A method and system for evaluating power grid resilience

Abstract

The present invention discloses a method and system for evaluating the resilience of a power grid, wherein the method comprises: inputting the collected monitoring index data and environmental index data of the power grid in the target area into a pre-trained power grid resilience evaluation model for evaluation and identification to obtain a resilience evaluation result; the monitoring index data comprises voltage data, current data, harmonic data and power data, and the environmental index data comprises temperature and humidity data, line loss data and frequency data; wherein the construction process of the power grid resilience evaluation model comprises respectively training the data expansion process, feature extraction model, feature dimensionality reduction model and classifier model therein to obtain the power grid resilience evaluation model. The present invention improves the accuracy of the evaluation of the evaluation model by optimizing the training process of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power maintenance equipment, and in particular to a power grid resilience assessment method and system. Background Art

[0002] With the rapid development of power grid systems and the improvement of their intelligence level, the safe and stable operation of campus power grids is particularly important. The facilities and equipment involved in campus power grids are complex, and they need to deal with various emergencies and environmental changes. This requires an effective assessment of the resilience of the power grid to prevent and mitigate the impact of potential failures or disasters.

[0003] Existing grid resilience assessment methods mostly rely on traditional data processing technology and simple mathematical models, which have certain limitations in terms of in-depth data analysis and adaptability to complex environments. In the process of feature extraction and dimensionality reduction of grid data, existing technologies cannot fully reflect the characteristics of data, which reduces the stability of the model and the accuracy of prediction. In addition, when processing grid data, traditional assessment models are prone to overfitting or inability to adapt to complex data structures. The robustness and accuracy of the model are low, making the final assessment results inaccurate, and thus unable to prevent faults in a timely manner.

[0004] It can be seen that how to effectively evaluate and analyze the resilience of the power grid has become a technical problem that needs to be urgently solved by technical personnel in this field. Summary of the invention

[0005] The present invention provides a power grid resilience assessment method and system to solve the problem of how to effectively assess and analyze power grid resilience and prevent power grid failures in a timely manner.

[0006] In order to solve the above technical problems, an embodiment of the present invention provides a power grid resilience assessment method, including:

[0007] Input the collected power monitoring index data and line environment index data of the target area power grid into the pre-trained power grid resilience assessment model for assessment and identification to obtain the resilience assessment results;

[0008] Among them, the power grid resilience assessment model is designed to be trained through data expansion, feature extraction, feature dimension reduction and classifier classification during the construction process to obtain the trained power grid resilience assessment model.

[0009] Furthermore, the construction process of the power grid resilience assessment model includes:

[0010] In the first stage, a macro-micro dual-kernel function is constructed to cluster the power monitoring index data and the line environment index data as sample data to obtain a preliminary clustering result, and a clustering characteristic analysis is performed on the preliminary clustering result to obtain a clustering characteristic analysis result;

[0011] Based on the clustering characteristic analysis result, the sample data is expanded using the SMO algorithm to obtain an indicator expansion data set;

[0012] In the second stage, the index expansion data set is input into the neural network model for training, the parameters of the model are optimized through the chaotic feedback mechanism and the regularization strategy to obtain a feature extraction model, and the first index feature data set is output based on the feature extraction model;

[0013] In the third step, the first indicator feature data set is input into the autoencoder and parameter reconstruction and optimization are performed through dynamic routing decision to construct a feature dimension reduction model, and a second indicator feature data set is output based on the feature dimension reduction model;

[0014] In the fourth stage, the second indicator feature data set is input into a high-order neural network model for training, and the model gradient update rate is dynamically adjusted to construct the power grid resilience assessment model.

[0015] Furthermore, the training process of the feature extraction model includes:

[0016] In the first training process, the learning rate of the neural network model is adjusted by the chaotic feedback mechanism, and the number of hidden layer nodes of the neural network model is dynamically adjusted based on the regularization strategy;

[0017] The first training process is iterated repeatedly until the parameter standard of the preset neural network model output is met, thereby obtaining the feature extraction model.

[0018] Furthermore, the training process of the feature dimensionality reduction model includes:

[0019] In the second training process, after the first indicator feature data set is input into the autoencoder, a dynamic routing decision is made on the importance of the output features of the activation function of each layer of the autoencoder through a dynamic routing function;

[0020] Reconstructing the output features using the loss function to obtain corresponding reconstruction errors, and updating and optimizing the autoencoder parameters using a back propagation algorithm based on the reconstruction errors;

[0021] The second training process is iterated repeatedly until the parameter standard of the preset autoencoder output is met, thereby obtaining the feature dimension reduction model.

[0022] Furthermore, the training process of the power grid resilience assessment model includes:

[0023] In the third training process, the activation output of each layer of the high-order neural network model is calculated based on the historical gradient accumulation information of the high-order neural network model;

[0024] Based on the activation output, calculating and iteratively adjusting the error gradient and target gradient accumulation information of each layer of the high-order neural network model;

[0025] When the number of iterations reaches the preset threshold, weight pruning and parameter optimization are performed;

[0026] The third training process is iterated repeatedly until the parameter standard of the preset high-order neural network model output is met, thereby obtaining the power grid resilience assessment model.

[0027] Another embodiment of the present invention provides a power grid resilience assessment system, the system being specifically used for:

[0028] Input the collected power monitoring index data and line environment index data of the target area power grid into the pre-trained power grid resilience assessment model for assessment and identification to obtain the resilience assessment results;

[0029] Among them, the power grid resilience assessment model is designed to be trained through data expansion, feature extraction, feature dimension reduction and classifier classification during the construction process to obtain the trained power grid resilience assessment model.

[0030] Furthermore, the system includes: a data expansion module, a feature extraction model construction module, a feature dimension reduction model construction module and an evaluation model construction module; wherein,

[0031] The data expansion module is used to construct a macro-micro dual-core function in the first stage to cluster the power monitoring index data and the line environment index data as sample data to obtain a preliminary clustering result, perform clustering characteristic analysis on the preliminary clustering result, and obtain a clustering characteristic analysis result; and

[0032] Based on the clustering characteristic analysis result, the sample data is expanded using the SMO algorithm to obtain an indicator expansion data set;

[0033] The feature extraction model building module is used to input the indicator expansion data set into the neural network model for training in the second stage, optimize the parameters of the model through the chaotic feedback mechanism and regularization strategy to obtain the feature extraction model, and output the first indicator feature data set based on the feature extraction model;

[0034] The feature dimension reduction model construction module is used to input the first indicator feature data set into the autoencoder and perform parameter reconstruction optimization through dynamic routing decision to construct a feature dimension reduction model in the third stage, and output a second indicator feature data set based on the feature dimension reduction model;

[0035] The evaluation model construction module is used to input the second indicator feature data set into a high-order neural network model for training in the fourth stage, and dynamically adjust the model gradient update rate to construct the power grid resilience evaluation model.

[0036] Furthermore, it is characterized in that the feature extraction model construction module is specifically used for:

[0037] In the first training process, the learning rate of the neural network model is adjusted by the chaotic feedback mechanism, and the number of hidden layer nodes of the neural network model is dynamically adjusted based on the regularization strategy;

[0038] The first training process is iterated repeatedly until the parameter standard of the preset neural network model output is met, thereby obtaining the feature extraction model.

[0039] Furthermore, the feature dimension reduction model construction module is specifically used to:

[0040] In the second training process, after the first indicator feature data set is input into the autoencoder, a dynamic routing decision is made on the importance of the output features of the activation function of each layer of the autoencoder through a dynamic routing function;

[0041] Reconstructing the output features using the loss function to obtain corresponding reconstruction errors, and updating and optimizing the autoencoder parameters using a back propagation algorithm based on the reconstruction errors;

[0042] The second training process is iterated repeatedly until the parameter standard of the preset autoencoder output is met, thereby obtaining the feature dimension reduction model.

[0043] Furthermore, the evaluation model building module is specifically used to:

[0044] In the third training process, the activation output of each layer of the high-order neural network model is calculated based on the historical gradient accumulation information of the high-order neural network model;

[0045] Based on the activation output, calculating and iteratively adjusting the error gradient and target gradient accumulation information of each layer of the high-order neural network model;

[0046] When the number of iterations reaches the preset threshold, weight pruning and parameter optimization are performed;

[0047] The third training process is iterated repeatedly until the parameter standard of the preset high-order neural network model output is met, thereby obtaining the power grid resilience assessment model.

[0048] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0049] (1) In the process of data expansion, the dynamic SMOTE algorithm with dual-core clustering is adopted. By dual analysis of the macro and micro structures of the data set, the coverage and detail of the samples in the data expansion process are enhanced, thereby effectively improving the representativeness and classification accuracy of the model training samples.

[0050] (2) During the training of the feature extraction model, the parameters of the neural network model are precisely adjusted through the chaos optimization algorithm, ensuring the stability and efficiency of the model in a complex power grid data environment. During the training of the feature dimension reduction model, the dynamic routing autoencoder is used to effectively optimize the feature transmission path, reduce the information loss in the feature processing process, and improve the data reconstruction accuracy and the generalization ability of the model. Finally, during the training of the resilience assessment model, the dynamic adjustment of the gradient update rate is used to enhance the adaptability of the model under various data conditions, thereby enhancing the judgment accuracy of the assessment model under different power grid conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic diagram of the process of constructing a power grid resilience assessment model in one embodiment of the present invention;

[0052] Figure 2 is a schematic diagram of a feature extraction model training process in one embodiment of the present invention;

[0053] Figure 3 It is a schematic diagram of the feature dimension reduction model training process in one embodiment of the present invention;

[0054] Figure 4 is a schematic diagram of the training and evaluation process of each model in one embodiment of the present invention;

[0055] Figure 5 It is a structural block diagram of a power grid resilience assessment system in one embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. 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.

[0057] In the description of this application, the terms "first", "second", "third", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first", "second", "third", etc. may explicitly or implicitly include one or more of the feature. In the description of this application, unless otherwise specified, "plurality" means two or more.

[0058] In the description of the present application, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the two elements. The terms "vertical", "horizontal", "left", "right", "upper", "lower" and similar expressions used herein are only for illustrative purposes, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0059] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as those commonly understood by those skilled in the art. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood by specific circumstances.

[0060] An embodiment of the present invention provides a power grid resilience assessment method, which inputs the collected power monitoring index data and line environment index data of the target area power grid into a pre-trained power grid resilience assessment model for assessment and identification to obtain a resilience assessment result.

[0061] Among them, the power grid resilience assessment model is designed to be trained through data expansion, feature extraction, feature dimension reduction and classifier classification during the construction process to obtain the trained power grid resilience assessment model.

[0062] It is understandable that in the process of model training, before data expansion, data collection and labeling must be performed first. In an embodiment of the present invention, the index data that can be collected in the target park power grid include the power monitoring index data and line environment index data of the power grid, that is, in an embodiment of the present invention, data is collected for parameters such as voltage, current, power, and equipment such as power grid monitoring equipment or dispatching equipment, as well as equipment failure, maintenance, and replacement records; and line aging data of power grid transmission lines, branch overload degree data, and temperature and humidity data of environmental monitoring equipment are collected.

[0063] Among them, the power monitoring indicator data includes, but is not limited to, voltage , Current , Harmonic content , Power Factor , Load distribution and load factor The line environment index data include but are not limited to line loss ,frequency ,temperature and humidity In some embodiments of the present invention, the collected data are stored in a structured JSON format.

[0064] In some embodiments of the present invention, as an example, the attribute value of the indicator data is, voltage is 220 volts, current For 50 amps, the harmonic content The power factor is 5.5%. is 0.98, load distribution The load factor is 45%. 75%, line loss 2%, frequency 50Hz, temperature 30 degrees Celsius, humidity It is 45%.

[0065] It should be noted that in practical applications, the attributes of data are usually more than 10 attributes, and the number of attributes of data may reach dozens or even hundreds.

[0066] Furthermore, the collected data is labeled. In some embodiments of the present invention, the labeling method is manual labeling. As an example, the labeling categories are: normal operation, warning state and fault state.

[0067] After collecting and labeling the indicator data, data expansion, feature extraction, and feature dimension reduction are performed, and then input into the classifier model to train the power grid resilience assessment model. After the power grid resilience assessment model is trained, new indicator data samples are input to conduct resilience assessment and obtain the assessment results.

[0068] In the embodiments of the present invention, specifically, please refer to Figure 1 As shown, Figure 1 1 is a schematic diagram of a process flow of constructing a power grid resilience assessment model in one embodiment of the present invention. It can be seen that the process of constructing a power grid resilience assessment model includes the following steps:

[0069] S1. In the first stage, a macro-micro dual-kernel function is constructed to cluster the power monitoring index data and the line environment index data as sample data to obtain a preliminary clustering result, and a clustering characteristic analysis is performed on the preliminary clustering result to obtain a clustering characteristic analysis result. It can be understood that the first stage is the data expansion stage.

[0070] As you can understand, the kernel function allows data to be mapped to a high-dimensional space, making problems that are difficult to handle in the original space more tractable.

[0071] In an embodiment of the present invention, the input power grid index sample data is mapped to the feature space defined by two kernels. First, two sets of kernel functions are constructed, namely, the macro kernel function based on the macro structure of the index data and the micro kernel function based on the micro structure, wherein the macro kernel function is represented by the following formula:

[0072]

[0073] in, and are the first data sample and the second data sample respectively, is the bandwidth parameter of the macro kernel function, is the macro kernel function.

[0074] In some embodiments of the present invention, the bandwidth parameter of the macro kernel function can be expressed by the following formula:

[0075]

[0076] In the formula, is the total number of samples in the original collected sample data set, is any sample, Indicates samples.

[0077] The constructed micro-kernel function is expressed by the following formula:

[0078]

[0079] in, is a regularization parameter used to increase the numerical stability of the kernel function; is the micro kernel function.

[0080] In some embodiments of the invention, the regularization parameter It is expressed by the following formula:

[0081]

[0082] In the formula, is a very small positive number, is the first data sample Features, is the first data sample Features.

[0083] After constructing two sets of kernel functions, the indicator sample data are mapped to the feature spaces of the two sets of kernel functions, which are expressed as follows:

[0084]

[0085]

[0086] in, The feature space function defined for the macro kernel function, represents the set of all samples in the dataset, Eigenspace functions defined for microscopic kernel functions.

[0087] Then, the clustering algorithms of the two kernel spaces are respectively executed to obtain preliminary clustering results, and the clustering results of the two kernel spaces are fused. In some embodiments of the present invention, the fusion is performed by weighted average calculation, and the calculation process is represented by the following formula:

[0088]

[0089] In the formula, and are the macro and micro clustering results, respectively. is the fusion weight, This is the clustering fusion result.

[0090] Based on the above clustering results, the statistical characteristics of each cluster are calculated to perform clustering characteristic analysis. The calculation process is expressed by the following formula:

[0091]

[0092] but,

[0093] In the formula, is the number of samples belonging to this cluster, is the cluster center, is the cluster covariance.

[0094] S2. Based on the clustering characteristic analysis result, the sample data is expanded using the SMO algorithm to obtain an indicator expansion data set.

[0095] It can be understood that the core idea of ​​the SMO algorithm (Synthetic Minority Oversampling Technology / SMOTE algorithm) is to increase the number of minority samples by synthesizing new minority samples, thereby balancing the data set. Therefore, in this embodiment, the process of obtaining the indicator expansion data set is expressed by the following formula:

[0096]

[0097] In the formula, is a randomly generated interpolation factor, is the generated sample, Indicates samples.

[0098] In some embodiments of the present invention, the interpolation factor is calculated via relative distance and cluster density, and is expressed by the following formula:

[0099]

[0100] In the formula, Indicates samples, Indicates samples.

[0101] The indicator expansion data set and the original indicator data samples are mapped to the new feature space through nonlinear feature projection. The mapping process is expressed by the following formula:

[0102]

[0103] In the formula, For samples enhanced by nonlinear mapping, is the adjustment parameter of nonlinear projection, is the hyperbolic tangent function.

[0104] It can be understood that the nonlinear projection can enhance the class distinction of the sample, so that the expanded data can fully reflect the essential characteristics of the data. From the above formula, it can be seen that the adjustment parameters of the nonlinear projection are adaptively adjusted according to the nonlinear degree of the data. It is expressed by the following formula:

[0105]

[0106] In the formula, It is a sample The characteristic variance of .

[0107] It is understandable that when the number of expanded samples reaches a preset sample number threshold, the sample expansion is stopped. As an example, if the original collected samples are 800, and the data expansion obtains 200 samples, the expanded data set contains 1000 samples.

[0108] S3. In the second stage, the index expansion data set is input into the neural network model for training, the parameters of the model are optimized through the chaotic feedback mechanism and the regularization strategy to obtain a feature extraction model, and the first index feature data set is output based on the feature extraction model. It can be understood that the second stage is the feature extraction stage.

[0109] This step is the process of data feature extraction. For details, please refer to Figure 2 As shown, Figure 2 It is a schematic diagram of the feature extraction model training process in one of the embodiments of the present invention. During the feature extraction process, the embodiment of the present invention adopts a neural network model to extract features. In order to avoid the problems of gradient disappearance, gradient explosion or falling into a local optimal solution during the training process, the embodiment of the present invention adopts a chaotic feedback mechanism to adjust the model learning rate and optimizes the model parameters through a regularization strategy.

[0110] The following is a detailed description of the training process of the above feature extraction model:

[0111] Specifically, in some implementations, the weights and biases of the neural network model are first initialized, and the initialization process can be expressed by the following formula:

[0112]

[0113]

[0114] In the formula, is the initial value of the weight of the neural network; is the initial value of the bias of the neural network; is the mean shift of the neural network weights; is the mean shift of the weight bias of the neural network; is the standard deviation of the neural network weight initialization; represents a normal distribution with a mean of 0 and a standard deviation of the identity matrix, is the identity matrix. Preferably, Set to 0.01, Set to 0.05, Set to 0.001.

[0115] In the first training process, i.e., the training process of the feature extraction model, the learning rate of the neural network model is adjusted by the chaotic feedback mechanism. Specifically, the dynamic update process of the self-adjusting learning rate is expressed by the following formula:

[0116]

[0117] In the formula, For the The learning rate of the neural network for each iteration; For the The learning rate of the neural network for each iteration; It is the regulating factor of the neural network, controlling the sensitivity of changes in the learning rate; is the change in the error of the neural network in the previous cycle.

[0118] After the learning rate is updated, the weights are updated:

[0119]

[0120] In the formula, It is The weights of the neural network at the iteration; It is The weights of the neural network at the iteration; is the error of the neural network, which is calculated by the preset Softmax function on the output features of the last layer of the neural network; is the symbol of partial derivative.

[0121] The process of dynamically adjusting the number of hidden layer nodes of the neural network model based on the regularization strategy is expressed by the following formula:

[0122]

[0123] In the formula, It is The number of nodes in the hidden layer of the neural network at each iteration; It is The number of nodes in the hidden layer of the neural network at each iteration; is the node adjustment step size of the hidden layer of the neural network; is the sign function of the error change, which is used to decide whether to increase or decrease the number of nodes, that is, for ,when When greater than zero, is 1, when When it is less than or equal to zero, for .

[0124] In some embodiments of the present invention, when training a neural network model, the determination of the update step size of the node adjustment step size (such as weights and biases) of the hidden layer is related to the structure and function of the network hierarchy and can be calculated by the following formula:

[0125]

[0126] In the formula, For rounding operation; is the influence coefficient of adjusting the adjustment step size; is the depth of the current network layer. Preferably, The value is 6. Set to 2.5.

[0127] The first training process is iterated repeatedly until the parameter standard of the preset neural network model output is met, and the feature extraction model is obtained. Preferably, the preset maximum number of iterations is set to 1000 times. When the iteration reaches the target number of times, if the parameters (such as accuracy, loss value, etc.) output by the model meet the preset standards, it means that the model has been trained. In this step, a regularization strategy is used to optimize parameters to prevent overfitting, and regularization sparsification is performed by dynamically adjusting the number of nodes in the hidden layer of the neural network; and the learning rate of the model is adjusted through a chaotic feedback mechanism to avoid overfitting or underfitting, thereby ensuring the generalization ability of the model.

[0128] S4. In the third step, the first indicator feature data set is input into the autoencoder and parameter reconstruction and optimization are performed through dynamic routing decision to construct a feature dimension reduction model, and a second indicator feature data set is output based on the feature dimension reduction model. In this embodiment, the third step is the feature dimension reduction step.

[0129] This step is the process of reducing the feature dimension of the data. For details, please refer to Figure 3 As shown, Figure 3 It is a schematic diagram of the feature dimensionality reduction model training process in one of the embodiments of the present invention. During the feature dimensionality reduction process, the embodiment of the present invention adopts an encoder model to achieve feature dimensionality reduction. In order to avoid the problem of data loss during the training process, the embodiment of the present invention adopts dynamic routing decision to reconstruct the model data.

[0130] The following is a detailed description of the training process of the above feature dimensionality reduction model:

[0131] First, the weights and biases of the autoencoder are initialized, which can be expressed by the following formula:

[0132]

[0133]

[0134] In the formula, The first The initial weights of the layer; The first The initial bias of the layer; The first The number of neurons in the layer; The first The number of neurons in the layer; To generate the shape Standard normal distribution random numbers; To generate a length of The zero vector of .

[0135] It can be understood that the data input to the autoencoder will be transmitted through each layer of the encoder. Therefore, in the embodiment of the present invention, in the second training process, i.e., the training process of the feature dimensionality reduction model, after the first indicator feature data set is input to the autoencoder, a dynamic routing function is used to dynamically determine the importance of the output features of the activation function of each layer of the autoencoder, which is specifically represented by the following formula:

[0136]

[0137]

[0138]

[0139] In the formula, The first The linear transformation output of the layer; is the output feature of the activation function of the autoencoder; The first The weight of the layer; The first Bias of the layer; is the ReLU activation function; The first The input features of the layer; A dynamic routing function based on feature importance; Importance scoring function for calculating the output features of the activation function of the autoencoder.

[0140] It can be seen from the above formula that the embodiment of the present invention makes a dynamic routing decision for each layer of the autoencoder according to the output of the previous layer and the feature importance score of the current layer.

[0141] In some embodiments of the present invention, the dynamic routing function integrates the activation outputs of each feature by weighted averaging, where the weight is determined by the importance score of the feature and is calculated by the following formula:

[0142]

[0143] The calculation process of the importance of the output feature of the activation function of the autoencoder is expressed by the following formula:

[0144]

[0145] In the formula, It is Tier The activation output of a neuron, It is Tier The activation output of each neuron; is the weight of the corresponding feature importance, obtained through training, indicating the The importance of each feature; is the weight of the corresponding feature importance, obtained through training, indicating the The importance of a feature; It is The total number of neurons in the layer.

[0146] Since the autoencoder includes an encoder and a decoder, in the decoder stage, the output features are reconstructed using the loss function to obtain the corresponding reconstruction error

[0147] In this step, the reconstruction error is calculated by the following formula:

[0148]

[0149] In the formula, is the loss function of the autoencoder; The number of samples entered for the current batch; is the first input to the autoencoder samples; After reconstruction samples.

[0150] Based on the reconstruction error, the autoencoder parameters are updated and optimized through the back propagation algorithm, that is, the weights and biases of the autoencoder are updated, which is specifically calculated by the following formula:

[0151]

[0152]

[0153]

[0154]

[0155] In the formula, The first The update increment of the layer weights; The first The update increment of the layer's bias; is the learning rate of the autoencoder; is the updated autoencoder The weight of the layer; is the updated autoencoder Bias of the layer; The first Sparsity measure of layer neurons; is the sparsity pruning threshold; is an indicator function, which takes 1 when the condition in the brackets is met, otherwise it takes 0. Preferably, It is set to 0.8, indicating that weights will be pruned when the sparsity measure is above 80%.

[0156] It can be understood that in the above parameter update optimization process, the embodiment of the present invention improves the computational efficiency and generalization ability of the model by dynamically pruning redundant weights, that is, in most cases, the weights corresponding to neurons with smaller activation values ​​will be dynamically pruned, thereby simplifying the model structure. Specifically, the importance of neurons is judged by statistically analyzing the sparsity of activation values, and the sparsity measure of the i-th layer of neurons in the autoencoder is It is calculated by the following formula:

[0157]

[0158] In the formula, For the The sample in The activation value of the layer; is the sparsity threshold. Preferably, Can be set to 0.1.

[0159] The second training process is iterated repeatedly until the preset parameter standard of the autoencoder output is met, and the feature dimension reduction model is obtained (similar to the iterative process of the above-mentioned feature extraction model, which will not be repeated here). In this step, a dynamic routing mechanism based on feature importance score is used to determine the transmission path of each feature in different layers of the encoder, thereby enhancing the model's ability to reconstruct the original data and reducing information loss.

[0160] S5. In the fourth step, the second indicator feature data set is input into a high-order neural network model for training, and the model gradient update rate is dynamically adjusted to construct the power grid resilience assessment model. In this embodiment, the fourth step is the classifier model construction step.

[0161] This step is the training process of the evaluation model, please refer to Figure 4 As shown, Figure 4 It is a schematic diagram of the training and evaluation process of each model in one embodiment of the present invention. It can be seen that in the process of constructing the evaluation model, the embodiment of the present invention trains the classifier to achieve the final evaluation process. In order to avoid fitting problems during the training process, the embodiment of the present invention dynamically adjusts the high-order neural network gradient update rate to achieve parameter optimization.

[0162] The following is a detailed description of the training process of the above-mentioned power grid resilience assessment model:

[0163] The weights and biases of high-order neural networks are initialized using a method based on random distribution of quantum states, which is specifically expressed by the following formula:

[0164]

[0165]

[0166] In the formula, Represents the initialized weight matrix of a high-order neural network; Represents the number of neurons in the previous layer of a high-order neural network; Represents the phase angle randomly generated according to the quantum state, ensuring the randomness and diversity of weight initialization; Represents the bias term for initialization of the high-order neural network; represents a small positive number, which is used to prevent the activation function from entering the saturation region in the initial state. Preferably, Set to 0.001.

[0167] It is understandable that since the feature data after input dimension reduction is propagated in the neural network, the output of each layer is used as the input of the next layer. In particular, each neuron in the high-order neural network will also receive the gradient accumulation information of the neuron in the previous several iterations to adjust its activation state. Therefore, in the third training process, that is, the training process of the power grid resilience assessment model, the activation output of each layer of the high-order neural network model is calculated based on the historical gradient accumulation information of the high-order neural network model, which is specifically expressed by the following formula:

[0168]

[0169] In the formula, Represents a high-order neural network The activation output of the layer; Represents a high-order neural network The weight matrix of the layer; Represents a high-order neural network The input of the layer; Represents a high-order neural network Bias of the layer; is the Sigmoid activation function; Represents a high-order neural network Layer The gradient information accumulated at the iteration; Is the weight coefficient, which adjusts the influence of the cumulative gradient.

[0170] In some embodiments of the invention, the high-order neural network Layer The accumulated gradient information at the iteration It is expressed by the following formula:

[0171]

[0172] In the formula, The attenuation factor The power is used to adjust the contribution of historical gradients to the current update; Represents the loss function of a high-order neural network.

[0173] Based on the activation output, the error gradient and target gradient accumulation information of each layer of the high-order neural network model are calculated and iteratively adjusted. Specifically, the error gradient is expressed by the following formula:

[0174]

[0175] In the formula, Represents a high-order neural network Error gradient of the layer; Represents a high-order neural network The weight matrix of the layer; Represents the derivative of the Sigmoid activation function; Represents a high-order neural network The weighted sum of the layer's inputs.

[0176] The target gradient accumulation information is expressed by the following formula:

[0177]

[0178] In the formula, Represents a high-order neural network Gradient accumulation of layers; Represents the learning rate of the high-order neural network; represents the decay coefficient of gradient accumulation; Represents an element-wise dot product operation.

[0179] In some embodiments of the present invention, the learning rate of the high-order neural network is adjusted by dynamically adjusting the learning rate. , specifically, It is calculated by the following formula:

[0180]

[0181] In the formula, represents the learning rate decay coefficient, Indicates The learning rate of the neural network is given by the iteration, Indicates the current iteration number.

[0182] It can be seen from the above formula that it can be understood that in the iteration process, the embodiment of the present invention realizes batch update of weights through the process of dynamically adjusting the gradient information accumulated in the above training process and the learning rate. The update process is represented by the following formula:

[0183]

[0184] In the formula, represents the step size of weight update, Represents a high-order neural network The weight matrix of the layer, Update operation for parameters.

[0185] Preferably, in some implementations, the weights are updated by using an adaptive step size control method. It can be expressed by the following formula:

[0186]

[0187] In the formula, represents the step length attenuation coefficient, Indicates The step size of the weight update for each iteration.

[0188] When the number of iterations reaches the preset threshold, weight pruning and parameter optimization are performed. Specifically, the weight pruning process is expressed by the following formula:

[0189]

[0190] In the formula, represents the indicator function, Indicates the threshold for weight pruning.

[0191] Preferably, in some implementations, the threshold for weight pruning is calculated by the following formula:

[0192]

[0193] In the formula, represents the trimming coefficient; Represents the standard deviation of the weight matrix and is used to dynamically adjust the threshold of weight pruning to meet the needs of different training stages. Preferably, Set to 0.1.

[0194] The third training process is iterated repeatedly until the parameter standard of the preset high-order neural network model output is met, and the power grid resilience assessment model is obtained (similar to the iterative process of the above-mentioned feature extraction model, which will not be repeated here). In this step, the instability and overfitting problems that may be encountered in the training process of traditional high-order neural networks are solved by dynamically adjusting the gradient update rate mechanism, thereby improving the accuracy of the model evaluation results.

[0195] In summary, the present invention constructs a power grid resilience assessment model to accurately assess and analyze power grid resilience. The process of constructing the power grid resilience assessment model includes: training the model for data expansion, feature extraction and feature dimensionality reduction, and classifier classification, respectively, aiming to improve the accuracy of the assessment model by optimizing the model training process.

[0196] An embodiment of the present invention provides a power grid resilience assessment system, wherein the system is specifically used for:

[0197] Input the collected monitoring index data and line environment index data of the target area power grid into a pre-trained grid resilience assessment model for assessment and identification to obtain resilience assessment results;

[0198] Among them, the power grid resilience assessment model is designed to be trained through data expansion, feature extraction, feature dimension reduction and classifier classification during the construction process to obtain the trained power grid resilience assessment model.

[0199] For details, see Figure 5 , Figure 5 Shown is a structural block diagram of a power grid resilience assessment system in one embodiment of the present invention, wherein the system includes a data expansion module, a feature extraction model construction module, a feature dimension reduction model construction module and an assessment model construction module.

[0200] As an example, the system uses a data expansion module M1 to construct a macro-micro dual-core function in the first stage to cluster the power monitoring index data and the power environment index data as sample data to obtain preliminary clustering results, and perform clustering characteristic analysis on the preliminary clustering results to obtain clustering characteristic analysis results; based on the clustering characteristic analysis results, the sample data is expanded using the SMO algorithm to obtain an indicator expansion data set.

[0201] After constructing the kernel function, the indicator sample data is mapped to the feature space of the two sets of kernel functions, as shown below:

[0202]

[0203]

[0204] in, The feature space function defined for the macro kernel function, represents the set of all samples in the dataset, Eigenspace functions defined for microscopic kernel functions.

[0205] Then, the clustering algorithms of the two kernel spaces are respectively executed to obtain preliminary clustering results, and the clustering results of the two kernel spaces are fused. In some embodiments of the present invention, the fusion is performed by weighted average calculation, and the calculation process is represented by the following formula:

[0206]

[0207] In the formula, and are the macro and micro clustering results, respectively. is the fusion weight.

[0208] Based on the above clustering results, the statistical characteristics of each cluster are calculated to perform clustering characteristic analysis. The calculation process is expressed by the following formula:

[0209]

[0210]

[0211] In the formula, is the number of samples belonging to this cluster, is the cluster center, is the cluster covariance.

[0212] In this embodiment, the process of obtaining the indicator expansion data set is expressed by the following formula:

[0213]

[0214] In the formula, is a randomly generated interpolation factor, is the generated sample, Indicates samples.

[0215] In some embodiments of the present invention, the interpolation factor is calculated via relative distance and cluster density, and is expressed by the following formula:

[0216]

[0217] In the formula, Indicates samples, Indicates samples.

[0218] The indicator expansion data set and the original indicator data samples are mapped to the new feature space through nonlinear feature projection. The mapping process is expressed by the following formula:

[0219]

[0220] In the formula, For samples enhanced by nonlinear mapping, is the adjustment parameter of nonlinear projection, is the hyperbolic tangent function.

[0221] It is understandable that when the number of expanded samples reaches a preset sample number threshold, the sample expansion is stopped. As an example, if the original collected samples are 800, and the data expansion obtains 200 samples, the expanded data set contains 1000 samples.

[0222] Furthermore, a feature extraction model construction module M2 is used to input the indicator expansion data set into the neural network model for training in the second stage, optimize the model parameters through the chaotic feedback mechanism and regularization strategy to obtain a feature extraction model, and output the first indicator feature data set based on the feature extraction model.

[0223] As an example, this embodiment uses a feature extraction model construction module M2 to adjust the learning rate of the neural network model through the chaotic feedback mechanism during the first training process, and dynamically adjust the number of hidden layer nodes of the neural network model based on the regularization strategy.

[0224] The following is a detailed description of the training process of the above feature extraction model:

[0225] Specifically, in some implementations, the weights and biases of the neural network model are first initialized, and the initialization process can be expressed by the following formula:

[0226]

[0227]

[0228] In the formula, is the initial value of the weight of the neural network; is the initial value of the bias of the neural network; is the mean shift of the neural network weights; is the mean shift of the weight bias of the neural network; is the standard deviation of the neural network weight initialization; represents a normal distribution with a mean of 0 and a standard deviation of the identity matrix, is the identity matrix. Preferably, Set to 0.01, Set to 0.05, Set to 0.001.

[0229] In the first training process, i.e., the training process of the feature extraction model, the learning rate of the neural network model is adjusted by the chaotic feedback mechanism. Specifically, the dynamic update process of the self-adjusting learning rate is expressed by the following formula:

[0230]

[0231] In the formula, For the The learning rate of the neural network for each iteration; For the The learning rate of the neural network for each iteration; It is the regulating factor of the neural network, controlling the sensitivity of changes in the learning rate; is the change in the error of the neural network in the previous cycle.

[0232] After the learning rate is updated, the weights are updated:

[0233]

[0234] In the formula, It is The weights of the neural network at the iteration; It is The weights of the neural network at the iteration; is the error of the neural network, which is calculated by the preset Softmax function on the output features of the last layer of the neural network; is the symbol of partial derivative.

[0235] The process of dynamically adjusting the number of hidden layer nodes of the neural network model based on the regularization strategy is expressed by the following formula:

[0236]

[0237] In the formula, It is The number of nodes in the hidden layer of the neural network at each iteration; It is The number of nodes in the hidden layer of the neural network at each iteration; is the node adjustment step size of the hidden layer of the neural network; is the sign function of the error change, which is used to decide whether to increase or decrease the number of nodes, that is, for ,when When greater than zero, is 1, when When it is less than or equal to zero, for .

[0238] In some embodiments of the present invention, when training a neural network model, the determination of the update step size of the node adjustment step size (such as weights and biases) of the hidden layer is related to the structure and function of the network hierarchy and can be calculated by the following formula:

[0239]

[0240] In the formula, For rounding operation; is the influence coefficient of adjusting the adjustment step size; is the depth of the current network layer. Preferably, The value is 6. Set to 2.5.

[0241] The first training process is iterated repeatedly until the parameter standard of the preset neural network model output is met, and the feature extraction model is obtained. Preferably, the preset maximum number of iterations is set to 1000 times. When the iteration reaches the target number of times, if the parameters output by the model (such as accuracy, loss value, etc.) meet the preset standards, it means that the model has been trained.

[0242] Furthermore, through the feature dimensionality reduction model construction module M3, in the third stage, the first indicator feature data set is input into the autoencoder and the parameters are reconstructed and optimized through dynamic routing decisions to construct the feature dimensionality reduction model, and the second indicator feature data set is output based on the feature dimensionality reduction model.

[0243] As an example, this embodiment constructs a module M3 through a feature dimensionality reduction model. In the second training process, after the first indicator feature data set is input into the autoencoder, a dynamic routing decision is made on the importance of the output features of the activation function of each layer of the autoencoder through a dynamic routing function; the output features are reconstructed using a loss function to obtain a corresponding reconstruction error, and based on the reconstruction error, the autoencoder parameters are updated and optimized through a back propagation algorithm.

[0244] The following is a detailed description of the training process of the above feature dimensionality reduction model:

[0245] First, the weights and biases of the autoencoder are initialized, which can be expressed by the following formula:

[0246]

[0247]

[0248] In the formula, The first The initial weights of the layer; The first The initial bias of the layer; The first The number of neurons in the layer; The first The number of neurons in the layer; To generate the shape Standard normal distribution random numbers; To generate a length of The zero vector of .

[0249] It can be understood that the data input to the autoencoder will be transmitted through each layer of the encoder. Therefore, in the embodiment of the present invention, in the second training process, i.e., the training process of the feature dimensionality reduction model, after the first indicator feature data set is input to the autoencoder, a dynamic routing function is used to dynamically determine the importance of the output features of the activation function of each layer of the autoencoder, which is specifically represented by the following formula:

[0250]

[0251]

[0252]

[0253] In the formula, The first The linear transformation output of the layer; is the output feature of the activation function of the autoencoder; The first The weight of the layer; The first Bias of the layer; is the ReLU activation function; The first The input features of the layer; A dynamic routing function based on feature importance; Importance scoring function for calculating the output features of the activation function of the autoencoder.

[0254] It can be seen from the above formula that the embodiment of the present invention makes a dynamic routing decision for each layer of the autoencoder according to the output of the previous layer and the feature importance score of the current layer.

[0255] In some embodiments of the present invention, the dynamic routing function integrates the activation outputs of each feature by weighted averaging, where the weight is determined by the importance score of the feature and is calculated by the following formula:

[0256]

[0257] The calculation process of the importance of the output feature of the activation function of the autoencoder is expressed by the following formula:

[0258]

[0259] In the formula, It is Tier The activation output of a neuron, It is Tier The activation output of each neuron; is the weight of the corresponding feature importance, obtained through training, indicating the The importance of each feature; is the weight of the corresponding feature importance, obtained through training, indicating the The importance of a feature; It is The total number of neurons in the layer.

[0260] Since the autoencoder includes an encoder and a decoder, in the decoder stage, the output features are reconstructed using the loss function to obtain the corresponding reconstruction error

[0261] In this step, the reconstruction error is calculated by the following formula:

[0262]

[0263] In the formula, is the loss function of the autoencoder; The number of samples entered for the current batch; is the first input to the autoencoder samples; After reconstruction samples.

[0264] Based on the reconstruction error, the autoencoder parameters are updated and optimized through the back propagation algorithm, that is, the weights and biases of the autoencoder are updated, which is specifically calculated by the following formula:

[0265]

[0266]

[0267]

[0268]

[0269] In the formula, The first The update increment of the layer weights; The first The update increment of the layer's bias; is the learning rate of the autoencoder; is the updated autoencoder The weight of the layer; is the updated autoencoder Bias of the layer; The first Sparsity measure of layer neurons; is the sparsity pruning threshold; is an indicator function, which takes 1 when the condition in the brackets is met, otherwise it takes 0. Preferably, It is set to 0.8, indicating that weights will be pruned when the sparsity measure is above 80%.

[0270] It can be understood that in the above parameter update optimization process, the embodiment of the present invention improves the computational efficiency and generalization ability of the model by dynamically pruning redundant weights, that is, in most cases, the weights corresponding to neurons with smaller activation values ​​will be dynamically pruned, thereby simplifying the model structure. Specifically, the importance of neurons is judged by statistically analyzing the sparsity of activation values, and the sparsity measure of the i-th layer of neurons in the autoencoder is It is calculated by the following formula:

[0271]

[0272] In the formula, For the The sample in The activation value of the layer; is the sparsity threshold. Preferably, Can be set to 0.1.

[0273] The second training process is iterated repeatedly until the parameter standard of the preset autoencoder output is met, thereby obtaining the feature dimensionality reduction model (which is similar to the iterative process of the above-mentioned feature extraction model and will not be repeated here).

[0274] Furthermore, through the evaluation model construction module M4, in the fourth stage, the second indicator feature data set is input into the high-order neural network model for training, and the model gradient update rate is dynamically adjusted to construct the power grid resilience evaluation model.

[0275] Exemplarily, by evaluating the model building module M4, in the third training process, the activation output of each layer of the high-order neural network model is calculated according to the historical gradient accumulation information of the high-order neural network model; based on the activation output, the error gradient and target gradient accumulation information of each layer of the high-order neural network model are calculated and iteratively adjusted; when the number of iterations reaches a preset threshold, weight pruning and parameter optimization are performed.

[0276] Specifically, the training process of the above-mentioned power grid resilience assessment model is described in detail:

[0277] The weights and biases of high-order neural networks are initialized using a method based on random distribution of quantum states, which is specifically expressed by the following formula:

[0278]

[0279]

[0280] In the formula, Represents the initialized weight matrix of a high-order neural network; Represents the number of neurons in the previous layer of a high-order neural network; Represents the phase angle randomly generated according to the quantum state, ensuring the randomness and diversity of weight initialization; Represents the bias term for initialization of the high-order neural network; represents a small positive number, which is used to prevent the activation function from entering the saturation region in the initial state. Preferably, Set to 0.001.

[0281] It is understandable that since the feature data after input dimension reduction is propagated in the neural network, the output of each layer is used as the input of the next layer. In particular, each neuron in the high-order neural network will also receive the gradient accumulation information of the neuron in the previous several iterations to adjust its activation state. Therefore, in the third training process, that is, the training process of the power grid resilience assessment model, the activation output of each layer of the high-order neural network model is calculated based on the historical gradient accumulation information of the high-order neural network model, which is specifically expressed by the following formula:

[0282]

[0283] In the formula, Represents a high-order neural network The activation output of the layer; Represents a high-order neural network The weight matrix of the layer; Represents a high-order neural network The input of the layer; Represents a high-order neural network Bias of the layer; is the Sigmoid activation function; Represents a high-order neural network Layer The gradient information accumulated at the iteration; Is the weight coefficient, which adjusts the influence of the cumulative gradient.

[0284] In some embodiments of the invention, the high-order neural network Layer The accumulated gradient information at the iteration It is expressed by the following formula:

[0285]

[0286] In the formula, The attenuation factor The power is used to adjust the contribution of historical gradients to the current update; Represents the loss function of a high-order neural network.

[0287] Based on the activation output, the error gradient and target gradient accumulation information of each layer of the high-order neural network model are calculated and iteratively adjusted. Specifically, the error gradient is expressed by the following formula:

[0288]

[0289] In the formula, Represents a high-order neural network Error gradient of the layer; Represents a high-order neural network The weight matrix of the layer; Represents the derivative of the Sigmoid activation function; Represents a high-order neural network The weighted sum of the layer's inputs.

[0290] The target gradient accumulation information is expressed by the following formula:

[0291]

[0292] In the formula, Represents a high-order neural network Gradient accumulation of layers; Represents the learning rate of the high-order neural network; represents the decay coefficient of gradient accumulation; Represents an element-wise dot product operation.

[0293] In some embodiments of the present invention, the learning rate of the high-order neural network is adjusted by dynamically adjusting the learning rate. , specifically, It is calculated by the following formula:

[0294]

[0295] In the formula, represents the learning rate decay coefficient, Indicates The learning rate of the neural network is given by the iteration, Indicates the current iteration number.

[0296] It can be seen from the above formula that it can be understood that in the iteration process, the embodiment of the present invention realizes batch update of weights through the process of dynamically adjusting the gradient information accumulated in the above training process and the learning rate. The update process is represented by the following formula:

[0297]

[0298] In the formula, represents the step size of weight update, Represents a high-order neural network The weight matrix of the layer, Update operation for parameters.

[0299] Preferably, in some implementations, the weights are updated by using an adaptive step size control method. It can be expressed by the following formula:

[0300]

[0301] In the formula, represents the step length attenuation coefficient, Indicates The step size of the weight update for each iteration.

[0302] When the number of iterations reaches the preset threshold, weight pruning and parameter optimization are performed. Specifically, the weight pruning process is expressed by the following formula:

[0303]

[0304] In the formula, represents the indicator function, Indicates the threshold for weight pruning.

[0305] Preferably, in some implementations, the threshold for weight pruning is calculated by the following formula:

[0306]

[0307] In the formula, represents the trimming coefficient; Represents the standard deviation of the weight matrix and is used to dynamically adjust the threshold of weight pruning to meet the needs of different training stages. Preferably, Set to 0.1.

[0308] The third training process is iterated repeatedly until the parameter standard of the preset high-order neural network model output is met, thereby obtaining the power grid resilience assessment model (which is similar to the iterative process of the above-mentioned feature extraction model and will not be repeated here).

[0309] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for evaluating power grid resilience, characterized in that: include: Input the collected power monitoring index data and line environment index data of the target area power grid into the pre-trained power grid resilience assessment model for assessment and identification to obtain the resilience assessment results; The power grid resilience assessment model is designed to be trained by data expansion, feature extraction, feature dimension reduction and classifier classification in the construction process to obtain the trained power grid resilience assessment model; The construction process of the power grid resilience assessment model includes: In the first stage, a macro-micro dual-kernel function is constructed to cluster the power monitoring index data and the line environment index data as sample data to obtain a preliminary clustering result, and a clustering characteristic analysis is performed on the preliminary clustering result to obtain a clustering characteristic analysis result; Based on the clustering characteristic analysis result, the sample data is expanded using the SMO algorithm to obtain an indicator expansion data set; In the second stage, the index expansion data set is input into the neural network model for training, the parameters of the model are optimized through the chaotic feedback mechanism and the regularization strategy to obtain a feature extraction model, and the first index feature data set is output based on the feature extraction model; In the third step, the first indicator feature data set is input into the autoencoder and parameter reconstruction and optimization are performed through dynamic routing decision to construct a feature dimension reduction model, and a second indicator feature data set is output based on the feature dimension reduction model; In the fourth stage, the second indicator feature data set is input into a high-order neural network model for training, and the model gradient update rate is dynamically adjusted to construct the power grid resilience assessment model.

2. The power grid resilience assessment method according to claim 1, characterized in that: The training process of the feature extraction model includes: In the first training process, the learning rate of the neural network model is adjusted by the chaotic feedback mechanism, and the number of hidden layer nodes of the neural network model is dynamically adjusted based on the regularization strategy; The first training process is iterated repeatedly until the parameter standard of the preset neural network model output is met, thereby obtaining the feature extraction model.

3. The power grid resilience assessment method according to claim 1, characterized in that: The training process of the feature dimensionality reduction model includes: In the second training process, after the first indicator feature data set is input into the autoencoder, a dynamic routing decision is made on the importance of the output features of the activation function of each layer of the autoencoder through a dynamic routing function; Reconstructing the output features using the loss function to obtain corresponding reconstruction errors, and updating and optimizing the autoencoder parameters using a back propagation algorithm based on the reconstruction errors; The second training process is iterated repeatedly until the parameter standard of the preset autoencoder output is met, thereby obtaining the feature dimension reduction model.

4. The power grid resilience assessment method according to claim 1, characterized in that: The training process of the power grid resilience assessment model includes: In the third training process, the activation output of each layer of the high-order neural network model is calculated based on the historical gradient accumulation information of the high-order neural network model; Based on the activation output, calculating and iteratively adjusting the error gradient and target gradient accumulation information of each layer of the high-order neural network model; When the number of iterations reaches the preset threshold, weight pruning and parameter optimization are performed; The third training process is iterated repeatedly until the parameter standard of the preset high-order neural network model output is met, thereby obtaining the power grid resilience assessment model.

5. A power grid resilience assessment system, characterized in that: The system is specifically used for: Input the collected power monitoring index data and line environment index data of the target area power grid into the pre-trained power grid resilience assessment model for assessment and identification to obtain the resilience assessment results; The power grid resilience assessment model is designed to be trained in the data expansion, feature extraction, feature dimension reduction and classifier classification links during the construction process to obtain the trained power grid resilience assessment model; the system includes: a data expansion module, a feature extraction model construction module, a feature dimension reduction model construction module and an assessment model construction module; wherein, The data expansion module is used to construct a macro-micro dual-core function in the first stage to cluster the power monitoring index data and the line environment index data as sample data to obtain a preliminary clustering result, perform clustering characteristic analysis on the preliminary clustering result, and obtain a clustering characteristic analysis result; and Based on the clustering characteristic analysis result, the sample data is expanded using the SMO algorithm to obtain an indicator expansion data set; The feature extraction model building module is used to input the indicator expansion data set into the neural network model for training in the second stage, optimize the parameters of the model through the chaotic feedback mechanism and regularization strategy to obtain the feature extraction model, and output the first indicator feature data set based on the feature extraction model; The feature dimension reduction model construction module is used to input the first indicator feature data set into the autoencoder and perform parameter reconstruction optimization through dynamic routing decision to construct a feature dimension reduction model in the third stage, and output a second indicator feature data set based on the feature dimension reduction model; The evaluation model construction module is used to input the second indicator feature data set into a high-order neural network model for training in the fourth stage, and dynamically adjust the model gradient update rate to construct the power grid resilience evaluation model.

6. The power grid resilience assessment system according to claim 5, characterized in that: The feature extraction model construction module is specifically used for: In the first training process, the learning rate of the neural network model is adjusted by the chaotic feedback mechanism, and the number of hidden layer nodes of the neural network model is dynamically adjusted based on the regularization strategy; The first training process is iterated repeatedly until the parameter standard of the preset neural network model output is met, thereby obtaining the feature extraction model.

7. The power grid resilience assessment system according to claim 5, characterized in that: The feature dimension reduction model construction module is specifically used for: In the second training process, after the first indicator feature data set is input into the autoencoder, a dynamic routing decision is made on the importance of the output features of the activation function of each layer of the autoencoder through a dynamic routing function; Reconstructing the output features using the loss function to obtain corresponding reconstruction errors, and updating and optimizing the autoencoder parameters using a back propagation algorithm based on the reconstruction errors; The second training process is iterated repeatedly until the parameter standard of the preset autoencoder output is met, thereby obtaining the feature dimension reduction model.

8. The power grid resilience assessment system according to claim 5, characterized in that: The evaluation model building module is specifically used for: In the third training process, the activation output of each layer of the high-order neural network model is calculated based on the historical gradient accumulation information of the high-order neural network model; Based on the activation output, calculating and iteratively adjusting the error gradient and target gradient accumulation information of each layer of the high-order neural network model; When the number of iterations reaches the preset threshold, weight pruning and parameter optimization are performed; The third training process is iterated repeatedly until the parameter standard of the preset high-order neural network model output is met, thereby obtaining the power grid resilience assessment model.

Images