inputting the pooling layer vector q into a Softmax classifier to output the probability of each alarm event category, selecting the category with the maximum probability as the identification result of the input monitoring alarm information of the section, wherein the expression is as follows:
p＝softmax(Wq·q+bq)
in the formula: wqIs the weight corresponding to event q; bqIs the bias term corresponding to event q.
Furthermore, parameters of the recognition model are gradually updated in an iterative mode by calculating gradient of a loss function in the deep learning recognition model training stage based on the combination of the long-short term memory network and the convolutional neural network, the loss function is set as a cross entropy loss function, an Adam algorithm is adopted for optimization to enable a model objective function to be converged to the minimum, the optimal weight and bias term are obtained, and meanwhile, a Dropout strategy is adopted for constraining model parameters to prevent an overfitting phenomenon from occurring in model training.
And (42) extracting more associated features hidden among local information by adopting a multi-granularity convolution window, forming convolution windows of different categories by changing the row number of the convolution matrix, and fully selecting the number of the convolution windows in each category to avoid the loss of information features in the training process.
Has the advantages that: compared with the prior art, the power grid monitoring alarm event identification method based on the convolution and long-short term memory network can convert monitoring alarm information into information vectors, realize identification of power grid monitoring alarm events based on texts and push event identification results. The monitoring mode that the existing power grid monitoring service depends on the one-by-one response of alarm information can be effectively changed, the problems of low artificial processing efficiency, high misjudgment rate and the like are solved, the screen monitoring pressure of power grid regulation and control personnel is reduced, and the working efficiency of daily monitoring and accident abnormity handling is improved.
Drawings
FIG. 1 is a diagram of a recognition model based on convolution and long-short term memory neural networks;
FIG. 2 is a diagram of a convolutional neural network architecture;
FIG. 3 is a diagram of a long-short term memory neural network.
Detailed Description
For the purpose of explaining the technical solution disclosed in the present invention in detail, the following description is further made with reference to the accompanying drawings and specific embodiments.
The method aims at the problems that the quantity of alarm information accessed to a regulation and control system is continuously increased, the acquired information is all displayed according to a time sequence and is not subjected to any inference and judgment processing, important information is easily omitted in a monitoring mode of one-by-one response of the existing power grid monitoring service, and the alarm event is missed and judged wrongly. The invention discloses a power grid monitoring alarm event identification method based on convolution and long-short term memory networks, which can identify input alarm information.
Example 1
Referring to fig. 1,2 and 3, the method of the present invention proceeds as follows:
the method comprises the steps that firstly, historical monitoring alarm information in a power grid monitoring system and a time scale of each piece of alarm information are collected, and all transformer substations and line names contained in the alarm information form a training data set required by a power grid monitoring alarm event identification model;
and secondly, performing data preprocessing on historical monitoring alarm information, performing unsupervised training on the monitoring alarm information through a word2vec model, and generating an information vector containing signal characteristics, wherein the specific process is as follows:
(1) word segmentation and stop word
And updating the power word stock, and referring to the collected power word stock through data, and importing the substation name and the line name which are derived from the historical monitoring alarm information into the word stock to be used as a power dictionary for word segmentation. And performing initial word segmentation by adopting an accurate mode of a Jieba word segmentation tool to generate monitoring alarm information consisting of a series of Chinese words and expressions in time sequence. And a stop word list is established, meaningless words in the alarm information are removed, and data cleaning is realized to improve the later training effect.
(2) Monitoring alarm information vectorization
The word2vec model is used for carrying out unsupervised training on the monitoring alarm information, distributed word vector representation of each word in the alarm information is calculated and generated, and the problem of high-dimensional sparse characteristics of the traditional model is solved. When an alarm event occurs, the monitoring alarm information is expressed in the form of a statement, so that the vectors of all words in one piece of monitoring alarm information are averaged to obtain the vector representation of the monitoring alarm information with the same dimension as the word vector, and the calculation formula is as follows:
Figure BDA0002373532610000061
in the formula: d represents a piece of monitoring alarm information; word _ num represents the number of words in d; t represents a word in the monitoring alarm information; word2vec (t) represents a vector of t; word2vec _ sum (d) represents a distributed vector representation of a piece of monitoring alarm information. And training and calculating to obtain the dimension of the alarm information vector of 300.
Thirdly, extracting event samples from the collected historical monitoring alarm information and constructing an alarm event sample library, wherein the specific implementation mode is as follows:
and taking the monitoring alarm information with the key word of 'switching off' as a mark, and extracting discrete monitoring alarm information of the same transformer substation or line within 15s before and after the information to form an alarm information set. After each alarm event is processed, a control worker writes a scheduling log to record occurrence time, event reasons, processing flows and event types, and also records various events of triggering or fault alarm in the existing power system to establish the scheduling log to improve operation management. And determining the event type of the alarm information set by contrasting the scheduling log to form 9 types of monitoring alarm event samples, wherein the 9 types of monitoring alarm event samples comprise bus faults, transient faults (successful reclosing), permanent faults (failed reclosing), permanent faults (not actuated reclosing), main transformer electrical quantity faults, main transformer body heavy gas faults, main transformer pressure regulating heavy gas faults and capacitive reactance device faults, namely grounding fault, and an alarm event sample library is constructed.
Fourthly, establishing a deep learning identification model based on the combination of the long-short term memory network and the convolutional neural network to monitor alarm information, and training the model by utilizing an alarm event sample, wherein the concrete process is as follows:
(1) the input to the LSTM layer is an alarm event sample, denoted X ═ X1,x2,…,xnIn which xiIs a distributed vector representation of the monitoring alarm information, i ═ 1,2, … n; n is the number of monitoring alarm information contained by the alarm event sample. Because the monitoring alarm information in the event is arranged according to the time sequence, each vector xiAll represent external input of the LSTM unit at a time step, and extract the time sequence characteristics of the whole monitoring alarm information sequence.
The input gate realizes the control of the input information at the current moment and determines how much information in the network input at the current moment is stored in the memory unit, and the calculation formula is as follows:
it＝σ(wxixt+whiht-1+bi)
in the formula: i.e. itIs the output of the input gate; x is the number oftAnd ht-1Current input and previous hidden layer output respectively; w is axiAnd whiAre respectively input xtAnd ht-1The weight of (c); biIs the offset of the input gate. σ denotes the sigmoid activation function.
The input gate also outputs a temporary memory unit c'tThe calculation formula is as follows:
c′t＝tanh(wxcxt+whcht-1+bc)
in the formula: w is axcAnd whcAre respectively input xtAnd ht-1The weight of (c); bcIs a temporary memory unit c'tIs offset from
The forgetting gate realizes the control of the memory unit at the previous moment and determines how much information of the memory unit at the previous moment is reserved in the memory unit at the current moment, and the calculation formula is as follows:
ft＝σ(wxfxt+whfht-1+bf)
in the formula: f. oftIs the output of the forget gate; w is axfAnd whfAre respectively input xtAnd ht-1The weight of (c); bfIs the bias of the forgetting gate. The second part is a temporary memory unit obtained by the input of the current time acting on the input gate.
The current moment memory cell calculation formula is:
ct＝ft·ct-1+it·c′t
in the formula: c. Ct-1Is the output value of the memory cell at the previous moment.
The output gate is input by the current moment, and the current moment memory unit and the previous moment hidden layer output are jointly determined:
ot＝σ(wxoxt+whoht-1+bo)
ht＝ot·tanh(ct)
in the formula: otAnd htRespectively representing the output gate and the output of the current hidden layer; w is axo，whoAnd boAre each xtWeight of (a), ht-1Weight sum o oftIs used to control the bias of (1).
(2) The hidden layer output matrix H on each time step in the long-short term memory network belongs to Rn×kAnd inputting the time sequence length of the alarm event sample, namely the quantity of monitoring alarm information contained in the event, and k is the vector dimension of an output value. Comparing the recognition accuracy rates of the output values with the vector dimensions of 64, 128 and 256 respectively, and finding out the current lengthWhen the dimension of the output vector of the phase memory network is 128, the recognition accuracy is highest. Adopting a convolution matrix W belonging to R and having the row number H and the column number same as that of the input layer matrix Hh×kPerforming convolution operation, wherein the convolution result is as follows:
ri＝W·Hi:i+h-1
in the formula: hi:i+h-1Representing a sub-matrix consisting of the ith to (i + H-1) th rows of the matrix H; the operational notation "·" is a dot product operation, i.e., multiplying and then summing the elements at the same position of two matrices. The result of each convolution after the nonlinear operation is:
ci＝ReLU(ri+bi)
in the formula: biIs a bias term; ReLU is an activation function, and the calculation formula is as follows:
ReLU＝(0,x)
all the results are sequentially arranged and accumulated to obtain a feature vector c epsilon R of the convolution layern-h+1N-h +1 is the number of convolution operations performed in common.
In order to avoid the loss of information characteristics in the training process, a multi-granularity convolution window is adopted to extract more relevant characteristics hidden among local information. Different types of convolution windows are formed by changing the line number of a convolution matrix, according to analysis of a monitoring alarm event text, 2-3 adjacent monitoring alarm information are found to have a local relevance characteristic, 3 types of convolution kernels are set in consideration of accompanying information possibly having interference, and the window sizes are 3, 4 and 5 respectively. The number of the convolution windows in each category is selected to be sufficient, and the number is 100.
(3) The pooling layer reduces the dimension of the feature vector through a down-sampling rule, improves the calculation efficiency of the classifier, and simultaneously realizes the further extraction of the alarm event features. Taking the maximum value in the feature vector c obtained by each convolution operation as a feature value by adopting a maximum pooling method (max-pooling):
cmax＝max{c}
all the characteristic values extracted by the pooling operation of different characteristic vectors are spliced to form a pooled layer output vector q ∈ RvWhere v is m.k, m being the class of the convolution windowThe number of classes, k, is the number of convolution windows of each class.
(4) Inputting the pooling layer vector q into a Softmax classifier to output the probability of each alarm event category, and selecting the category with the maximum probability as the identification result of the input monitoring alarm information of the section:
p＝softmax(Wq·q+bq)
in the formula: wqIs the weight corresponding to event q; bqIs the bias term corresponding to event q.
(5) In the model training stage, an Adam algorithm is adopted for training, namely, a model objective function is converged to the minimum through continuous iteration, and the optimal weight and bias term are obtained. In addition, in order to prevent the overfitting phenomenon in the training of the model, a Dropout strategy is adopted to constrain the model parameters, namely, a part of trained parameters are randomly selected to be discarded every time of updating, and the discarding rate is 0.5.
And fifthly, recognizing the monitoring alarm information by using the trained deep learning model, and outputting the event type with the maximum probability as a recognition result. In the application process, the model can still be correctly identified under the condition of partial information error or information loss, and has strong robustness and certain fault tolerance. Meanwhile, the deep learning model has self-learning capability, and with the continuous expansion of a sample library, the newly defined event type is subjected to iterative training, the parameter structure is adjusted and perfected, and the self stability and the identification accuracy are enhanced.
Example 2
By taking more than 1400 historical monitoring alarm information of a certain municipal power grid company 2016 and 2017 as an original corpus, 9 types of alarm event samples are extracted from the original corpus to train and test the recognition model. 90% of the alarm event samples of each type are used as a training set, 10% are used as a test set, and the types of the alarm events and the number of samples of each type are shown in table 1.
TABLE 1 alarm event sample number
Figure BDA0002373532610000091
In the classification task of events, the classification result of the recognition model is generally expressed by a confusion matrix, and the meanings of two classification confusion matrices are shown in table 2.
TABLE 2 confusion matrix in event recognition
Figure BDA0002373532610000101
The confusion matrix divides all events into four types according to the actual attribution and the identification attribution, four indexes of Accuracy (Accuracy), Precision (Precision), Recall (Recall) and F1 are defined to measure the identification effect of the model, the value ranges of the four indexes are all [0,1], and the closer the value is to 1, the better the identification effect of the model is.
The calculation formulas are respectively as follows:
Figure BDA0002373532610000102
Figure BDA0002373532610000103
Figure BDA0002373532610000104
Figure BDA0002373532610000105
in order to verify the classification effect of the proposed LSTM-CNN combined recognition model, several single deep learning models and a typical machine learning model are selected for comparison verification. The deep learning model selects CNN, LSTM and Bidirectional Long short-term Memory network (BilSTM), and the machine learning model selects Support Vector Machines (SVM), Logistic Regression (LR) and Random Forest (RF) models. The recognition effect of alarm events for different models is shown in table 3.
Table 3 comparison of the recognition results of the models herein with other models
Figure BDA0002373532610000106
As can be seen from table 3, the minimum CNN accuracy in the four deep learning models is 92.69%, and the maximum random forest model accuracy in the three machine learning models is 91.18%, which indicates that the recognition effect of the deep learning model is better than that of the machine learning model. The LSTM-CNN combined recognition model is better than other deep learning models in each index, and the accuracy rate is respectively 1.69% and 5.61% higher than that of a single LSTM and CNN. The accuracy, recall rate and F1 values are respectively 1.68%, 1.69% and 2.05% higher than that of single LSTM and 5.58%, 5.61% and 6% higher than that of single CNN; meanwhile, the highest identification accuracy rate in other models is BiLSTM, which reaches 96.75%, and the LSTM-CNN combined identification model is still 1.55% higher than the BiLSTM.
The LSTM-CNN combined recognition model is practically applied, monitoring alarm events triggered in an information intensive period 13:27-13:31 when the typhoon passes through a certain city in 8, 8 and 17 months in 2018 and triggered in 17 days in Wimbiya are used as application objects, 6 monitoring alarm events suitable for the model are extracted from 4146 pieces of monitoring alarm information in the period, all the monitoring alarm events are correctly recognized by the model, and the recognition time is 0.5 s.

Claims (5)

1. A power grid monitoring alarm event identification method based on convolution and long-short term memory network is characterized by comprising the following steps:
(1) collecting historical monitoring alarm information and the time scale of each piece of alarm information in a power grid monitoring system, wherein all transformer substations and line names contained in the alarm information form a training data set required by a power grid monitoring alarm event identification model;
(2) carrying out data preprocessing on historical monitoring alarm information, carrying out unsupervised training on the monitoring alarm information through a word2vec model, and generating an information vector containing signal characteristics;
(3) extracting a monitoring alarm information set from the collected historical monitoring alarm information according to a sliding time window, determining the event type of the alarm information set and the sign word of the alarm event, obtaining various labeled monitoring alarm event samples, and constructing an alarm event sample library;
(4) establishing a deep learning identification model based on a combination of a long-term and short-term memory network and a convolutional neural network, setting various hyper-parameters, establishing a target function and selecting an optimization algorithm, performing iterative training on the model by using samples in a monitoring alarm event sample library, and gradually iteratively updating the parameters of the identification model by calculating the gradient of a loss function;
(5) and identifying the monitoring alarm information by using the trained deep learning model, and outputting an event type with the maximum probability as an alarm event identification result.
2. The convolution and long-short term memory network based power grid monitoring alarm event recognition method according to claim 1, wherein the method comprises the following steps: the specific processes of data preprocessing and generation vector generation of the historical monitoring alarm information in the step (2) are as follows:
(21) word segmentation and stop word
(211) Updating the power word bank, and importing the transformer substation name and the line name derived from the historical monitoring alarm information into the power word bank as a power dictionary used for word segmentation;
(212) performing initial word segmentation according to the accurate mode of the Jieba word segmentation tool, and generating monitoring alarm information which is composed of a series of Chinese words and is sorted according to time;
(213) establishing a stop word list, and removing the stopped electric power words in the monitoring alarm information to realize data cleaning;
(22) monitoring alarm information vectorization
(221) Carrying out unsupervised training on the monitoring alarm information by using a word2vec model, and calculating to generate distributed word vector representation of each word in the alarm information;
(222) calculating the vector of all words in the monitoring alarm information when the alarm event occurs, and averaging to obtain the vector representation of the monitoring alarm information with the same dimension as the word vector, wherein the calculation formula is as follows:
Figure FDA0002373532600000021
in the formula: d represents a piece of monitoring alarm information; word _ num represents the number of words in d; t represents a word in the monitoring alarm information; word2vec (t) represents a vector of t; word2vec _ sum (d) represents a distributed vector representation of a piece of monitoring alarm information.
3. The convolution and long-short term memory network based power grid monitoring alarm event recognition method according to claim 1, wherein the method comprises the following steps: the specific process of establishing the deep learning identification model based on the combination of the long-short term memory network and the convolutional neural network in the step (4) is as follows:
(41) the input to the LSTM layer is an alarm event sample, denoted X ═ X1,x2,…,xnIn which xiIs a distributed vector representation of the monitoring alarm information, i ═ 1,2, … n; n is the number of monitoring alarm information contained in the alarm event sample;
the information quantity calculation formula stored in the memory unit in the network input at the current moment in the input gate is as follows:
it＝σ(wxixt+whiht-1+bi)
in the formula: i.e. itIs the output of the input gate; x is the number oftAnd ht-1Current input and previous hidden layer output respectively; w is axiAnd whiAre respectively input xtAnd ht-1The weight of (c); biIs the offset of the input gate; sigma represents a sigmoid activation function;
the input gate outputs a temporary memory unit c'tThe calculation formula of (a) is as follows:
c′t＝tanh(wxcxt+whcht-1+bc)
in the formula: w is axcAnd whcAre respectively input xtAnd ht-1The weight of (c); bcIs a temporary memory unit c'tBias of (3);
the formula for calculating the information quantity of the memory unit reserved to the current moment by the memory unit at the previous moment in the forgetting gate is as follows:
ft＝σ(wxfxt+whfht-1+bf)
in the formula: f. oftIs the output of the forget gate; w is axfAnd whfAre respectively input xtAnd ht-1The weight of (c); bfIs the bias of the forgetting gate; the second part is a temporary memory unit obtained by inputting and acting on the input gate at the current moment;
the current memory cell calculation formula is as follows:
ct＝ft·ct-1+it·c′t
in the formula: c. Ct-1Is the output value of the memory cell at the previous moment;
the output expressions of the output gates and the hidden layer are as follows:
ot＝σ(wxoxt+whoht-1+bo)
ht＝ot·tanh(ct)
in the formula: otAnd htRespectively representing the output gate and the output of the current hidden layer; w is axo，whoAnd boAre each xtWeight of (a), ht-1Weight sum o oftBias of (3);
(42) the hidden layer output matrix H on each time step in the long-short term memory network belongs to Rn×kInputting the time sequence length of an alarm event sample into a convolutional layer to extract local characteristics of alarm information, wherein n is the time sequence length of the alarm event sample and represents the quantity of monitoring alarm information contained in an event, k is the vector dimension of an output value, a convolutional matrix W belonging to R and having the row number of H and the column number same as that of an input layer matrix H is adoptedh ×kPerforming convolution operation, wherein the expression of the convolution result is as follows:
ri＝W·Hi:i+h-1
in the formula: hi:i+h-1Representing a sub-matrix consisting of the ith to (i + H-1) th rows of the matrix H; the operational symbol ". cndot.Multiplying elements at the same position of each matrix and then summing;
the result of each convolution after the nonlinear operation is:
ci＝ReLU(ri+bi)
in the formula: biIs a bias term; ReLU is an activation function, and the calculation formula is as follows:
ReLU＝(0,x)
all the results are sequentially arranged and accumulated to obtain a feature vector c epsilon R of the convolution layern-h+1N-h +1 is the number of convolution operations performed in common;
(43) the pooling layer performs dimensionality reduction on the feature vectors through a down-sampling rule, and takes the maximum value in each feature vector c obtained through convolution operation as a feature value through a maximum pooling method, and the feature value is expressed as follows:
cmax＝max{c}
all the characteristic values extracted by different characteristic vectors through pooling operation are spliced to form a pooled layer output vector q E RvWhere v is m · k, m is the number of classes of convolution windows, and k is the number of convolution windows of each class;
(44) inputting the pooling layer vector q into a Softmax classifier to output the probability of each alarm event category, selecting the category with the maximum probability as the identification result of the input monitoring alarm information of the section, wherein the expression is as follows:
p＝softmax(Wq·q+bq)
in the formula: wqIs the weight corresponding to event q; bqIs the bias term corresponding to event q.
4. The convolution and long-short term memory network based grid monitoring alarm event recognition method according to claim 3, wherein the method comprises the following steps: in the deep learning recognition model training stage based on the combination of the long-short term memory network and the convolutional neural network, parameters of a recognition model are gradually updated in an iterative mode by calculating the gradient of a loss function, the loss function is set as a cross entropy loss function, optimization is carried out based on an Adam algorithm to enable a model target function to be converged to the minimum, the optimal weight and bias term are obtained, and meanwhile, a Dropout strategy is adopted to restrain model parameters to prevent an overfitting phenomenon from occurring in model training.
5. The convolution and long-short term memory network based grid monitoring alarm event recognition method according to claim 3, wherein the method comprises the following steps: and (42) extracting the associated features hidden in the local information by adopting a multi-granularity convolution window, and forming different types of convolution windows by changing the row number of the convolution matrix so as to avoid the loss of the information features in the training process.
CN202010058261.4A 2020-01-19 2020-01-19 Power grid monitoring alarm event identification method based on convolution and long-short term memory network Active CN111274395B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010058261.4A CN111274395B (en) 2020-01-19 2020-01-19 Power grid monitoring alarm event identification method based on convolution and long-short term memory network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010058261.4A CN111274395B (en) 2020-01-19 2020-01-19 Power grid monitoring alarm event identification method based on convolution and long-short term memory network

Publications (2)

Publication Number Publication Date
CN111274395A true CN111274395A (en) 2020-06-12
CN111274395B CN111274395B (en) 2021-11-12

Family

ID=71003164

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010058261.4A Active CN111274395B (en) 2020-01-19 2020-01-19 Power grid monitoring alarm event identification method based on convolution and long-short term memory network

Country Status (1)

Country Link
CN (1) CN111274395B (en)

Cited By (27)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111865947A (en) * 2020-07-08 2020-10-30 浙江大学 Method for generating abnormal data of power terminal based on transfer learning
CN111970489A (en) * 2020-08-05 2020-11-20 北京必可测科技股份有限公司 Intelligent monitoring management method and system based on man-machine two-way
CN112001482A (en) * 2020-08-14 2020-11-27 佳都新太科技股份有限公司 Vibration prediction and model training method and device, computer equipment and storage medium
CN112187413A (en) * 2020-08-28 2021-01-05 中国人民解放军海军航空大学航空作战勤务学院 SFBC (Small form-factor Block code) identifying method and device based on CNN-LSTM (convolutional neural network-Link State transition technology)
CN112418014A (en) * 2020-11-09 2021-02-26 南京信息工程大学滨江学院 Modulation signal identification method based on wavelet transformation and convolution long-short term memory neural network
CN112487406A (en) * 2020-12-02 2021-03-12 中国电子科技集团公司第三十研究所 Network behavior analysis method based on machine learning
CN112507720A (en) * 2020-11-12 2021-03-16 西安交通大学 Graph convolution network root identification method based on causal semantic relation transfer
CN112598023A (en) * 2020-12-02 2021-04-02 国网青海省电力公司海东供电公司 Intelligent identification and automatic mapping method and device for power grid monitoring alarm information
CN112838946A (en) * 2020-12-17 2021-05-25 国网江苏省电力有限公司信息通信分公司 Method for constructing intelligent sensing and early warning model based on communication network faults
CN113032917A (en) * 2021-03-03 2021-06-25 安徽大学 Electromechanical bearing fault detection method based on generation countermeasure and convolution cyclic neural network and application system
CN113065352A (en) * 2020-06-29 2021-07-02 国网浙江省电力有限公司杭州供电公司 Operation content identification method for power grid dispatching work text
CN113157859A (en) * 2021-04-06 2021-07-23 北京理工大学 Event detection method based on upper concept information
CN113239638A (en) * 2021-06-29 2021-08-10 百维金科(上海)信息科技有限公司 Overdue risk prediction method for optimizing multi-core support vector machine based on dragonfly algorithm
CN113315790A (en) * 2021-07-29 2021-08-27 湖南华菱电子商务有限公司 Intrusion flow detection method, electronic device and storage medium
CN113487010A (en) * 2021-05-21 2021-10-08 国网浙江省电力有限公司杭州供电公司 Power grid network security event analysis method based on machine learning
CN113610762A (en) * 2021-07-07 2021-11-05 中国科学院国家空间科学中心 Early warning method and early warning system for solar flare
CN113743512A (en) * 2021-09-07 2021-12-03 上海观安信息技术股份有限公司 Autonomous learning judgment method and system for safety alarm event
CN113786205A (en) * 2021-11-17 2021-12-14 华南理工大学 Portable early warning system and method based on convolutional network and long-short term memory network
CN113825148A (en) * 2020-06-18 2021-12-21 ***通信集团浙江有限公司 Method and device for determining alarm level of network node and computing equipment
CN113821408A (en) * 2021-09-23 2021-12-21 中国建设银行股份有限公司 Server alarm processing method and related equipment
CN113962294A (en) * 2021-09-30 2022-01-21 西南电子技术研究所（中国电子科技集团公司第十研究所） Multi-type event prediction model
CN114726751A (en) * 2022-04-07 2022-07-08 广州汇智通信技术有限公司 Intelligent early warning method, system, equipment and storage medium for resource quality monitoring
WO2022213600A1 (en) * 2021-04-08 2022-10-13 中广核检测技术有限公司 Method for recognizing type of vortex signal of evaporator of nuclear power plant on basis of lstm-cnn
CN115474108A (en) * 2022-09-16 2022-12-13 南京品淳通信科技有限公司 Event monitoring system and method based on edge calculation
WO2022257423A1 (en) * 2021-06-08 2022-12-15 天翼云科技有限公司 Warning information association method and apparatus, and electronic device and readable storage medium
CN116980181A (en) * 2023-06-21 2023-10-31 江南信安（北京）科技有限公司 Method and system for detecting associated alarm event
CN117408165A (en) * 2023-12-14 2024-01-16 中国石油大学(华东) Fracturing process complex event intelligent early warning optimization method based on machine learning

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106569998A (en) * 2016-10-27 2017-04-19 浙江大学 Text named entity recognition method based on Bi-LSTM, CNN and CRF
CN107562784A (en) * 2017-07-25 2018-01-09 同济大学 Short text classification method based on ResLCNN models
CN108595632A (en) * 2018-04-24 2018-09-28 福州大学 A kind of hybrid neural networks file classification method of fusion abstract and body feature
CN109635083A (en) * 2018-11-27 2019-04-16 北京科技大学 It is a kind of for search for TED speech in topic formula inquiry document retrieval method
CN110110707A (en) * 2019-05-24 2019-08-09 苏州闪驰数控***集成有限公司 Artificial intelligence CNN, LSTM neural network dynamic identifying system
CN110263172A (en) * 2019-06-26 2019-09-20 国网江苏省电力有限公司南京供电分公司 A kind of evented autonomous classification method of power system monitor warning information

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106569998A (en) * 2016-10-27 2017-04-19 浙江大学 Text named entity recognition method based on Bi-LSTM, CNN and CRF
CN107562784A (en) * 2017-07-25 2018-01-09 同济大学 Short text classification method based on ResLCNN models
CN108595632A (en) * 2018-04-24 2018-09-28 福州大学 A kind of hybrid neural networks file classification method of fusion abstract and body feature
CN109635083A (en) * 2018-11-27 2019-04-16 北京科技大学 It is a kind of for search for TED speech in topic formula inquiry document retrieval method
CN110110707A (en) * 2019-05-24 2019-08-09 苏州闪驰数控***集成有限公司 Artificial intelligence CNN, LSTM neural network dynamic identifying system
CN110263172A (en) * 2019-06-26 2019-09-20 国网江苏省电力有限公司南京供电分公司 A kind of evented autonomous classification method of power system monitor warning information

Cited By (41)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113825148B (en) * 2020-06-18 2023-08-18 ***通信集团浙江有限公司 Method and device for determining alarm grade of network node and computing equipment
CN113825148A (en) * 2020-06-18 2021-12-21 ***通信集团浙江有限公司 Method and device for determining alarm level of network node and computing equipment
CN113065352B (en) * 2020-06-29 2022-07-19 国网浙江省电力有限公司杭州供电公司 Method for identifying operation content of power grid dispatching work text
CN113065352A (en) * 2020-06-29 2021-07-02 国网浙江省电力有限公司杭州供电公司 Operation content identification method for power grid dispatching work text
CN111865947B (en) * 2020-07-08 2021-07-20 浙江大学 Method for generating abnormal data of power terminal based on transfer learning
CN111865947A (en) * 2020-07-08 2020-10-30 浙江大学 Method for generating abnormal data of power terminal based on transfer learning
CN111970489A (en) * 2020-08-05 2020-11-20 北京必可测科技股份有限公司 Intelligent monitoring management method and system based on man-machine two-way
CN112001482B (en) * 2020-08-14 2024-05-24 佳都科技集团股份有限公司 Vibration prediction and model training method, device, computer equipment and storage medium
CN112001482A (en) * 2020-08-14 2020-11-27 佳都新太科技股份有限公司 Vibration prediction and model training method and device, computer equipment and storage medium
CN112187413A (en) * 2020-08-28 2021-01-05 中国人民解放军海军航空大学航空作战勤务学院 SFBC (Small form-factor Block code) identifying method and device based on CNN-LSTM (convolutional neural network-Link State transition technology)
CN112187413B (en) * 2020-08-28 2022-05-03 中国人民解放军海军航空大学航空作战勤务学院 SFBC (Small form-factor Block code) identifying method and device based on CNN-LSTM (convolutional neural network-Link State transition technology)
CN112418014A (en) * 2020-11-09 2021-02-26 南京信息工程大学滨江学院 Modulation signal identification method based on wavelet transformation and convolution long-short term memory neural network
CN112418014B (en) * 2020-11-09 2023-11-17 南京信息工程大学滨江学院 Modulated signal identification method based on wavelet transformation and convolution long-term and short-term memory neural network
CN112507720A (en) * 2020-11-12 2021-03-16 西安交通大学 Graph convolution network root identification method based on causal semantic relation transfer
CN112598023A (en) * 2020-12-02 2021-04-02 国网青海省电力公司海东供电公司 Intelligent identification and automatic mapping method and device for power grid monitoring alarm information
CN112598023B (en) * 2020-12-02 2024-01-26 国网青海省电力公司海东供电公司 Intelligent identification and automatic mapping method and device for power grid monitoring alarm information
CN112487406A (en) * 2020-12-02 2021-03-12 中国电子科技集团公司第三十研究所 Network behavior analysis method based on machine learning
CN112838946A (en) * 2020-12-17 2021-05-25 国网江苏省电力有限公司信息通信分公司 Method for constructing intelligent sensing and early warning model based on communication network faults
CN113032917A (en) * 2021-03-03 2021-06-25 安徽大学 Electromechanical bearing fault detection method based on generation countermeasure and convolution cyclic neural network and application system
CN113157859A (en) * 2021-04-06 2021-07-23 北京理工大学 Event detection method based on upper concept information
GB2616996A (en) * 2021-04-08 2023-09-27 Cgnpc Inspection Tech Co Ltd Method for recognizing type of vortex signal of evaporator of nuclear power plant on basis of LSTM-CNN
WO2022213600A1 (en) * 2021-04-08 2022-10-13 中广核检测技术有限公司 Method for recognizing type of vortex signal of evaporator of nuclear power plant on basis of lstm-cnn
CN113487010A (en) * 2021-05-21 2021-10-08 国网浙江省电力有限公司杭州供电公司 Power grid network security event analysis method based on machine learning
CN113487010B (en) * 2021-05-21 2024-01-05 国网浙江省电力有限公司杭州供电公司 Power grid network security event analysis method based on machine learning
WO2022257423A1 (en) * 2021-06-08 2022-12-15 天翼云科技有限公司 Warning information association method and apparatus, and electronic device and readable storage medium
CN113239638A (en) * 2021-06-29 2021-08-10 百维金科(上海)信息科技有限公司 Overdue risk prediction method for optimizing multi-core support vector machine based on dragonfly algorithm
CN113610762B (en) * 2021-07-07 2024-02-23 中国科学院国家空间科学中心 Early warning method and early warning system for solar flare
CN113610762A (en) * 2021-07-07 2021-11-05 中国科学院国家空间科学中心 Early warning method and early warning system for solar flare
CN113315790A (en) * 2021-07-29 2021-08-27 湖南华菱电子商务有限公司 Intrusion flow detection method, electronic device and storage medium
CN113743512A (en) * 2021-09-07 2021-12-03 上海观安信息技术股份有限公司 Autonomous learning judgment method and system for safety alarm event
CN113821408A (en) * 2021-09-23 2021-12-21 中国建设银行股份有限公司 Server alarm processing method and related equipment
CN113962294B (en) * 2021-09-30 2023-05-16 西南电子技术研究所（中国电子科技集团公司第十研究所） Multi-type event prediction model
CN113962294A (en) * 2021-09-30 2022-01-21 西南电子技术研究所（中国电子科技集团公司第十研究所） Multi-type event prediction model
CN113786205A (en) * 2021-11-17 2021-12-14 华南理工大学 Portable early warning system and method based on convolutional network and long-short term memory network
CN114726751A (en) * 2022-04-07 2022-07-08 广州汇智通信技术有限公司 Intelligent early warning method, system, equipment and storage medium for resource quality monitoring
CN115474108B (en) * 2022-09-16 2023-09-22 南京品淳通信科技有限公司 Event monitoring system and method based on edge calculation
CN115474108A (en) * 2022-09-16 2022-12-13 南京品淳通信科技有限公司 Event monitoring system and method based on edge calculation
CN116980181A (en) * 2023-06-21 2023-10-31 江南信安（北京）科技有限公司 Method and system for detecting associated alarm event
CN116980181B (en) * 2023-06-21 2024-02-20 江南信安（北京）科技有限公司 Method and system for detecting associated alarm event
CN117408165A (en) * 2023-12-14 2024-01-16 中国石油大学(华东) Fracturing process complex event intelligent early warning optimization method based on machine learning
CN117408165B (en) * 2023-12-14 2024-03-15 中国石油大学(华东) Fracturing process complex event intelligent early warning optimization method based on machine learning

Also Published As

Publication number Publication date
CN111274395B (en) 2021-11-12

Similar Documents

Publication Publication Date Title
CN111274395B (en) Power grid monitoring alarm event identification method based on convolution and long-short term memory network
CN110263172B (en) Power grid monitoring alarm information evenized autonomous identification method
Pinzón et al. Real-time multi-state classification of short-term voltage stability based on multivariate time series machine learning
Yang et al. Combined wireless network intrusion detection model based on deep learning
CN110428005B (en) Umbrella algorithm-based dynamic security misclassification constraint method for power system
CN110646708B (en) 10kV single-core cable early state identification method based on double-layer long-and-short-term memory network
CN113438114A (en) Method, device, equipment and storage medium for monitoring running state of Internet system
CN110766215B (en) Wind power climbing event prediction method based on feature adaptive selection and WDNN
CN115221233A (en) Transformer substation multi-class live detection data anomaly detection method based on deep learning
CN111783464A (en) Electric power-oriented domain entity identification method, system and storage medium
CN112559741B (en) Nuclear power equipment defect record text classification method, system, medium and electronic equipment
CN113899987A (en) Power grid fault diagnosis method based on deep pyramid convolutional neural network
CN117668493A (en) Tobacco equipment fault prediction method and system
CN113221442A (en) Construction method and device of health assessment model of power plant equipment
CN116933860A (en) Transient stability evaluation model updating method and device, electronic equipment and storage medium
Boicea et al. A novel approach for power load forecast based on GAN data augmentation
CN113961708B (en) Power equipment fault tracing method based on multi-level graph convolutional network
CN116775423A (en) Method for cluster failure prediction
CN114881172A (en) Software vulnerability automatic classification method based on weighted word vector and neural network
Jianli et al. Wind power forecasting by using artificial neural networks and Grubbs criterion
CN104573003A (en) Method for predicating financial time series based on news topic information retrieval
CN117688367B (en) Wind power generation ultra-short term power prediction method and device based on instant learning
Yu et al. Intelligent classification and automatic annotation of violations based on neural network language model
Yang et al. Text classification method for analysing accidents in power system operation
CN117592789B (en) Power grid environment fire risk assessment method and equipment based on time sequence analysis

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant