Power network security risk prediction method based on neural network
Through the neural network-based power network security risk prediction method, the CNN-Bi-LSTM-FC model combined with the attention mechanism is used to solve the problem that traditional methods cannot predict potential risks, and the accurate risk assessment and real-time early warning of the power network are realized. Improves the stability and flexibility of the power network.
Patent Information
- Application Number
- CN202411112163.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-14
- Publication Date
- 2024-11-26
AI Technical Summary
Traditional power network security risk prevention methods cannot effectively predict potential risks, and cannot adapt to the flexible and changing needs of power networks, and cannot detect and respond to complex cyber attacks in real time.
Adopt the neural network security risk prediction method, select the risk indicators of the power network, perform data preprocessing and model construction, and use the CNN-Bi-LSTM-FC model to combine the attention mechanism to conduct risk prediction and analyze network data in real time , provide early warning information.
It realizes accurate prediction of power network security risks, reduces false alarm rates and missed alarm rates, provides cross-level and cross-field risk assessments, and improves the stability and flexibility of power networks.
Smart Images

Figure CN119026907A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of power network security and artificial intelligence, and in particular to a power network security risk prediction method based on neural network. Background Art
[0002] The power network is an important foundation for the stable operation of the power system. With the continuous evolution and complexity of cyber attack methods, key infrastructure in the power network, such as data centers, substations, distribution stations, data centers, etc., are easily targeted by hacker attacks. These attacks may cause power supply interruptions, system paralysis, and even cause large-scale social chaos and economic losses, seriously threatening national security and social stability. Therefore, how to effectively predict potential network security threats is an urgent and key consideration in the current power network security protection and research process.
[0003] Traditional methods of defending against power network security risks include installing firewalls, antivirus software, intrusion detection, etc. Although these methods have played a certain role, they require real-time software updates and only target specific viruses and malicious codes, which cannot meet the flexible and changeable needs of power networks. In addition, traditional methods focus on detection and response and cannot predict potential risks in power networks.
[0004] As an emerging technology, artificial intelligence can effectively learn the correlation between complex data. Therefore, how to apply artificial intelligence to the prediction of power network security risks and mine abnormal behaviors and hidden attacks in the network by analyzing the characteristics and behavior patterns of the network environment is of great research significance. Summary of the invention
[0005] In view of the above problems, the present invention provides a method for predicting power network security risks based on neural network.
[0006] The technical solution of the present invention is: a method for predicting power network security risks based on neural network, comprising the following steps:
[0007] Step 1: Select power network security risk indicators and form risk data.
[0008] Step 2: Risk data preprocessing, including data cleaning, data missing value filling and data fusion.
[0009] Step 3: Build a power network risk prediction model to predict power network security risks based on preprocessed risk data.
[0010] Furthermore, step 1 specifically includes:
[0011] Select risk indicator sources. For the security assessment and risk management of power networks, select traffic statistics, vulnerability information, services, hosts, local networks, macro networks and users as the sources of power network risk indicators.
[0012] A risk indicator evaluation system is constructed based on the source of indicators, including primary indicators and secondary indicators. The primary indicators are the specific quantification of risks, and the secondary indicators are refined into four indicators: vulnerability, threat, disaster tolerance and stability based on the primary indicators.
[0013] The vulnerability includes six first-level indicators, namely: the number and level of network vulnerabilities, the number and level of important equipment vulnerabilities, the number of security equipment in the subnet, the operating system type and version of each important equipment in the subnet, the total number of development ports of each important equipment in the subnet, and the network topology;
[0014] The threat level includes eight first-level indicators, namely: number of alarms, subnet broadband usage, historical frequency of safety events in the subnet, subnet data inflow, subnet inflow growth rate, distribution of packets of different protocols in the subnet, distribution of packets of different sizes in the subnet, and IP distribution;
[0015] The disaster tolerance includes five first-level indicators, namely: network bandwidth, number of security equipment in the subnet, network topology, number of threads, and number of concurrent threads supported by the server in the subnet;
[0016] The stability includes six first-level indicators, namely: subnet traffic change rate, subnet data flow total amount, subnet traffic change rate, subnet mean time between failures, number of important equipment surviving in the subnet, and IP distribution of output subnet data packets.
[0017] Furthermore, step 2 specifically includes:
[0018] Data cleaning: Data cleaning mainly involves unified processing of network risk time series data from different sources and in different formats, removing duplicate data through reasonable methods, and filling in missing data.
[0019] Data missing value filling. In the case of random missing data, the present invention adopts the random forest method to fill the missing values. Ensemble learning is performed by constructing multiple decision trees, each of which is trained based on a random subset of the data set and a feature subset. When filling the missing values, the random forest model uses other non-missing features to predict the value of the missing feature.
[0020] Data fusion. This invention adopts the method of feature-level data fusion, focusing on integrating the information of different risk indicators. First, different indicators are vectorized to form feature vectors that reflect the information or attributes of different indicators; then, considering the correlation between different indicators, the feature vectors are spliced together by series connection; finally, in order to reduce the data dimension, the PCA method is used to reduce the dimension of the feature vector to form a vector matrix.
[0021] Furthermore, step 3 specifically includes:
[0022] 3-1 Construct a CNN-Bi-LSTM-FC power network risk prediction model with an added attention mechanism. The model input is the power network indicator vector matrix, and the output is the risk prediction result. Among them, CNN extracts deep features in the spatial dimension by convolution; the Attention layer is used for the neural network to autonomously learn key information; Bi-LSTM can effectively capture time series dynamic information. FC is used to combine the features obtained by the previous network for the final classification.
[0023] (1) CNN layer
[0024] CNN is a typical deep learning network with the characteristics of weight sharing and local connection. CNN consists of convolutional layers and pooling layers. The convolutional layer is used for feature extraction. The convolution kernel is gradually slid along the data window to perform convolution operations. The convolutional layer usually uses activation functions to introduce nonlinearity to enhance the expression ability of the model, such as relu, sigmoid, etc. The pooling layer summarizes the information after convolution and reduces the dimension of the data by taking the average or maximum value.
[0025] (2) Attention layer
[0026] Considering that CNN focuses on local features during feature extraction and lacks the capture of global information.
[0027] Therefore, the present invention allocates attention weights by calculating the similarity between different positions in the feature map after the convolution layer in CNN, and generates a new feature representation, as shown in the attached figure. Figure 5 The specific process is as follows:
[0028] 1) Generate Q, K, V matrices
[0029] The input matrix X is transformed linearly to generate three new matrices: query matrix (Q), key matrix (K)
[0030] The sum value matrix (V) is shown in equations (1), (2) and (3).
[0031] Q=WqX (1)
[0032] K=W k X (2)
[0033] V=W v X (3)
[0034] Where X is the input vector matrix, Wq, Wk and Wv are all linear transformation matrices;
[0035] 2) Calculate attention weight
[0036] Calculate the dot product of the transposed matrices of Q and K to obtain the attention weight w, as shown in formula (4).
[0037] w=QK T (4)
[0038] 3) Weighted sum
[0039] The attention weights are normalized using the softmax normalization function so that the sum of all scores is 1. In this way, each position is assigned a weight based on the degree of its relationship with other positions. These weights are then multiplied with the value matrix V to obtain a weighted vector.
[0040] (3) Bi-LSTM layer
[0041] The LSTM network is an improved structural model based on RNN. The LSTM network introduces storage units and gate mechanisms, which can use the information in the vector of the previous moment at the current moment. The LSTM network consists of four parts: input gate i t 、Forget gate t and output gate o t and storage memory unit C t The forget gate determines the amount of information to be discarded, which is determined by formula (5):
[0042] f t =σ(W t h t-1 +U t x t +b t ) (5)
[0043] Where: f t is the output of the forget gate, σ is the sigmoid function, x t is the input at the current time t; h t-1 is the output of the hidden layer at time t-1; U t x is the forget gate t The weight of t h is the forget gate t-1 The weight of b t is the bias of the forget gate.
[0044] The input gate determines how much information the LSTM network will save when updating a cell, which is determined by equations (6) to (8):
[0045] i t =σ(W i h i-1 +U i x i +b i ) (6)
[0046] C=W c h t-1 +U c x t +b c (7)
[0047]
[0048] Where: i t is the output of the input gate, σ is the value sigmoid function, U i is the input gate x i The weight of i is the input gate h i-1 The weight of; C is the temporary unit state; U c x of C t The weight of c h for C t-1 The weights of bi and b c are the bias parameters of i and C respectively; C t is the current unit state; C t-1 is the unit state at the previous moment; It is the vector obtained after the elements in C are activated by the tanh function.
[0049] The output gate determines the output of the LSTM model:
[0050] o t =σ(W o h t-1 +U o x t +b o ) (9)
[0051] h t =o t tanh(C t ) (10)
[0052] Where: t is the output of the output gate, σ is the value of the sigmoid function, W o h is the output gate t-1 The weight of Uo is the output gate x t The weight of o is the bias parameter of the output gate; h t is the final output of the current LSTM.
[0053] In the LSTM network structure, the state is transmitted unidirectionally from the front to the back. Therefore, LSTM only obtains the forward features of the data, but cannot obtain the backward features of the data. In order to achieve complete feature information acquisition, the present invention uses a Bi-LSTM network. The Bi-LSTM network obtains the forward information h of each vector t-1 and backward information h t+1 , 2 different feature representations.
[0054] Specifically, the Bi-LSTM neural network structure model is divided into two independent LSTMs. The input sequence is input into the two LSTM neural networks in forward and reverse order for feature extraction, and the two output vectors (i.e., the extracted feature vectors) are concatenated to form the final feature expression. The design concept of the Bi-LSTM model is to make the feature data obtained at time t have information between the past and the future at the same time.
[0055] (4) FC layer
[0056] The role of FC is to combine the output features of Bi-LSTM so that the model can perform classification based on all the features. In FC, each neuron is connected to all neurons in the previous layer, and the output value is the weighted sum of the neurons in the previous layer, as shown in formula (11).
[0057] y=Wx+b (11)
[0058] In the formula, y represents the output of FC, W represents the weight, x represents the input, and b represents the bias.
[0059] In the present invention, when the power network risk prediction model is used to predict network risks in real time, the model will continuously receive real-time data from the network monitoring system. These data are pre-processed and converted into vector matrices suitable for neural network processing before being input into the model. The risk prediction model uses the learned weights and time series relationships to conduct an in-depth analysis of the current power network data and predict the security risks that may exist in the future, providing network administrators with early warning information so that they can respond quickly and take corresponding defensive measures.
[0060] The present invention introduces artificial intelligence technology and utilizes the powerful nonlinear modeling and autonomous learning capabilities of neural networks to deeply explore the hidden patterns and associations in the power network operation data, realize early prediction of safety risks, and provide strong pre-emptive protection for the stable operation of the power network. BRIEF DESCRIPTION OF THE DRAWINGS
[0061] Figure 1 It is a flow chart of a power network risk prediction method based on neural network;
[0062] Figure 2 This is a schematic diagram of the sources of power network risk indicators;
[0063] Figure 3 It is the principle diagram of the present invention;
[0064] Figure 4 Supplementary schematic diagram for missing values based on random forest;
[0065] Figure 5 Schematic diagram of the prediction model based on CNN-Bi-LSTM-FC;
[0066] Figure 6 This is the Bi-LSTM structure diagram. DETAILED DESCRIPTION
[0067] The technical solution of the present invention is now further described in conjunction with the accompanying drawings and embodiments.
[0068] The present invention Figure 1-6 As shown, a method for predicting power network security risks based on a neural network is provided, comprising:
[0069] Step 1: Selection of power network security risk indicators.
[0070] Step 1-1: Select risk indicator sources. For the security assessment and risk management of power networks, select traffic statistics, vulnerability information, services, hosts, local networks, macro networks and users as the sources of power network risk indicators, as shown in the attached Figure 2 shown.
[0071] Step 1-2: Construct a risk indicator evaluation system based on the indicator source, including primary indicators and secondary indicators. The primary indicators are the specific quantification of risks, and the secondary indicators are refined into four indicators: vulnerability, threat, disaster tolerance and stability based on the primary indicators.
[0072]
[0073]
[0074] Step 2: Risk data preprocessing, including data cleaning, data missing value filling and data fusion.
[0075] Data cleaning: Data cleaning mainly involves unified processing of network risk time series data from different sources and in different formats, removing duplicate data through reasonable methods, and filling in missing data.
[0076] Data missing value filling. In the case of random missing data, the present invention adopts the random forest method to fill the missing values. Ensemble learning is performed by constructing multiple decision trees, each of which is trained based on a random subset of the data set and a feature subset. When filling missing values, the random forest model uses other non-missing features to predict the value of the missing feature, such as the attached Figure 3 shown.
[0077] Data fusion. The present invention adopts the method of feature-level data fusion, focusing on integrating the information of different risk indicators. First, different indicators are vectorized to form feature vectors that reflect the information or attributes of different indicators; then, considering the correlation between different indicators, the feature vectors are spliced together by series connection; finally, in order to reduce the data dimension, the PCA method is used to reduce the dimension of the feature vector, thereby forming a vector matrix.
[0078] Step 3: Build a CNN-Bi-LSTM-FC neural network with an attention mechanism to predict power network security risks.
[0079] Construct a CNN-Bi-LSTM-FC power network risk prediction model with added attention mechanism, as shown in the attached figure. Figure 4 As shown in the figure. The model input is the power network indicator vector matrix, and the output is the risk prediction result. Among them, CNN extracts deep features in the spatial dimension by convolution; the Attention layer is used for the neural network to autonomously learn key information; Bi-LSTM can effectively capture the dynamic information of time series. FC is used to combine the features obtained by the previous network for the final classification.
[0080] (1) CNN layer
[0081] CNN is a typical deep learning network with the characteristics of weight sharing and local connection. CNN consists of convolutional layers and pooling layers. The convolutional layer is used for feature extraction. The convolution kernel is gradually slid along the data window to perform convolution operations. The convolutional layer usually uses activation functions to introduce nonlinearity to enhance the expression ability of the model, such as relu, sigmoid, etc. The pooling layer summarizes the information after convolution and reduces the dimension of the data by taking the average or maximum value.
[0082] (2) Attention layer
[0083] Considering that CNN focuses on local features during feature extraction and lacks the capture of global information.
[0084] Therefore, the present invention allocates attention weights by calculating the similarity between different positions in the feature map after the convolution layer in CNN, and generates a new feature representation, as shown in the attached figure. Figure 5 The specific process is as follows:
[0085] 1) Generate Q, K, V matrices
[0086] The input matrix X is transformed linearly to generate three new matrices: query matrix (Q), key matrix (K)
[0087] The sum value matrix (V) is shown in equations (1), (2) and (3).
[0088] Q=WqX (1)
[0089] K=W k X (2)
[0090] V=W v X (3)
[0091] Where X is the input vector matrix, Wq, Wk and Wv are all linear transformation matrices;
[0092] 2) Calculate attention weight
[0093] Calculate the dot product of the transposed matrices of Q and K to obtain the attention weight w, as shown in formula (4).
[0094] w=Q·K T (4)
[0095] 3) Weighted sum
[0096] The attention weights are normalized using the softmax normalization function so that the sum of all scores is 1. In this way, each position is assigned a weight based on the degree of its relationship with other positions. These weights are then multiplied with the value matrix V to obtain a weighted vector.
[0097] (1) Bi-LSTM layer
[0098] The LSTM network is an improved structural model based on RNN. The LSTM network introduces storage units and gate mechanisms, which can use the information in the vector of the previous moment at the current moment. The LSTM network consists of four parts: input gate i t 、Forget gate t and output gate o t and storage memory unit C tThe forget gate determines the amount of information to be discarded, which is determined by formula (5):
[0099] f t =σ(W t h t-1 +U t x t +b t ) (5)
[0100] Where: f t is the output of the forget gate, σ is the sigmoid function, x t is the input at the current time t; h t-1 is the output of the hidden layer at time t-1; U t x is the forget gate t The weight of t h is the forget gate t-1 The weight of b t is the bias of the forget gate.
[0101] The input gate determines how much information the LSTM network will save when updating a cell, which is determined by equations (6) to (8):
[0102] i t =σ(W i h i-1 U i x i +b i ) (6)
[0103] C=W c h t-1 +U c x t +b c (7)
[0104]
[0105] Where: i t is the output of the input gate, σ is the value sigmoid function, U i is the input gate x i The weight of i is the input gate h i-1 The weight of; C is the temporary unit state; U c x of C t The weight of c h for C t-1 The weights of bi and b c are the bias parameters of i and C respectively; C t is the current unit state; C t-1 is the unit state at the previous moment; It is the vector obtained after the elements in C are activated by the tanh function.
[0106] The output gate determines the output of the LSTM model:
[0107] o t =σ(W o h t-1 +U o x t +b o ) (9)
[0108] h t =o t tanh(C t ) (10)
[0109] In the formula: o t is the output of the output gate, σ is the value of the sigmoid function, W o h is the output gate t-1 The weight of U o is the output gate x t The weight of o is the bias parameter of the output gate; h t is the final output of the current LSTM.
[0110] In the LSTM network structure, the state is transmitted unidirectionally from the front to the back. Therefore, LSTM only obtains the forward features of the data, but cannot obtain the backward features of the data. In order to achieve complete feature information acquisition, the present invention uses a Bi-LSTM network. The Bi-LSTM network obtains the forward information h of each vector t-1 and backward information h t+1 , 2 different feature representations.
[0111] Specifically, the Bi-LSTM neural network structure model is divided into two independent LSTMs. The input sequence is input into the two LSTM neural networks in forward and reverse order for feature extraction. The two output vectors (i.e., the extracted feature vectors) are concatenated to form the final feature expression, as shown in the attached figure. Figure 6 The design concept of the Bi-LSTM model is to make the feature data obtained at time t contain information between the past and the future.
[0112] (2) FC layer
[0113] The role of FC is to combine the output features of Bi-LSTM so that the model can perform classification based on all the features. In FC, each neuron is connected to all neurons in the previous layer, and the output value is the weighted sum of the neurons in the previous layer, as shown in formula (22).
[0114] y=Wx+b (11)
[0115] In the formula, y represents the output of FC, W represents the weight, x represents the input, and b represents the bias.
[0116] When using the trained power network risk prediction model to predict network risks in real time, the model will continuously receive real-time data from the network monitoring system. After preprocessing, these data are converted into vector matrices suitable for neural network processing and then input into the model. The risk prediction model uses its learned weights and time series relationships to conduct in-depth analysis of current power network data and predict possible security risks in the future, providing network administrators with early warning information so that they can respond quickly and take appropriate defensive measures.
[0117] The research of the power network risk prediction method based on neural network in this invention has many important significances. First, the method uses advanced neural network models to accurately capture the complex and changeable security risk factors in the power network environment, including but not limited to network traffic anomalies, equipment status changes, malware activities, etc., so as to achieve accurate prediction of potential security threats and significantly reduce false alarm and missed alarm rates. Secondly, the method can comprehensively consider the multi-dimensional information of the power network, such as hosts, services, etc., to achieve comprehensive risk assessment across levels and fields. Finally, compared with targeted defense software such as antivirus software and firewalls, this method can dig out potential correlations and models between data, and is more flexible.
[0118] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A method for predicting power network security risks based on neural network, characterized in that: The following steps are involved: Step 1: Select power network security risk indicators and form risk data; Step 2: Risk data preprocessing; Step 3: Build a power network risk prediction model to predict power network security risks based on preprocessed risk data.
2. The method for predicting power network security risks based on neural network according to claim 1, characterized in that: The risk data in step 1 includes selecting the source of risk indicators and constructing a risk indicator evaluation system based on the indicator source.
3. The method for predicting power network security risks based on neural network according to claim 1, characterized in that: Traffic statistics, vulnerability information, services, hosts, local networks, macro networks and users are selected as sources of power network risk indicators.
4. The method for predicting power network security risks based on neural network according to claim 1, characterized in that: The risk indicator evaluation system includes four secondary indicators, namely vulnerability, threat, disaster tolerance and stability. The vulnerability includes six first-level indicators, namely: the number and level of network vulnerabilities, the number and level of important equipment vulnerabilities, the number of security equipment in the subnet, the operating system type and version of each important equipment in the subnet, the total number of development ports of each important equipment in the subnet, and the network topology; The threat level includes eight first-level indicators, namely: number of alarms, subnet broadband usage, historical frequency of safety events in the subnet, subnet data inflow, subnet inflow growth rate, distribution of packets of different protocols in the subnet, distribution of packets of different sizes in the subnet, and IP distribution; The disaster tolerance includes five first-level indicators, namely: network bandwidth, number of security equipment in the subnet, network topology, number of threads, and number of concurrent threads supported by the server in the subnet; The stability includes six first-level indicators, namely: subnet traffic change rate, subnet data flow total amount, subnet traffic change rate, subnet mean time between failures, number of important equipment surviving in the subnet, and IP distribution of output subnet data packets.
5. The method for predicting power network security risks based on neural network according to claim 1, characterized in that: The risk data preprocessing in step 2 includes: data cleaning, data missing value filling and data fusion.
6. A method for predicting power network security risks based on neural network according to claim 5, characterized in that: The data cleaning is used to remove duplicate data; The missing data value supplement is used to supplement the missing values when the data is randomly missing.
7. The method for predicting power network security risks based on neural network according to claim 5, characterized in that: The data fusion integrates the information of different risk indicators by adopting the feature-level data fusion method; First, different indicators are vectorized to form feature vectors that reflect the information or attributes of different indicators; Then, considering the correlation between different indicators, the feature vectors are concatenated in series; finally, the PCA method is used to reduce the dimension of the feature vector to form a vector matrix.
8. The method for predicting power network security risks based on neural network according to claim 1, characterized in that: In step 3, a CNN-Bi-LSTM-FC power network risk prediction model with added attention mechanism is constructed. The model input is the power network indicator vector matrix, and the output is the risk prediction result. Among them, CNN extracts deep features in the spatial dimension through convolution; the Attention mechanism is used for the neural network to autonomously learn key information; Bi-LSTM is used to capture dynamic information of time series; FC is used to combine the features obtained by the previous network for final classification.
9. A method for predicting power network security risks based on neural network according to claim 8, characterized in that: The power network risk prediction model includes: (1) CNN layer It consists of convolutional layers and pooling layers. The convolutional layer is used for feature extraction. The convolution kernel is gradually slid along the data window to perform convolution operations. The convolutional layer uses activation functions to introduce nonlinearity to enhance the expressiveness of the model. The pooling layer summarizes the information after convolution and reduces the dimension of the data by taking the average or maximum value. (2) Attention layer After the convolutional layer in CNN, the input of this layer is the feature map output by the convolutional layer. The attention weight is assigned by calculating the similarity between different positions in the feature map, and a new feature representation is generated. The specific process is as follows: 1) Generate Q, K, V matrices The input matrix X is transformed linearly to generate three new matrices: query matrix (Q), key matrix (K) The sum value matrix (V) is shown in equations (1), (2) and (3): Q = W q X (1) K = W k X (2) V = W v X (3) Where X is the input vector matrix, Wq, Wk and Wv are all linear transformation matrices; 2) Calculate attention weight Calculate the dot product of the transposed matrices of Q and K to obtain the attention weight w, as shown in formula (4): w=Q﹒K T (4) 3) Weighted sum The attention weights are normalized using the softmax normalization function; in this way, each position is assigned a weight based on the degree of its relationship with other positions; these weights are then multiplied with the value matrix V to obtain a weighted vector; (3) Bi-LSTM layer The LSTM network consists of four parts: input gate i t 、Forget gate t and output gate o t and storage memory unit C t ; The forget gate determines the amount of information to be discarded, which is determined by formula (5): f t =σ(W t h t-1 +U t x t +b t ) (5) Where: f t is the output of the forget gate, σ is the sigmoid function, x t is the input at the current time t; h t-1 is the output of the hidden layer at time t-1; U t x is the forget gate t The weight of t h is the forget gate t-1 The weight of b t is the bias of the forget gate; The input gate determines how much information the LSTM network will save when updating a cell, which is determined by equations (6) to (8): i t =σ(W i h i-1 +U i x i +b i ) (6) C=W c h t-1 +U c x t +b c (7) Where: i t is the output of the input gate, σ is the value sigmoid function, U i is the input gate x i The weight of i is the input gate h i-1 The weight of; C is the temporary unit state; U c x of C t The weight of c h for C t-1 The weight of i and b c i t and C bias parameters; C t is the current unit state; C t-1 is the unit state at the previous moment; is the vector obtained after the elements in C are activated by the tanh function; The output gate determines the output of the LSTM model: the t =σ(W o h t-1 +∪ o x t +b o ) (9) h t =o t fishy(C) t ) (10) In the formula: o t is the output of the output gate, σ is the value of the sigmoid function, W o h is the output gate t-1 The weight of U o is the output gate x t The weight of o is the bias parameter of the output gate; h t is the final output of the current LSTM; In order to achieve complete feature information acquisition, a Bi-LSTM network is used. The Bi-LSTM network obtains the forward information h of each vector t-1 and backward information h t+1 , two different feature representations; (4) FC layer The role of FC is to combine the output features of Bi-LSTM so that the model can perform classification based on all the features. In FC, each neuron is connected to all neurons in the previous layer, and the output value is the weighted sum of the neurons in the previous layer, as shown in formula (11): y=Wx+b (11) In the formula, y represents the output of FC, W represents the weight, x represents the input, and b represents the bias.
Citation Information
Patent Citations
CNN-LSTM-based power system network security situation prediction method and system
CN116112283A
Power system network security situation detection method and system
CN116346433A
Cited By
Rockburst risk prediction method and system in combination with LSTM and convolutional neural network
CN119886818A
Power transaction service security risk prediction method and device based on time sequence detection
CN120180088A
Power transaction service security risk prediction method and device based on timing detection
CN120180088B
Electric power information network security risk assessment method and system
CN120185855A
Communication network security situation prediction method and system based on big data
CN120223439A
