Method and device for predicting critical point of security domain of power system based on memory enhancement

Abstract

The invention discloses an electric power system security domain critical point prediction method and device based on memory enhancement, and the method comprises the steps: calling PSD-BPA in batches through matlab, and obtaining original measurement data; based on a Dense 1D-CNN network structure, constructing a dynamic security domain critical point prediction initial model; the historical parameter memory ability of the dynamic security domain critical point prediction initial model is enhanced through a Fisher information matrix; reconstructing the historical data through a generative model VAE to obtain generated data; mixing the generated data with historical data through a playback technology to obtain mixed data; performing iterative training on the dynamic security domain critical point prediction initial model through the mixed data to obtain a trained dynamic security domain critical point prediction model; and performing prediction processing on the original measurement data through the trained dynamic security domain critical point prediction model to obtain a dynamic security domain critical point in a set scene. According to the method, the critical point of the dynamic safety domain can be quickly and accurately predicted.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic security domain of power system, and in particular to a method and device for predicting critical points of security domain of power system based on memory enhancement. Background Art

[0002] Both the power supply side and the load side of the new power system show strong randomness, which places higher requirements on the safe and stable operation of the power system.

[0003] At present, the traditional calculation method of dynamic safety domain is based on repeated trials based on dichotomy in the process of predicting critical points, which often takes a long time and is difficult to provide timely and accurate predictions. The unprecedented changes in digital technology in recent years have also brought new possibilities for the construction of new power systems. The rapid development of artificial intelligence technology has prompted data-driven dynamic safety assessment methods to gradually gain attention. It solves the problems of high computational complexity and poor real-time performance of traditional simulation methods based on physical models. As a powerful nonlinear modeling tool, neural networks can learn system operation characteristics from historical data and make rapid predictions of critical points. They are meaningful and valuable in promoting the research of artificial intelligence technology in the field of dynamic safety domains.

[0004] Therefore, how to invent a power system security domain critical point prediction method that can quickly and accurately predict the dynamic security domain critical point has become an urgent problem to be solved. Summary of the invention

[0005] To this end, the present invention provides a method and device for predicting critical points in the safety domain of an electric power system based on memory enhancement. By constructing a neural network model with fault parameter information, a reference point, a stable injection power, and a search direction as input, the dynamic safety domain critical point is used as the prediction target of the neural network, and the dynamic characteristics and pattern information in the historical data are fully utilized. After training, the dynamic safety domain critical point can be predicted quickly and accurately, which provides an efficient and intelligent technical means for the dynamic safety domain analysis of the electric power system.

[0006] In order to achieve the above object, the present invention provides the following technical solution: a method for predicting critical points in power system security domain based on memory enhancement, comprising:

[0007] Call PSD-BPA in batches through MATLAB to obtain the original measurement data;

[0008] Based on the Dense 1D-CNN network structure, an initial model for predicting the critical point of the dynamic security domain is constructed;

[0009] The historical parameter memory capability of the dynamic safety domain critical point prediction initial model is enhanced through the Fisher information matrix;

[0010] Reconstruct historical data through the generative model VAE to obtain generated data;

[0011] Mixing the generated data with the historical data by a playback technology to obtain mixed data;

[0012] Iteratively training the dynamic safety domain critical point prediction initial model through the mixed data to obtain a trained dynamic safety domain critical point prediction model;

[0013] The original measurement data is predicted and processed by the trained dynamic safety domain critical point prediction model to obtain the dynamic safety domain critical point under the set scenario.

[0014] As a preferred solution of the power system safety domain critical point prediction method based on memory enhancement, the original measurement data includes: a reference point under a fault scenario, a stable injection power, a search direction and a corresponding critical injection power.

[0015] As a preferred solution of the power system security domain critical point prediction method based on memory enhancement, the dynamic security domain critical point prediction initial model includes: an input layer, a convolution layer, a dense connection, a fully connected layer, an output layer and a loss function;

[0016] The expression of the input feature vector of the input layer is:

[0017] X=[x 1 ,x 2 ,...,x n ],x i ∈R d

[0018] Where X∈R n×d is the input feature vector; n is the number of features; d is the dimension of each feature;

[0019] The expression of the convolution kernel of the convolution layer is:

[0020] H (l) =RELU(W (l) *H (l-1) +b (l) )

[0021] In the formula, H (l-1) is the output of the previous layer (for the first layer, H (0) =X); W (l) ∈R k×d is the convolution kernel weight of the lth layer; * is the one-dimensional convolution operation; b (l) is the bias term; RELU is the activation function;

[0022] The expression of the dense connection is:

[0023] H (l) =ReLU([H (0) ,H (1) ,...,H (l-1) ]·W (l) +b (l) )

[0024] Where [·] is the feature concatenation operation;

[0025] The expression of the fully connected layer is:

[0026] Z = ReLU(H (L) ·W fc +b fc )

[0027] In the formula, H (L) is a high-dimensional feature map; W fc is the weight matrix of the fully connected layer; b fc is the bias term;

[0028] The expression of the output layer is:

[0029]

[0030] In the formula, is the critical point of the dynamic security domain; W out and b out are the weight and bias of the output layer respectively; σ(·) is the linear activation function;

[0031] The loss function is the mean square error, expressed as:

[0032]

[0033] In the formula, is the predicted value of the network; y i is the true critical point label; N is the number of samples.

[0034] As a preferred solution of the power system security domain critical point prediction method based on memory enhancement, in the process of enhancing the historical parameter memory capability of the dynamic security domain critical point prediction initial model through the Fisher information matrix, the enhancement steps are:

[0035] Obtain the importance value of the set weight by calculation;

[0036] Dynamically adjusting the learning rate according to the importance value to obtain an adjusted learning rate;

[0037] Preserving long-term memory through an exponential decay mechanism and dynamically updating the importance value;

[0038] By adding a regularization term based on the Fisher information matrix to the loss function, key parameters are protected;

[0039] According to the adjusted learning rate, the set weight is optimized and updated to obtain the optimized weight.

[0040] As a preferred solution of the power system security domain critical point prediction method based on memory enhancement, in the process of reconstructing the historical data through the generation model VAE to obtain the generated data, the expression of the generated data is:

[0041] D gen ={x 1 ',x' 2 ,...,x' N},x i '~p θ (x|z),z~p(z)

[0042] Where D gen is the generated dataset generated by VAE; x i ' is the generated data; p θ (x|z) is the conditional probability distribution generated by the decoder; p(z) is the prior distribution; z is the latent variable.

[0043] As a preferred solution of the power system security domain critical point prediction method based on memory enhancement, in the process of iteratively training the dynamic security domain critical point prediction initial model through the mixed data, the dynamic security domain critical point prediction initial model is iteratively trained through a joint loss function; the expression of the joint loss function is:

[0044] L total =L task +λ 1 L VAE

[0045] Where, L task The loss of the task for dynamic security domain prediction; L VAE is the loss of the generative model; 1 is the weight for generating the model loss.

[0046] The present invention also provides a power system security domain critical point prediction device based on memory enhancement, based on the above power system security domain critical point prediction method based on memory enhancement, including:

[0047] The original measurement data acquisition module is used to call PSD-BPA in batches through matlab to obtain the original measurement data;

[0048] Dynamic safety domain critical point prediction initial model construction module, used to build a dynamic safety domain critical point prediction initial model based on the Dense 1D-CNN network structure;

[0049] A Fisher information matrix strengthening module, used for strengthening the historical parameter memory capability of the dynamic safety domain critical point prediction initial model through the Fisher information matrix;

[0050] The generated data acquisition module is used to reconstruct historical data through the generation model VAE to obtain generated data;

[0051] A mixed data acquisition module, used for mixing the generated data with the historical data through a playback technology to obtain mixed data;

[0052] A dynamic safety domain critical point prediction initial model training module is used to iteratively train the dynamic safety domain critical point prediction initial model through the mixed data to obtain a trained dynamic safety domain critical point prediction model;

[0053] The dynamic safety domain critical point acquisition module is used to predict the original measurement data through the trained dynamic safety domain critical point prediction model to obtain the dynamic safety domain critical point under a set scenario.

[0054] As a preferred solution of the power system safety domain critical point prediction device based on memory enhancement, in the original measurement data acquisition module, the original measurement data includes: a reference point under a fault scenario, a stable injection power, a search direction and a corresponding critical injection power.

[0055] As a preferred solution of the power system security domain critical point prediction device based on memory enhancement, in the dynamic security domain critical point prediction initial model construction module, the dynamic security domain critical point prediction initial model includes: an input layer, a convolution layer, a dense connection, a fully connected layer, an output layer and a loss function;

[0056] The expression of the input feature vector of the input layer is:

[0057] X=[x 1 ,x 2 ,...,x n ],x i ∈R d

[0058] Where X∈R n×d is the input feature vector; n is the number of features; d is the dimension of each feature;

[0059] The expression of the convolution kernel of the convolution layer is:

[0060] H(l) =RELU(W (l) *H (l-1) +b (l) )

[0061] In the formula, H (l-1) is the output of the previous layer (for the first layer, H (0) =X); W (l) ∈R k×d is the convolution kernel weight of the lth layer; * is the one-dimensional convolution operation; b (l) is the bias term; RELU is the activation function;

[0062] The expression of the dense connection is:

[0063] H (l) =ReLU([H (0) ,H (1) ,...,H (l-1) ]·W (l) +b (l) )

[0064] Where [·] is the feature concatenation operation;

[0065] The expression of the fully connected layer is:

[0066] Z = ReLU(H (L) ·W fc +b fc )

[0067] In the formula, H (L) is a high-dimensional feature map; W fc is the weight matrix of the fully connected layer; b fc is the bias term;

[0068] The expression of the output layer is:

[0069]

[0070] In the formula, is the critical point of the dynamic security domain; W out and b out are the weight and bias of the output layer respectively; σ(·) is the linear activation function;

[0071] The loss function is the mean square error, expressed as:

[0072]

[0073] In the formula, is the predicted value of the network; y i is the true critical point label; N is the number of samples.

[0074] As a preferred solution of the power system security domain critical point prediction device based on memory enhancement, in the Fisher information matrix enhancement module, the submodule of the Fisher information matrix for enhancing the historical parameter memory capability of the dynamic security domain critical point prediction initial model includes:

[0075] The importance value acquisition submodule is used to obtain the importance value of the set weight by calculation;

[0076] A learning rate dynamic adjustment submodule, used to dynamically adjust the learning rate according to the importance value to obtain an adjusted learning rate;

[0077] An important property dynamic update submodule, for preserving long-term memory through an exponential decay mechanism and dynamically updating the importance value;

[0078] The regularization term adding submodule is used to protect key parameters by adding regularization terms based on the Fisher information matrix in the loss function;

[0079] The weight optimization submodule is used to optimize and update the set weight according to the adjusted learning rate to obtain the optimized weight.

[0080] As a preferred solution of the power system security domain critical point prediction device based on memory enhancement, in the generated data acquisition module, in the process of reconstructing the historical data through the generation model VAE to obtain the generated data, the expression of the generated data is:

[0081] D gen ={x 1 ',x' 2 ,...,x' N},x i '~p θ (x|z),z~p(z)

[0082] Where D gen is the generated dataset generated by VAE; x i ' is the generated data; p θ (x|z) is the conditional probability distribution generated by the decoder; p(z) is the prior distribution; z is the latent variable.

[0083] As a preferred solution of the power system security domain critical point prediction device based on memory enhancement, in the dynamic security domain critical point prediction initial model training module, in the process of iteratively training the dynamic security domain critical point prediction initial model through the mixed data, the dynamic security domain critical point prediction initial model is iteratively trained through a joint loss function; the expression of the joint loss function is:

[0084] L total =L task +λ 1 L VAE

[0085] Where, L task The loss of the task for dynamic security domain prediction; L VAE is the loss of the generative model; 1 is the weight for generating the model loss.

[0086] The present invention has the following advantages: the present invention uses matlab to call PSD-BPA in batches to obtain original measurement data; based on the Dense 1D-CNN network structure, a dynamic safety domain critical point prediction initial model is constructed; the historical parameter memory capability of the dynamic safety domain critical point prediction initial model is enhanced through the Fisher information matrix; wherein the enhancement step is: obtaining the importance value of the set weight by calculation; dynamically adjusting the learning rate according to the importance value to obtain the adjusted learning rate; preserving long-term memory through an exponential decay mechanism, and dynamically updating the importance value; protecting key parameters by adding a regularization term based on the Fisher information matrix in the loss function; optimizing and updating the set weight according to the adjusted learning rate to obtain the optimized weight. The historical data is reconstructed through the generative model VAE to obtain generated data; the generated data is mixed with the historical data through playback technology to obtain mixed data; the dynamic safety domain critical point prediction initial model is iteratively trained through the mixed data to obtain a trained dynamic safety domain critical point prediction model; the original measurement data is predicted and processed through the trained dynamic safety domain critical point prediction model to obtain the dynamic safety domain critical point under the set scenario. The present invention firstly introduces the Dense 1D-CNN structure, and innovatively models the complex mapping relationship between the fault parameters, reference points, stable injection power and search direction and the critical injection power. Then, the parameters of the Dense 1D-CNN model are adjusted by regularization of the Fisher information matrix, and the key historical parameters are effectively constrained while optimizing the loss function, thereby ensuring the stability of the model. Finally, the historical data are reconstructed using a generative model (VAE), and the generated data and the real historical data are mixed and trained in proportion through playback technology, which enriches the training samples and improves the model performance under data scarcity conditions. A Dense 1D-CNN dynamic safety domain critical point prediction method based on memory enhancement and generative playback is given, which not only improves the computational efficiency, but also provides efficient and reliable technical support for dynamic safety analysis of power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0088] The structures, proportions, sizes, etc. illustrated in this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with the technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantial technical significance. Any structural modification, change in proportion or adjustment of size shall still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0089] Figure 1 A schematic flow chart of a method for predicting critical points in a power system security domain based on memory enhancement provided in Embodiment 1 of the present invention;

[0090] Figure 2 It is a schematic diagram of a specific implementation process of the power system security domain critical point prediction method based on memory enhancement provided in Example 1 of the present invention;

[0091] Figure 3 A schematic diagram of the IEEE39 structure in a possible embodiment provided in Embodiment 1 of the present invention;

[0092] Figure 4 This is a schematic diagram of the architecture of the power system security domain critical point prediction device based on memory enhancement provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0093] The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0094] Example 1

[0095] See also Figure 1 and Figure 2 Embodiment 1 of the present invention provides a method for evaluating and analyzing transient stability of a power system based on a wavelet transform image, comprising the following steps:

[0096] S1. Call PSD-BPA in batches through MATLAB to obtain the original measurement data;

[0097] S2. Based on the Dense 1D-CNN network structure, an initial model for predicting the critical point of the dynamic security domain is constructed;

[0098] S3, strengthening the historical parameter memory capability of the dynamic safety domain critical point prediction initial model through the Fisher information matrix;

[0099] S4, reconstruct the historical data through the generative model VAE to obtain generated data;

[0100] S5. Mixing the generated data with the historical data by a playback technology to obtain mixed data;

[0101] S6. Iteratively train the dynamic security domain critical point prediction initial model using the mixed data to obtain a trained dynamic security domain critical point prediction model;

[0102] S7. Predictively process the original measurement data using the trained dynamic safety domain critical point prediction model to obtain the dynamic safety domain critical point under a set scenario.

[0103] In this embodiment, in step S1, PSD-BPA is called in batches through matlab to obtain original measurement data;

[0104] The original measurement data includes: a reference point under a fault scenario, a stable injection power, a search direction and a corresponding critical injection power.

[0105] In this embodiment, in step S2, based on the Dense 1D-CNN network structure, an initial model for predicting the critical point of the dynamic security domain is constructed;

[0106] Specifically, the dynamic safety domain critical point prediction initial model includes: an input layer, a convolutional layer, a dense connection, a fully connected layer, an output layer and a loss function; the training set of the model is the relevant data obtained by the traditional dichotomy method.

[0107] The expression of the input feature vector of the input layer is:

[0108] X=[x 1 ,x 2 ,...,x n ],x i ∈R d

[0109] Where X∈R n×d is the input feature vector; n is the number of features (such as fault parameter information, reference point, stable injection power, search direction); d is the dimension of each feature (can be a one-dimensional scalar or a multi-dimensional array);

[0110] The expression of the convolution kernel of the convolution layer is:

[0111] H (l) =RELU(W (l) *H (l-1) +b (l) )

[0112] In the formula, H (l-1) is the output of the previous layer (for the first layer, H (0) =X); W (l) ∈R k×d is the convolution kernel weight of the lth layer; * is the one-dimensional convolution operation; b (l) is the bias term; RELU is the activation function;

[0113] ReLU(z)=max(0,z)

[0114] The expression of the dense connection is:

[0115] H (l) =ReLU([H (0) ,H (1) ,...,H (l-1) ]·W (l) +b (l) )

[0116] Where [·] is the feature concatenation operation;

[0117] The expression of the fully connected layer is:

[0118] Z = ReLU(H (L) ·W fc +b fc )

[0119] In the formula, H (L) is a high-dimensional feature map; W fc is the weight matrix of the fully connected layer; b fc is the bias term;

[0120] The expression of the output layer is:

[0121]

[0122] In the formula, is the critical point of the dynamic security domain; W out and b out are the weight and bias of the output layer respectively; σ(·) is the linear activation function;

[0123] The loss function is the mean square error, expressed as:

[0124]

[0125] In the formula, is the predicted value of the network; y i is the true critical point label; N is the number of samples.

[0126] In this embodiment, the input layer is used to receive multi-dimensional input features such as fault parameters, reference points, stable injection power and search direction. The hidden layer adopts a 4-layer densely connected convolutional layer structure. The number of convolution kernels in each layer is 128, 64, 32 and 16 respectively, and the convolution kernel size is set to 3. The activation function uses ReLU to enhance the nonlinear fitting ability and accelerate the training convergence. The output layer activation function uses a linear function to output the predicted value of the critical point of the dynamic safety domain. During the model training process, the loss function uses mean square error (MSE), the optimization algorithm uses Adam, the learning rate is set to 0.001, the batch size is 64, and the training rounds are 200. In order to improve the generalization ability of the model and prevent overfitting, Dropout regularization is added to the hidden layer, the ratio is set to 0.3, and L2 regularization is introduced, and the weight attenuation coefficient is 0.0005.

[0127] In this embodiment, in step S3, the historical parameter memory capability of the dynamic safety domain critical point prediction initial model is enhanced by using the Fisher information matrix;

[0128] Specifically, by introducing the Fisher information matrix, an adaptive memory mechanism based on parameter importance can be provided for the initial model of dynamic safety domain critical point prediction, thereby strengthening the network's memory and dependence on key parameters. During the training process, the prediction accuracy and stability can be improved by optimizing the Fisher information matrix.

[0129] The strengthening steps are:

[0130] S31, obtaining the importance value of the set weight by calculation;

[0131] Specifically, the calculation formula of the importance value is:

[0132]

[0133] In the formula, ω j is the weight; j is the weight ω j The importance value of D s is the number of samples in the current data set; x i is the input data; i Output for the target; is the loss function L for weight ω j gradient.

[0134] S32, dynamically adjusting the learning rate according to the importance value to obtain an adjusted learning rate;

[0135] Specifically, according to the importance value ρ of the weight j Dynamically adjust the learning rate η j , to reduce the impact on key parameters;

[0136] η j =η o ·(1-α·ρ j )

[0137] Where η o is the initial learning rate; α is a tuning parameter used to control the dynamic range of the learning rate; for important weights (ρ j Larger), η j Small to avoid over-update; for unimportant weights (ρ j Smaller), η j Larger, improving adaptability.

[0138] S33, preserving long-term memory through an exponential decay mechanism, and dynamically updating the importance value;

[0139] Specifically, the exponential decay mechanism is used to preserve long-term memory and dynamically update the importance of weights:

[0140]

[0141] Where β is the decay rate that controls the importance of historical weights.

[0142] S34, by adding a regularization term based on the Fisher information matrix in the loss function, the key parameters are protected;

[0143] Specifically, a regularization term based on the Fisher information matrix is ​​added to the loss function to protect key parameters:

[0144]

[0145] Where, L total is the total loss function, L new is the loss function of the current task; λ is the regularization coefficient, which weighs the predicted new data loss and the historical data parameter protection; F ii is the Fisher information matrix for parameter θ i The importance estimate of i is the i-th parameter in the newly predicted data; is the best parameter value in historical data.

[0146] S35. Optimize and update the set weights according to the adjusted learning rate to obtain optimized weights.

[0147] Specifically, based on the dynamically adjusted learning rate η j , optimize the weights:

[0148]

[0149] Where η j is the learning rate.

[0150] In this embodiment, in step S4, the historical data is reconstructed by the generative model VAE to obtain generated data;

[0151] Among them, variational autoencoder (VAE) is a generative model that is usually used to learn the potential distribution of data and generate new samples from it. The core of VAE is to map the input sample to the latent space through the encoder, and then reconstruct the input sample from the latent variable through the decoder.

[0152] Specifically, given an input sample x, it is mapped to the distribution parameters of the latent space through the encoder. The expression of this process is:

[0153] z~q φ (z|x)

[0154] In the formula, z is the latent variable; q φ (z|x) is the approximate posterior distribution learned by the encoder network, usually assumed to be a Gaussian distribution:

[0155] q φ (z|x)=N(z|μ(x),σ 2 (x)

[0156] Where μ(x) and σ(x) are the mean and standard deviation calculated by the encoder network, respectively.

[0157] Generate a reconstruction of the sample by sampling the latent variable z from the latent space:

[0158] x'~p θ (x|z)

[0159] In the formula, p θ (x|z) is the conditional probability distribution generated by the decoder, usually a Gaussian distribution:

[0160] p θ (x|z)=N(x|f θ (z),σ 2 I)

[0161] In the formula, f θ(z) is the output of reconstructing the input x from the latent variable z through the decoder neural network.

[0162] The loss function of VAE consists of two parts: reconstruction error and KL divergence. The reconstruction error ensures that the reconstructed sample x' is consistent with the original sample x, and the KL divergence makes the potential distribution q φ (z|x) is close to the prior distribution p(z):

[0163]

[0164] Wherein, the first term is the log-likelihood reconstruction error, which measures the difference between the generated samples and the real samples; the second term is the KL divergence, which measures the difference between the potential distribution and the prior distribution.

[0165] By training VAE, we can capture the distribution of historical data and generate historical samples that are close to reality to supplement the training data set:

[0166] D gen ={x 1 ',x' 2 ,...,x' N},x i '~p θ (x|z),z~p(z)

[0167] Where D gen is the generated dataset generated by VAE; x i 'Generate data for the generation.

[0168] In this embodiment, in step S5, the generated data is mixed with the historical data by a playback technology to obtain mixed data;

[0169] Specifically, the VAE-based replay technology optimizes the initial model for predicting the critical point of the dynamic safety domain by combining historical data and current generated data through the associative replay mechanism:

[0170] D train =D old ∪D gen

[0171] Where D train is a mixed data set; D old is a historical data set; D gen Generative datasets generated for generative models.

[0172] In this embodiment, in step S6, the dynamic security domain critical point prediction initial model is iteratively trained through the mixed data to obtain a trained dynamic security domain critical point prediction model;

[0173] Specifically, the dynamic safety domain critical point prediction initial model is iteratively trained through a joint loss function; the expression of the joint loss function is:

[0174] L total =L task +λ 1 L VAE

[0175] Where, L task The loss of the task for dynamic security domain prediction; L VAE is the loss of the generative model; 1 is the weight for generating the model loss.

[0176] The initial model for dynamic safety domain critical point prediction is updated by using mixed data training until the dynamic safety domain critical point error reaches the expected value, thus obtaining a trained dynamic safety domain critical point prediction model.

[0177] In this embodiment, in step S7, the original measurement data is predicted and processed by using the trained dynamic safety domain critical point prediction model to obtain the dynamic safety domain critical point in a set scenario.

[0178] Specifically, the trained dynamic safety domain critical point prediction model is used to perform prediction processing on the original measurement data to predict the dynamic safety domain critical point under a set scenario.

[0179] In a possible embodiment, an example of prediction of critical points in the power system security domain is provided as follows:

[0180] The entire dynamic safety domain critical point prediction experiment is carried out on the IEEE 39 standard example. Figure 3 As shown. The selected fault scenarios are Case 1: a three-phase short circuit fault occurs at the head end of line 10-11, and the fault duration is 6 cycles; Case 2: a three-phase short circuit fault occurs at the head end of line 26-27, and the fault duration is 6 cycles; Case 3: a three-phase short circuit fault occurs at the head end of line 28-29, and the fault duration is 6 cycles. All experiments were completed on an Intel(R) Core(TM) i7-14700F desktop computer. The Dense1D-CNN model was implemented and trained using the Pytorch framework, using GPU acceleration, with a learning rate set to 0.001, an optimizer of Adam, a batch size of 64, and 200 epochs.

[0181] The experimental results are shown in Table 1:

[0182]

[0183]

[0184] Table 1 Comparison of performance between traditional method and the method of the present invention

[0185] By using the techniques of generation replay and memory enhancement, the present invention can generate more critical points and significantly improve the accuracy of critical point prediction. In Case 3, the error was reduced to 1.39e-15, which is much lower than the 2.81e-6 of the traditional method. In addition, the present invention also improves the computational efficiency, especially in Case 2 and Case 3, the computational time is reduced by about 54% and 42%, respectively. In general, the present invention shows significant advantages in accuracy and computational efficiency when predicting critical points by utilizing Dense1D-CNN, combined with memory enhancement and generation replay techniques. Compared with the traditional dichotomy method to obtain critical point data, the present invention can not only predict more dynamic safety domain critical points, but also provide more accurate prediction results, while greatly shortening the calculation time. These advantages make the present invention have significant application value in dynamic safety analysis and real-time early warning of power systems, and can effectively support the stability assessment and emergency decision-making of the system.

[0186] In summary, the present invention batch calls PSD-BPA through matlab to obtain original measurement data; constructs a dynamic safety domain critical point prediction initial model based on the Dense1D-CNN network structure; strengthens the historical parameter memory ability of the dynamic safety domain critical point prediction initial model through the Fisher information matrix; wherein, the strengthening step is: obtain the importance value of the set weight by calculation; dynamically adjust the learning rate according to the importance value to obtain the adjusted learning rate; preserve long-term memory through the exponential decay mechanism, and dynamically update the importance value; protect key parameters by adding a regularization term based on the Fisher information matrix in the loss function; optimize and update the set weight according to the adjusted learning rate to obtain the optimized weight. The historical data is reconstructed through the generative model VAE to obtain generated data; the generated data is mixed with the historical data through playback technology to obtain mixed data; the dynamic safety domain critical point prediction initial model is iteratively trained through the mixed data to obtain a trained dynamic safety domain critical point prediction model; the original measurement data is predicted and processed through the trained dynamic safety domain critical point prediction model to obtain the dynamic safety domain critical point under the set scenario. The present invention firstly introduces the Dense 1D-CNN structure, and innovatively models the complex mapping relationship between the fault parameters, reference points, stable injection power and search direction and the critical injection power. Then, the parameters of the Dense 1D-CNN model are adjusted by regularization of the Fisher information matrix, and the key historical parameters are effectively constrained while optimizing the loss function, thereby ensuring the stability of the model. Finally, the historical data are reconstructed using a generative model (VAE), and the generated data and the real historical data are mixed and trained in proportion through playback technology, which enriches the training samples and improves the model performance under data scarcity conditions. A Dense 1D-CNN dynamic safety domain critical point prediction method based on memory enhancement and generative playback is given, which not only improves the computational efficiency, but also provides efficient and reliable technical support for dynamic safety analysis of power systems.

[0187] It should be noted that the method of the embodiment of the present disclosure can be performed by a single device, such as a computer or a server. The method of the present embodiment can also be applied in a distributed scenario and completed by multiple devices cooperating with each other. In the case of such a distributed scenario, one of the multiple devices can only perform one or more steps in the method of the embodiment of the present disclosure, and the multiple devices will interact with each other to complete the described method.

[0188] It should be noted that the above describes some embodiments of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0189] Example 2

[0190] See also Figure 4 Embodiment 2 of the present invention also provides a power system security domain critical point prediction device based on memory enhancement, including:

[0191] The original measurement data acquisition module 001 is used to call PSD-BPA in batches through matlab to obtain the original measurement data;

[0192] Dynamic security domain critical point prediction initial model construction module 002, used to construct a dynamic security domain critical point prediction initial model based on a Dense 1D-CNN network structure;

[0193] Fisher information matrix strengthening module 003, used for strengthening the historical parameter memory capability of the dynamic safety domain critical point prediction initial model through the Fisher information matrix;

[0194] A generated data acquisition module 004 is used to reconstruct historical data through a generation model VAE to obtain generated data;

[0195] A mixed data acquisition module 005 is used to mix the generated data with the historical data through a playback technology to obtain mixed data;

[0196] A dynamic security domain critical point prediction initial model training module 006 is used to iteratively train the dynamic security domain critical point prediction initial model through the mixed data to obtain a trained dynamic security domain critical point prediction model;

[0197] The dynamic safety domain critical point acquisition module 007 is used to perform prediction processing on the original measurement data through the trained dynamic safety domain critical point prediction model to obtain the dynamic safety domain critical point under a set scenario.

[0198] In this embodiment, in the original measurement data acquisition module 001, the original measurement data includes: a reference point in a fault scenario, a stable injection power, a search direction and a corresponding critical injection power.

[0199] In this embodiment, in the dynamic safety domain critical point prediction initial model construction module 002, the dynamic safety domain critical point prediction initial model includes: an input layer, a convolution layer, a dense connection, a fully connected layer, an output layer and a loss function;

[0200] The expression of the input feature vector of the input layer is:

[0201] X=[x 1 ,x 2 ,...,x n ],x i ∈R d

[0202] Where X∈R n×d is the input feature vector; n is the number of features; d is the dimension of each feature;

[0203] The expression of the convolution kernel of the convolution layer is:

[0204] H (l) =RELU(W (l) *H (l-1) +b (l) )

[0205] In the formula, H (l-1) is the output of the previous layer (for the first layer, H (0) =X); W (l) ∈R k×d is the convolution kernel weight of the lth layer; * is the one-dimensional convolution operation; b (l) is the bias term; RELU is the activation function;

[0206] The expression of the dense connection is:

[0207] H (l) =ReLU([H (0) ,H (1) ,...,H (l-1) ]·W (l) +b (l) )

[0208] Where [·] is the feature concatenation operation;

[0209] The expression of the fully connected layer is:

[0210] Z = ReLU(H (L) ·W fc +b fc )

[0211] In the formula, H (L) is a high-dimensional feature map; W fc is the weight matrix of the fully connected layer; b fc is the bias term;

[0212] The expression of the output layer is:

[0213]

[0214] In the formula, is the critical point of the dynamic security domain; W out and b out are the weight and bias of the output layer respectively; σ(·) is the linear activation function;

[0215] The loss function is the mean square error, expressed as:

[0216]

[0217] In the formula, is the predicted value of the network; y i is the true critical point label; N is the number of samples.

[0218] In this embodiment, in the Fisher information matrix strengthening module 003, the submodule of the Fisher information matrix for strengthening the historical parameter memory capability of the dynamic safety domain critical point prediction initial model includes:

[0219] The importance value acquisition submodule 031 is used to obtain the importance value of the set weight by calculation;

[0220] A learning rate dynamic adjustment submodule 032, used to dynamically adjust the learning rate according to the importance value to obtain an adjusted learning rate;

[0221] The importance property dynamic update submodule 033 is used to store long-term memory through an exponential decay mechanism and dynamically update the importance value;

[0222] A regularization term adding submodule 034 is used to protect key parameters by adding a regularization term based on the Fisher information matrix in the loss function;

[0223] The weight optimization submodule 035 is used to optimize and update the set weight according to the adjusted learning rate to obtain the optimized weight.

[0224] In this embodiment, in the generated data acquisition module 004, in the process of reconstructing the historical data through the generative model VAE to obtain the generated data, the expression of the generated data is:

[0225] D gen ={x 1 ',x' 2 ,...,x' N},x i'~p θ (x|z),z~p(z)

[0226] Where D gen is the generated dataset generated by VAE; x i ' is the generated data; p θ (x|z) is the conditional probability distribution generated by the decoder; p(z) is the prior distribution; z is the latent variable.

[0227] In this embodiment, in the dynamic security domain critical point prediction initial model training module 006, in the process of iteratively training the dynamic security domain critical point prediction initial model through the mixed data, the dynamic security domain critical point prediction initial model is iteratively trained through a joint loss function; the expression of the joint loss function is:

[0228] L total =L task +λ 1 L VAE

[0229] Where, L task The loss of the task for dynamic security domain prediction; L VAE is the loss of the generative model; 1 is the weight for generating the model loss.

[0230] It should be noted that the information interaction, execution process and other contents between the modules of the above-mentioned system are based on the same concept as the method embodiment in Example 1 of the present application, and the technical effects they bring are the same as those of the method embodiment of the present application. For specific contents, please refer to the description in the method embodiment shown above in the present application, and will not be repeated here.

[0231] Example 3

[0232] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium, in which a program code of a method for predicting critical points in the safety domain of a power system based on memory enhancement is stored, and the program code includes instructions for executing the method for predicting critical points in the safety domain of a power system based on memory enhancement of embodiment 1 or any possible implementation thereof.

[0233] The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that includes one or more available media. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0234] Example 4

[0235] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0236] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the memory-enhanced power system security domain critical point prediction method of Example 1 or any possible implementation thereof.

[0237] Specifically, the processor can be implemented by hardware or by software. When implemented by hardware, the processor can be a logic circuit, an integrated circuit, etc.; when implemented by software, the processor can be a general-purpose processor implemented by reading software codes stored in a memory. The memory can be integrated in the processor or can be located outside the processor and exist independently.

[0238] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable systems. The computer instructions can be stored in a computer-readable storage medium, or transmitted from a computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center.

[0239] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing system, they can be concentrated on a single computing system, or distributed on a network composed of multiple computing systems, and optionally, they can be implemented by a program code executable by a computing system, so that they can be stored in a storage system and executed by the computing system, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0240] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A power system security domain critical point prediction method based on memory enhancement, characterized in that: include: Call PSD-BPA in batches through MATLAB to obtain the original measurement data; Based on the Dense 1D-CNN network structure, an initial model for predicting the critical point of the dynamic security domain is constructed; The historical parameter memory capability of the dynamic safety domain critical point prediction initial model is enhanced through the Fisher information matrix; Reconstruct historical data through the generative model VAE to obtain generated data; Mixing the generated data with the historical data by a playback technology to obtain mixed data; Iteratively training the dynamic safety domain critical point prediction initial model through the mixed data to obtain a trained dynamic safety domain critical point prediction model; The original measurement data is predicted and processed by the trained dynamic safety domain critical point prediction model to obtain the dynamic safety domain critical point under the set scenario.

2. The method for predicting critical points of power system security domain based on memory enhancement according to claim 1 is characterized in that: The original measurement data includes: a reference point under a fault scenario, a stable injection power, a search direction and a corresponding critical injection power.

3. The method for predicting critical points of power system security domain based on memory enhancement according to claim 2 is characterized in that: The dynamic safety domain critical point prediction initial model includes: an input layer, a convolution layer, a dense connection, a fully connected layer, an output layer and a loss function; The expression of the input feature vector of the input layer is: X=[x1,x2,…,x n ],x i ∈R d Where X∈R n×d is the input feature vector; n is the number of features; d is the dimension of each feature; The expression of the convolution kernel of the convolution layer is: A (l) =RELU(W (l) *H (l-1) +b (l) ) In the formula, H (l-1) is the output of the previous layer (for the first layer, H (0) =X); W (l) ∈R k×d is the convolution kernel weight of the lth layer; * is the one-dimensional convolution operation; b (l) is the bias term; RELU is the activation function; The expression of the dense connection is: A (l) =ReLU([H (0) ,H (1) ,...,H (l-1) ]·W (l) +b (l) ) Where [·] is the feature concatenation operation; The expression of the fully connected layer is: Z=ReLU(H (L) ·IN fc +b fc ) In the formula, H (L) is a high-dimensional feature map; W fc is the weight matrix of the fully connected layer; b fc is the bias term; The expression of the output layer is: In the formula, is the critical point of the dynamic security domain; W out and b out are the weight and bias of the output layer respectively; σ(·) is the linear activation function; The loss function is the mean square error, expressed as: In the formula, is the predicted value of the network; y i is the true critical point label; N is the number of samples.

4. The method for predicting critical points of power system security domain based on memory enhancement according to claim 3 is characterized in that: In the process of strengthening the historical parameter memory capability of the dynamic safety domain critical point prediction initial model through the Fisher information matrix, the strengthening steps are: Obtain the importance value of the set weight by calculation; Dynamically adjusting the learning rate according to the importance value to obtain an adjusted learning rate; Preserving long-term memory through an exponential decay mechanism and dynamically updating the importance value; By adding a regularization term based on the Fisher information matrix to the loss function, key parameters are protected; According to the adjusted learning rate, the set weight is optimized and updated to obtain the optimized weight.

5. The method for predicting critical points of power system security domain based on memory enhancement according to claim 4 is characterized in that: In the process of reconstructing the historical data through the generative model VAE to obtain the generated data, the expression of the generated data is: D gen ={x1',x'2,...,x' N },x i '~p θ (x|z),z~p(z) Where D gen is the generated dataset generated by VAE; x i ' is the generated data; p θ (x|z) is the conditional probability distribution generated by the decoder; p(z) is the prior distribution; z is the latent variable.

6. The method for predicting critical points of power system security domain based on memory enhancement according to claim 5 is characterized in that: In the process of iteratively training the dynamic safety domain critical point prediction initial model through the mixed data, the dynamic safety domain critical point prediction initial model is iteratively trained through a joint loss function; the expression of the joint loss function is: L total =L task +λ1L VAE Where, L task The loss of the task for dynamic security domain prediction; L VAE is the loss of the generative model; λ1 is the weight of the generative model loss.

7. A power system security domain critical point prediction device based on memory enhancement, adopting a power system security domain critical point prediction method based on memory enhancement according to any one of claims 1 to 6, characterized in that: include: The original measurement data acquisition module is used to call PSD-BPA in batches through Matlab to obtain the original measurement data; Dynamic security domain critical point prediction initial model construction module, used to build a dynamic security domain critical point prediction initial model based on the Dense 1D-CNN network structure; A Fisher information matrix strengthening module, used for strengthening the historical parameter memory capability of the dynamic safety domain critical point prediction initial model through the Fisher information matrix; The generated data acquisition module is used to reconstruct historical data through the generation model VAE to obtain generated data; A mixed data acquisition module, used for mixing the generated data with the historical data through a playback technology to obtain mixed data; A dynamic safety domain critical point prediction initial model training module is used to iteratively train the dynamic safety domain critical point prediction initial model through the mixed data to obtain a trained dynamic safety domain critical point prediction model; The dynamic safety domain critical point acquisition module is used to predict the original measurement data through the trained dynamic safety domain critical point prediction model to obtain the dynamic safety domain critical point under a set scenario.

8. The power system security domain critical point prediction device based on memory enhancement according to claim 7 is characterized in that: In the original measurement data acquisition module, the original measurement data includes: a reference point under a fault scenario, a stable injection power, a search direction and a corresponding critical injection power.

9. The power system security domain critical point prediction device based on memory enhancement according to claim 8 is characterized in that: In the dynamic safety domain critical point prediction initial model construction module, the dynamic safety domain critical point prediction initial model includes: an input layer, a convolution layer, a dense connection, a fully connected layer, an output layer and a loss function; The expression of the input feature vector of the input layer is: X=[x1,x2,…,x n ],x i ∈R d Where X∈R n×d is the input feature vector; n is the number of features; d is the dimension of each feature; The expression of the convolution kernel of the convolution layer is: A (l) =RELU(W (l) *H (l-1) +b (l) ) In the formula, H (l-1) is the output of the previous layer (for the first layer, H (0) =X); W (l) ∈R k×d is the convolution kernel weight of the lth layer; * is the one-dimensional convolution operation; b (l) is the bias term; RELU is the activation function; The expression of the dense connection is: A (l) =ReLU([H (0) ,H (1) ,...,H (l-1) ]·W (l) +b (l) ) Where [·] is the feature concatenation operation; The expression of the fully connected layer is: Z=ReLU(H (L) ·IN fc +b fc ) In the formula, H (L) is a high-dimensional feature map; W fc is the weight matrix of the fully connected layer; b fc is the bias term; The expression of the output layer is: In the formula, is the critical point of the dynamic security domain; W out and b out are the weight and bias of the output layer respectively; σ(·) is the linear activation function; The loss function is the mean square error, expressed as: In the formula, is the predicted value of the network; y i is the true critical point label; N is the number of samples.

10. The power system security domain critical point prediction device based on memory enhancement according to claim 9 is characterized in that: In the Fisher information matrix strengthening module, the submodule of the Fisher information matrix for strengthening the historical parameter memory capability of the dynamic safety domain critical point prediction initial model includes: The importance value acquisition submodule is used to obtain the importance value of the set weight by calculation; A learning rate dynamic adjustment submodule, used to dynamically adjust the learning rate according to the importance value to obtain an adjusted learning rate; An important property dynamic update submodule, for preserving long-term memory through an exponential decay mechanism and dynamically updating the importance value; The regularization term adding submodule is used to protect key parameters by adding regularization terms based on the Fisher information matrix in the loss function; The weight optimization submodule is used to optimize and update the set weight according to the adjusted learning rate to obtain the optimized weight.

Images