Power grid characteristic event online identification method based on multi-source data driving

Abstract

The invention relates to the technical field of power system operation control, in particular to a power grid characteristic event online identification method based on multi-source data driving. The method comprises the following steps: firstly, collecting multi-source electric quantity information in an electric power device information system, and constructing a CNN-BiGRU-AM fusion model to carry out data processing and identification; extracting local features of data by using a convolutional neural network and realizing dimension reduction; capturing the time sequence correlation and dependency relationship before and after the fault through a bidirectional gating circulation unit; and carrying out adaptive weighting on the key features through an attention mechanism so as to output an identification result. According to the method, through the cooperative advantages in the aspects of spatial feature extraction, long-sequence two-way dependence learning and key information focusing, the identification precision and model robustness of the power grid feature event in a complex data environment are improved. According to the method, online automatic identification of the power grid characteristic event can be realized, the fault response time can be shortened, and the intelligent level and the operation efficiency of power grid operation and maintenance can be improved.

Description

Technical Field

[0001] This invention relates to the field of power system operation and control technology, and in particular to an online identification method for power grid characteristic events based on multi-source data. Background Technology

[0002] Maintaining safe and stable operation at all times has long been a core objective pursued by power systems. After a grid fault occurs, comparing and analyzing system characteristics, control response processes, and simulation models is a crucial aspect of the daily operation and safety management of power systems. However, with the rapid development of new power systems, the grid structure and its operating characteristics are becoming increasingly complex. The continuous integration of massive control resources, including conventional generating units, new energy generation units, energy storage devices, and flexible loads, leads to the intertwining of various stability control strategies, such as system safety and stability control, DC frequency limiting control, primary frequency regulation, and new energy grid connection control. This results in an exponential increase in the types and scale of system response information after a fault.

[0003] Currently, power grid fault analysis and event identification mainly rely on manual offline acquisition of various data, including fault information, system response, control measures actions, and system operation modes, and offline methods are used for stability response and simulation analysis. Although information systems such as synchronous phasor measurement devices, data acquisition and monitoring control systems, and automatic safety devices can collect and upload massive amounts of multi-source monitoring information, including control commands and operating status, traditional analysis methods lack the ability to automatically collect and effectively integrate multi-source information, making it difficult to quickly and accurately identify characteristic events and control and protection action types in the power grid.

[0004] Especially with the high proportion of new energy access and the highly intelligent control of the power system, the characteristics of power grid faults exhibit stronger nonlinearity and time-series correlation. Traditional identification methods based on human experience or single models are no longer suitable for multi-source, high-dimensional, and strongly coupled data environments, which can easily lead to problems such as insufficient identification accuracy and delayed response. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] This invention provides an online identification method for power grid characteristic events based on multi-source data, to overcome the problems of existing technologies that rely on manual offline operation and cannot quickly and accurately identify power grid characteristic events.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, this invention provides an online identification method for power grid characteristic events based on multi-source data, comprising the following steps:

[0009] Step S1: Collect multi-source electrical quantity information data from the power device information system. The multi-source electrical quantity information data includes: three-phase voltage, three-phase current, voltage phase, current phase, and frequency.

[0010] Step S2: Input the multi-source electrical quantity information data into the CNN-BiGRU-AM fusion model for training and identification. The CNN-BiGRU-AM fusion model is composed of a convolutional neural network (CNN), a bidirectional gated recurrent unit (BiGRU) network, and an attention mechanism (AM) network connected in series.

[0011] Step S3: Extract local features of the multi-source electrical quantity information data and perform data dimensionality reduction using the convolutional neural network (CNN);

[0012] Step S4: Capture the feature correlation between electrical quantity data before and after the fault and extract timing feature information through the bidirectional gated cyclic unit BiGRU network;

[0013] Step S5: The attention mechanism AM network is used to perform attention weighting and identification output on the feature events to obtain the online identification results of the power grid feature events.

[0014] Preferably, the steps of extracting local features of the data and performing dimensionality reduction using a CNN network include:

[0015] Step S31: Input the multi-source electrical quantity information data into a one-dimensional multi-channel convolutional layer for convolution operation;

[0016] Step S32: Perform batch normalization on the result of the convolution operation;

[0017] Step S33: Perform a nonlinear transformation on the batch-normalized result using the ReLU activation function;

[0018] Step S34: Perform global average pooling on the result after the activation function transformation to compress the data dimension.

[0019] Preferably, the convolution operation in step S31 includes:

[0020] Multi-source datasets The input is fed into a one-dimensional multi-channel convolutional layer for convolution operations. The convolution operation formula is as follows:

[0021]

[0022] in, For feature vectors in channels With position The output value at that time For channel The bias term, Input the number of channels. For the input sequence in the channel With position The value of time, For the convolution kernel in the output channel Input channel is And the position is Weighting of time.

[0023] Preferably, the batch normalization formula is:

[0024]

[0025] in, This serves as the index for each batch of samples during the training process. For the first Batch average of each channel, For the first Batch variance of each channel, For the smoothing term parameters, and The first Each channel can learn scaling and offset parameters. This is the final output of the batch normalization operation.

[0026] Preferably, the formula for performing the nonlinear transformation using the ReLU activation function in step S33 is as follows:

[0027] .

[0028] Preferably, the global average pooling calculation formula is as follows:

[0029] .

[0030] Preferably, the extraction of temporal feature information through the bidirectional gated recurrent unit (BiGRU) network includes forward propagation and backward propagation;

[0031] The forward propagation formula is:

[0032]

[0033] in, This is the scalar sequence obtained by serializing the output of the global average pooling layer. and These are the Sigmoid and Tanh activation functions, respectively. It is in a forward-hidden state. Update the hidden state of the gate in the previous step in the forward GRU. The weight, and These are the weight matrix and bias vector of the forward update gate, respectively. , and These are the parameters for resetting the door forward. , and These are the parameters of the forward candidate hidden state. This represents the operator for element-wise multiplication;

[0034] The backpropagation formula is:

[0035]

[0036] in, It is in a backward hidden state. , and These are the weight matrix and bias vector of the backward update gate, respectively. , and These are the parameters for resetting the door backwards. , and These are the back-view candidate hidden state parameters;

[0037] The global output calculation formula for the bidirectional gated recurrent unit (BiGRU) network is as follows:

[0038]

[0039] in, This represents the element-wise addition operator.

[0040] Preferably, the attention mechanism AM network is used to perform attention weighting and identification output on feature events, including:

[0041] Step S51: Encode the global output sequence of the bidirectional gated recurrent unit (BiGRU) network into a sequence vector. The attention score is calculated using a dot product scaling model, and the formula is as follows:

[0042]

[0043] in, , and All of these are learnable parameter matrices. For sequence vectors and The dimension;

[0044] Step S52: Calculate the attention weights of the feature event identification model for each sequence vector using the Soft-Max function. The final identification output is obtained by weighted summation. The calculation formula is:

[0045] .

[0046] (III) Beneficial Effects

[0047] This invention provides an online identification method for power grid characteristic events based on multi-source data. It can achieve comprehensive collection of multi-source data and combine local feature extraction of CNN network, temporal correlation capture of BiGRU network and attention weighting mechanism of AM network. The model can deeply integrate key information before and after the fault, effectively overcome the shortcomings of traditional methods in perceiving complex features, realize online automatic identification of power grid characteristic events, shorten fault response time, and improve the intelligence level and efficiency of power grid operation and maintenance.

[0048] The CNN-BiGRU-AM model combines feature dimensionality reduction, long sequence dependency learning, and key information focusing capabilities. The CNN layer compresses data dimensionality to reduce computational burden, the BiGRU network captures temporal features before and after a fault bidirectionally, and the AM mechanism dynamically allocates attention weights, making the model more sensitive to key event signals. This achieves the goal of improving the accuracy of identifying power grid characteristic events and the robustness of the model under complex data. Attached Figure Description

[0049] Figure 1 This diagram illustrates the workflow of an online identification method for power grid characteristic events based on multi-source data driven by the present invention.

[0050] Figure 2 This shows a simulation example built in PSASP;

[0051] Figure 3 This illustrates the basic structure of a GRU network;

[0052] Figure 4 The training flowchart of the BiGRU network is shown;

[0053] Figure 5 The network structure diagram of the CNN-BiGRU-AM fusion model is shown.

[0054] Figure 6 The graph shows the stability test results of the power grid characteristic event identification model;

[0055] Figure 7 The training flowchart of the power grid characteristic event identification model is shown. Detailed Implementation

[0056] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] In the description of this invention, it is necessary to understand that the orientations or positional relationships indicated by terms such as "upper," "lower," "left," "right," "inner," "outer," "top," and "bottom" are based on the orientations or positional relationships shown in the accompanying drawings. They are intended only to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the components referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0058] like Figure 1-7 As shown, this invention provides an online identification method for power grid characteristic events based on multi-source data, comprising the following steps:

[0059] Step S1: Collect multi-source electrical quantity information data from the power device information system. This multi-source electrical quantity information data includes: three-phase voltage, three-phase current, voltage phase, current phase, and frequency.

[0060] Step S2: Input the multi-source electrical quantity information data into the CNN-BiGRU-AM fusion model for training and identification. The CNN-BiGRU-AM fusion model is composed of a convolutional neural network (CNN), a bidirectional gated recurrent unit (BiGRU) network, and an attention mechanism (AM) network connected in series.

[0061] Step S3: Extract local features of the multi-source electrical quantity information data and perform data dimensionality reduction using the convolutional neural network (CNN);

[0062] The steps involved in extracting local features from the data using a CNN network and then performing dimensionality reduction include:

[0063] Step S31: Input the multi-source electrical quantity information data into a one-dimensional multi-channel convolutional layer for convolution operation;

[0064] Step S32: Perform batch normalization on the result of the convolution operation;

[0065] Step S33: Perform a nonlinear transformation on the batch-normalized result using the ReLU activation function;

[0066] Step S34: Perform global average pooling on the result after the activation function transformation to compress the data dimension.

[0067] Step S31 is used to efficiently extract local key features from multi-source electrical quantity information data through convolution operations of a one-dimensional multi-channel convolutional layer. This convolution operation includes:

[0068] Multi-source datasets The input is fed into a one-dimensional multi-channel convolutional layer for convolution operations. The convolution operation formula is as follows:

[0069]

[0070] in, For feature vectors in channels With position The output value at that time For channel The bias term, Input the number of channels. For the input sequence in the channel With position The value of time, For the convolution kernel in the output channel Input channel is And the position is Weighting of time.

[0071] The batch normalization formula in step S32 is as follows:

[0072]

[0073] in, This serves as the index for each batch of samples during the training process. For the first Batch average of each channel, For the first Batch variance of each channel, For the smoothing term parameters, and The first Each channel can learn scaling and offset parameters. This is the final output of the batch normalization operation.

[0074] In step S33, the ReLU activation function is used to perform a nonlinear transformation. The formula is as follows:

[0075] ;

[0076] The formula for calculating global average pooling is as follows:

[0077] .

[0078] Step S4: Capture the feature correlation between electrical quantity data before and after the fault and extract timing feature information through the bidirectional gated cyclic unit BiGRU network, including forward propagation and backward propagation;

[0079] The forward propagation formula is:

[0080]

[0081] in, This is the scalar sequence obtained by serializing the output of the global average pooling layer. and These are the Sigmoid and Tanh activation functions, respectively. It is in a forward-hidden state. Update the hidden state of the gate in the previous step in the forward GRU. The weight, and These are the weight matrix and bias vector of the forward update gate, respectively. , and These are the parameters for resetting the door forward. , and These are the parameters of the forward candidate hidden state. This represents the operator for element-wise multiplication;

[0082] The backpropagation formula is:

[0083]

[0084] in, It is in a backward hidden state. , and These are the weight matrix and bias vector of the backward update gate, respectively. , and These are the parameters for resetting the door backwards. , and These are the back-view candidate hidden state parameters;

[0085] The global output calculation formula for the bidirectional gated recurrent unit (BiGRU) network is as follows:

[0086]

[0087] in, This represents the element-wise addition operator.

[0088] Step S5: The attention mechanism AM network is used to perform attention weighting and identification output on the feature events to obtain the online identification results of the power grid feature events. The attention weighting and identification output on the feature events through the attention mechanism AM network includes:

[0089] Step S51: Encode the global output sequence of the bidirectional gated recurrent unit (BiGRU) network into a sequence vector. The attention score is calculated using a dot product scaling model, and the formula is as follows:

[0090]

[0091] in, , and All of these are learnable parameter matrices. For sequence vectors and The dimension;

[0092] Step S52: Calculate the attention weights of the feature event identification model for each sequence vector using the Soft-Max function. The final identification output is obtained by weighted summation. The calculation formula is:

[0093] .

[0094] This invention, in its embodiment, establishes a simulation test system comprising three machines and nine nodes within the PSASP power system analysis and synthesis program. A simplified diagram of the calculation example is shown below. Figure 2 As shown, generator No. 2 was built using the PSASP No. 14 Type 2 photovoltaic power station model, generator No. 3 was built using the PSASP No. 12 Type 2 doubly fed wind turbine model, and generator No. 1 was a constant potential synchronous generator model and served as the system's balancing node.

[0095] In the simulation computing platform, power grid characteristic events such as generator tripping, line tripping, transformer tripping, photovoltaic voltage disconnection, photovoltaic frequency disconnection, wind turbine voltage disconnection, and wind turbine frequency disconnection were simulated. Three-phase voltage, current, and frequency data before and after different power grid characteristic events were collected to simulate the operating status monitoring data collected by synchronous phasor measurement devices, data acquisition and monitoring control systems, and safety devices. This data ultimately forms multi-source electrical quantity information data for model training and testing. The CNN-BiGRU-AM network, which uses this multi-source electrical quantity information data as input and power grid characteristic events as output, is as follows:

[0096] The core feature of this CNN network is that it extracts local features from the data through convolutional operations and reduces the number of network parameters by using weight sharing and pooling mechanisms to improve the model's efficiency and generalization ability. This multi-source dataset... The input is first fed into a one-dimensional multi-channel convolutional layer. The convolution operation formula is:

[0097]

[0098] in, For feature vectors in channels With position The output value at that time; For channel The bias term; Input the number of channels; For the input sequence in the channel With position The value at time; For the convolution kernel in the output channel Input channel is And the position is Weighting of time.

[0099] To prevent data from falling into the activation function saturation region during training of the power grid feature event identification model due to data distribution shifts, batch normalization of the convolution operation results is used to improve the model's convergence speed and avoid the gradient vanishing problem during training. The batch normalization formula is as follows:

[0100]

[0101] in, This serves as an index for each batch of samples during the training process. For the first Batch average of each channel; For the first Batch variance of each channel; For smoothing term parameters; and The first Each channel can learn scaling and offset parameters; This is the final output of the batch normalization operation.

[0102] To further enhance the ability of deep neural networks to perceive nonlinear features of data, the ReLU activation function is used to nonlinearly transform the batch-normalized output information. The calculation formula is as follows:

[0103]

[0104] To further improve the feature localization robustness of the power grid characteristic event identification model and reduce data dimensionality and parameter count, a global average pooling layer is used to process the output vector of the activation function. The formula for global average pooling is as follows:

[0105]

[0106] To capture the feature correlations between electrical quantity data before and after a fault and extract temporal feature information from multi-source data, the serialized output of the global average pooling layer of a CNN network is used as the input to a BiGRU network. This BiGRU network is a deep learning model based on the GRU network, combining a bidirectional recurrent mechanism and a gated unit structure. It can efficiently identify the correlations between sequential data, effectively avoiding the problem of traditional recurrent neural networks' weak perception of long sequence data. Furthermore, it has advantages over long short-term memory networks in terms of training cost and convergence. The training process of BiGRU includes three steps: forward propagation, back propagation, and output synthesis. The calculation formula for forward propagation is:

[0107]

[0108] in, That is, the scalar sequence obtained after serializing the output of the global average pooling layer; and These are the Sigmoid and Tanh activation functions, respectively. It is in a forward-hidden state; Update the hidden state of the gate in the previous step in the forward GRU. The weights; and These are the weight matrix and bias vector of the forward update gate, respectively; , and These are the parameters for resetting the forward door; , and These are the parameters of the forward candidate hidden state; To represent the element-wise multiplication operator; the backpropagation calculation formula is:

[0109]

[0110] in, It is in a backward hidden state; , and These are the weight matrix and bias vector of the backward update gate, respectively; , and These are the parameters for resetting the door backwards; , and These are the backward candidate hidden state parameters; finally, the global output calculation formula of the BiGRU network is:

[0111]

[0112] in, This represents the element-wise addition operator.

[0113] To improve the performance and identification accuracy of the power grid feature event identification model by focusing its attention more on useful information, an AM network is introduced to assign different attention weights to the output sequence information of the BiGRU network. The basic structure of the AM model consists of a set of query-key-value pairs. Therefore, in this process, it is first necessary to process the final global output sequence of the BiGRU network. Encode into sequence vectors The attention score is calculated using a dot product scaling model, as shown in the following formula:

[0114]

[0115] In the formula, , and All are learnable parameter matrices; For sequence vectors and The dimension, and the value vector Maintain consistency. Then, the AM model calculates the attention weights of the feature event identification model for each sequence vector using the Soft-Max function. The final feature temporal identification output of the model is obtained by calculating the weighted sum of the weights. The calculation formula is as follows:

[0116]

[0117] After the training process described above, which is based on a multi-source data and a CNN-BiGRU-AM fusion neural network model, an online identification model that can effectively determine the characteristic events of the power grid before and after a fault state can be obtained.

[0118] During model training, the stability of the power grid characteristic event identification model's performance on different data subsets was verified using 10-fold cross-validation. This involved dividing the original multi-source information data into 10 equal parts, using each part alternately as the test set, and the remaining data as the training set for model training. The cross-validation results showed that the power grid characteristic event identification model did not fall into overfitting or underfitting, and the model exhibited an overall stable trend across different data subsets.

[0119] Then, 70% of the original multi-source information data was selected as the training set to train the power grid feature event identification model. As the number of training rounds increased, the model loss gradually decreased and approached 0, and the model gradually stabilized and converged after about 1000 rounds.

[0120] Finally, the remaining 30% of the original multi-source information data was used as the test set to statistically analyze the accuracy of the trained model in identifying power grid characteristic events. The identification results show that the multi-source information data, which includes changes in electrical quantities before and after faults, sent by various information systems can effectively identify power grid characteristic events. The applied model can achieve relatively accurate automatic identification of power grid characteristic events.

[0121] It is understood that the various embodiments mentioned above in this invention can be combined with each other to form combined embodiments without violating the principle and logic. Due to space limitations, this invention will not elaborate further.

[0122] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0123] This invention provides an online identification method for power grid characteristic events based on multi-source data. It can achieve comprehensive collection of multi-source data and combine local feature extraction of CNN network, temporal correlation capture of BiGRU network and attention weighting mechanism of AM network. The model can deeply integrate key information before and after the fault, effectively overcome the shortcomings of traditional methods in perceiving complex features, realize online automatic identification of power grid characteristic events, shorten fault response time, and improve the intelligence level and efficiency of power grid operation and maintenance.

[0124] The CNN-BiGRU-AM model combines feature dimensionality reduction, long sequence dependency learning, and key information focusing capabilities. The CNN layer compresses data dimensionality to reduce computational burden, the BiGRU network captures temporal features before and after a fault bidirectionally, and the AM mechanism dynamically allocates attention weights, making the model more sensitive to key event signals. This achieves the goal of improving the accuracy of identifying power grid characteristic events and the robustness of the model under complex data.

[0125] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for online identification of power grid characteristic events based on multi-source data, characterized in that, Includes the following steps: Step S1: Collect multi-source electrical quantity information data from the power device information system. The multi-source electrical quantity information data includes: three-phase voltage, three-phase current, voltage phase, current phase, and frequency. Step S2: Input the multi-source electrical quantity information data into the CNN-BiGRU-AM fusion model for training and identification. The CNN-BiGRU-AM fusion model is composed of a convolutional neural network (CNN), a bidirectional gated recurrent unit (BiGRU) network, and an attention mechanism (AM) network connected in series. Step S3: Extract local features of the multi-source electrical quantity information data and perform data dimensionality reduction using the convolutional neural network (CNN); Step S4: Capture the feature correlation between electrical quantity data before and after the fault and extract timing feature information through the bidirectional gated cyclic unit BiGRU network; Step S5: The attention mechanism AM network is used to perform attention weighting and identification output on the feature events to obtain the online identification results of the power grid feature events.

2. The online identification method for power grid characteristic events based on multi-source data as described in claim 1, characterized in that, The steps involved in extracting local features from data and performing dimensionality reduction using a CNN network include: Step S31: Input the multi-source electrical quantity information data into a one-dimensional multi-channel convolutional layer for convolution operation; Step S32: Perform batch normalization on the result of the convolution operation; Step S33: Perform a nonlinear transformation on the batch-normalized result using the ReLU activation function; Step S34: Perform global average pooling on the result after the activation function transformation to compress the data dimension.

3. The online identification method for power grid characteristic events based on multi-source data as described in claim 2, characterized in that, The convolution operation in step S31 includes: Multi-source datasets The input is fed into a one-dimensional multi-channel convolutional layer for convolution operations. The convolution operation formula is as follows: in, For feature vectors in channels With position The output value at that time For channel The bias term, Input the number of channels. For the input sequence in the channel With position The value of time, For the convolution kernel in the output channel Input channel is And the position is Weighting of time.

4. The online identification method for power grid characteristic events based on multi-source data as described in claim 3, characterized in that, The batch normalization formula is as follows: in, This serves as the index for each batch of samples during the training process. For the first Batch average of each channel, For the first Batch variance of each channel, For the smoothing term parameters, and The first Each channel can learn scaling and offset parameters. This is the final output of the batch normalization operation.

5. The online identification method for power grid characteristic events based on multi-source data driving according to claim 4, characterized in that, The formula for performing the nonlinear transformation using the ReLU activation function in step S33 is as follows: 。 6. The online identification method for power grid characteristic events based on multi-source data as described in claim 5, characterized in that, The formula for global average pooling is: 。 7. The online identification method for power grid characteristic events based on multi-source data as described in claim 6, characterized in that, The temporal feature information extracted through the bidirectional gated recurrent unit (BiGRU) network includes forward propagation and backward propagation. The forward propagation formula is: in, This is the scalar sequence obtained by serializing the output of the global average pooling layer. and These are the Sigmoid and Tanh activation functions, respectively. It is in a forward-hidden state. Update the hidden state of the gate in the previous step in the forward GRU. The weight, and These are the weight matrix and bias vector of the forward update gate, respectively. , and These are the parameters for resetting the door forward. , and These are the parameters of the forward candidate hidden state. This represents the operator for element-wise multiplication; The backpropagation formula is: in, It is in a backward hidden state. , and These are the weight matrix and bias vector of the backward update gate, respectively. , and These are the parameters for resetting the door backwards. , and These are the back-view candidate hidden state parameters; The global output calculation formula for the bidirectional gated recurrent unit (BiGRU) network is as follows: in, This represents the element-wise addition operator.

8. The online identification method for power grid characteristic events based on multi-source data as described in claim 7, characterized in that, The attention mechanism AM network performs attention-weighted and identification outputs on feature events, including: Step S51: Encode the global output sequence of the bidirectional gated recurrent unit (BiGRU) network into a sequence vector. The attention score is calculated using a dot product scaling model, and the calculation formula is as follows: in, , and All of these are learnable parameter matrices. For sequence vectors and The dimension; Step S52: Calculate the attention weights of the feature event identification model for each sequence vector using the Soft-Max function. The final identification output is obtained by weighted summation. The calculation formula is: 。

Images