Power grid accident early warning and processing method based on multi-factor coupling
By combining deep convolutional self-coding network and deep noise reduction self-coding network, low-dimensional characteristics of power grid accidents are extracted and standardized process assisted decisions are made, which solves the limitations of power grid accident early warning methods in the existing technology, and improves early warning accuracy and Processing efficiency ensures the safety and stability of the power grid.
Patent Information
- Application Number
- CN202411040748.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-26
AI Technical Summary
The existing grid accident warning methods cannot comprehensively and effectively predict complex and changeable grid accidents, and insufficient data utilization and processing have affected the timeliness and accuracy of early warnings. The existing systems respond to speed and decision-making in emergencies and emergencies. Insufficient support capabilities.
Using a combination of deep convolutional self-coding network and deep noise reduction self-coding network, low-dimensional abstract features are extracted by collecting power grid equipment operating status, load data and operation data, and auxiliary decision-making is made based on standardized processes during the accident handling process to verify faults. Security of disposal policies.
It improves the accuracy and reliability of grid accident warning, enhances the efficiency and safety guarantee capabilities of accident handling, and ensures the safety and stability of grid operation.
Smart Images

Figure CN119027270A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of power systems and automation thereof, and in particular to a power grid accident early warning and processing method based on multi-factor coupling. Background Art
[0002] Existing power grid accident early warning methods usually only consider a single factor or a simple multi-factor linear relationship, and cannot comprehensively and effectively predict power grid accidents. These methods often show limitations when dealing with complex and changeable power grid accidents, and it is difficult to provide sufficient accuracy and reliability. Most existing methods rely on traditional physical models, statistical models or simple machine learning algorithms. Although these methods can provide certain early warning capabilities under specific conditions, they are limited in effectiveness when faced with complex power grid operating environments and changeable external influencing factors. Equipment failures may be exacerbated by operational errors, while load fluctuations and changes in power demand may be affected by the external environment and market dynamics.
[0003] At the same time, the existing early warning system also has deficiencies in data utilization and processing. It cannot make full use of the real-time data in the operation of the power grid, lacks the ability to deeply mine and analyze the data, and the timeliness and accuracy of the early warning are affected. Especially in emergencies and emergency situations, the response speed and decision support capabilities of the existing system are obviously insufficient, and it is difficult to provide effective early warnings and processing suggestions in a timely manner. Chinese patent CN116910633A discloses a power grid fault prediction method based on multimodal knowledge hybrid reasoning, which uses a knowledge reasoning algorithm based on distributed representation to infer the type of power grid fault in the current environment and obtain the prediction result. Although the power grid fault prediction method considers multimodal knowledge and hybrid reasoning, it still has some deficiencies. First, the method has a high processing complexity for multimodal data, which may lead to insufficient real-time operation of the system. Secondly, in practical applications, the acquisition and processing of data may be subject to various restrictions, especially data that needs to be updated in real time. Finally, the method may have different processing and identification effects on different fault types, and some fault types may be difficult to accurately predict, thereby affecting the overall early warning effect. Chinese patent CN113553420A discloses a method and system for recommending power grid fault handling procedures based on knowledge graphs, which stores fault-related knowledge in a topological form in a graph database, and searches the graph database according to the fault text to avoid accidents caused by human error, effectively shorten the accident handling time, improve the dispatching personnel's ability to handle power grid accidents, and improve the intelligent level of power grid accident handling. Although this power grid fault prediction method uses knowledge graph technology to improve the intelligent level of fault handling, it still has some shortcomings. The processing complexity of multimodal data is high, resulting in insufficient real-time operation of the system. In practical applications, data acquisition and processing may be subject to various restrictions, especially data that needs to be updated in real time, which increases the difficulty of system maintenance. This method relies on a pre-trained model, and the recognition and processing effects are not ideal for new types of faults or unseen fault modes, affecting the accuracy and timeliness of the overall warning. Summary of the invention
[0004] Purpose of the invention: The purpose of the invention is to provide a power grid accident early warning and processing method based on multi-factor coupling.
[0005] Technical solution: The present invention comprises the following steps:
[0006] (1) Collect the operating status, load data and operation data of power grid equipment;
[0007] (2) Use deep convolutional autoencoder networks to extract low-dimensional abstract features of power grid accidents;
[0008] (3) Establish a deep noise reduction autoencoding network for power grid accident warning and processing;
[0009] (4) Assist decision-making based on standardized processes during accident handling;
[0010] (5) Verify the safety of the fault handling strategy through the power grid flow calculation model.
[0011] Furthermore, the step (2) of adding a deep convolutional autoencoder network to perform feature extraction includes:
[0012] (2.1) Network construction: Construct a deep convolutional autoencoder network, including an encoder and a decoder, where the encoder converts the input into a low-dimensional feature vector and the decoder maps the feature vector back to the input;
[0013] (2.2) Hyperparameter setting: Perform preliminary setting of network hyperparameters, fix the values of hyperparameters that do not need to be optimized, and set the objective function and the range of hyperparameters to be optimized for the White Whale optimization algorithm;
[0014] (2.3) Network training: Use the White Whale optimization algorithm to train the network and find the optimal hyperparameters. In each round of iteration, use the training set to perform the back propagation algorithm to update the network parameters, use the validation set to evaluate the performance of the model and adjust the optimal hyperparameters, and save the model weights and biases after training is completed;
[0015] (2.4) Feature extraction: The trained deep convolutional autoencoder network is used to extract features from the dataset, and the features are visualized using t-SNE technology to intuitively demonstrate the feature extraction effect of the method used in this chapter.
[0016] Furthermore, in step (2), except for the network input layer, each layer receives the output of the previous layer as the input of the layer, and adopts a layer-by-layer training method during training, that is, the input data is compressed into a low-dimensional representation layer by layer, and the low-dimensional representation is restored to the original data dimension through a decoder. In the process of layer-by-layer training, each layer of the network will learn the features of different levels of the input data.
[0017] Furthermore, in step (2), L2 regularization is used in the deep convolutional autoencoder network. When the loss function reaches a sufficiently small value, it is considered that the features in the original input data are already included in the output of the encoding part:
[0018]
[0019] The first part is the mean square error loss function, and the second part is the L2 regularization term, where λ represents the weight decay parameter, which is used to adjust the relative importance of the two parts in the formula; nl represents the number of hidden layers in the network, and s l represents the number of neurons in layer l, is the weight coefficient between the ith neuron in the lth layer and the jth neuron in the l+1th layer; the role of the L2 regularization term is to control and adjust the weight size between each layer.
[0020] Furthermore, the step (3) uses a deep noise reduction autoencoder network to perform unsupervised training and supervised fine-tuning to perform power grid accident warning, including:
[0021] (3.1) Feature splicing: The features extracted by the deep convolutional autoencoder network for the three types of disaster factors are spliced in the same dimension, and the 60*1-dimensional feature vector is used as the input of the deep denoising autoencoder network;
[0022] (3.2) Network construction: Construct the structure of the deep denoising autoencoder network, which consists of two parts: the encoding part and the decoding part;
[0023] (3.3) Hyperparameter setting: fix the hyperparameters of the network that are not involved in the optimization, including the slope of LeakyReLu, the mean and standard deviation of the network's randomly initialized weights, and the relevant parameters of the adam optimizer that are not involved in the optimization;
[0024] (3.4) Set the objective function and the hyperparameter range to be optimized for the White Whale optimization algorithm. The four hyperparameters to be optimized are: the learning rate of training, the number of batch samples and the number of training rounds, and the weight decay coefficient of the Adam optimizer;
[0025] (3.5) Network training: Use the Beluga optimization algorithm to train the network and automatically optimize the hyperparameters. Use the training set to adjust the network weights and biases. Use the validation set to evaluate the model performance and use the Beluga optimization algorithm to adjust the hyperparameters. Save the model weights and biases after all training rounds are completed. Each round of training consists of two parts: unsupervised training and supervised fine-tuning.
[0026] Furthermore, the unsupervised training in step (3.5) includes:
[0027] (3.5.1) Randomly initialize the weights and biases of the deep convolutional autoencoder network;
[0028] (3.5.2) The input data is passed to the convolutional autoencoder of the first layer. The autoencoder of this layer consists of an encoder and a decoder. The encoder is responsible for compressing the input data into a set of feature vectors, while the decoder converts the feature vectors back to input data. The loss function of the first layer of autoencoder is calculated by the gradient descent method. When the loss function is less than the set value, the training of the network of this layer can be stopped, and the weights of the network of this layer can be saved.
[0029] (3.5.3) The output of the first layer of convolutional autoencoder after training is used as the input of the second layer of convolutional autoencoder. The second layer network is trained using the gradient descent method until the loss function of this layer is less than the set value, and the weight parameters of the second layer network are obtained.
[0030] (3.5.4) Repeat the above steps and train the network layer by layer until all convolutional autoencoders are trained and the parameters of each layer of the network are obtained. The encoder output of the last layer of convolutional autoencoder is used as the feature vector of the disaster factor to complete the feature extraction.
[0031] Furthermore, in the step (3.2) of constructing the structure of the deep denoising autoencoder network, the loss function is improved by adding a sparsity penalty term to the loss function to suppress the activation of the hidden layer neurons.
[0032] Furthermore, adding a sparsity penalty term to the loss function includes:
[0033] The output of a neuron is defined as the activation of the neuron, that is, for m given inputs, the average activation of the yth neuron in the hidden layer is:
[0034]
[0035] For the Sigmoid activation function, a neuron is considered activated when its output is close to 1 after the Sigmoid function, and is considered inhibited when its output is close to 0. Calculate the average activation of the neurons in the hidden layer ρ j , assuming that the sparsity parameter is wide ρ j = 0.05, if the average activation of hidden layer neuron j is ρ j The closer it is to ρ, the closer the neuron is to being unactivated. The sparse penalty term uses the KL divergence to measure the difference between the average activation of the neuron and the sparsity parameter:
[0036]
[0037] The loss function after adding the sparse penalty term is:
[0038]
[0039] Among them, β is the weight of controlling the sparsity penalty factor.
[0040] Furthermore, the KL divergence is used to measure the j The KL divergence is an indicator of the relative entropy between two Bernoulli random variables with the same mean. As the difference between ρ and ρ increases, the value of KL divergence also increases. When the difference between the values of ρ and ρ is smaller, the KL divergence will also be smaller. If ρ j =p then KL(ρ j||ρ)=0.
[0041] Furthermore, the step (4) of assisting decision making based on the standardized process includes:
[0042] (4.1) Acquisition of raw data and data normalization preprocessing. Collect power grid equipment data, load data, and operation data to associate and match the three types of data with power grid accidents, and construct a power grid accident data set. In order to solve the problem of many types of factors and large dimensional differences, the raw data is normalized, and the values of each type of disaster-causing factor are scaled to the interval [0, 1];
[0043] (4.2) Feature extraction based on deep convolutional autoencoder networks: Three deep convolutional autoencoder networks are established for the three types of disaster factors, and unsupervised training is used to extract features from the expanded data set;
[0044] (4.3) Constructing a power grid accident prediction model, the feature vectors extracted by the three feature extraction sub-networks are concatenated, and the concatenated feature vectors are used as the input of the deep denoising autoencoder network to construct a power grid accident prediction model based on the deep denoising autoencoder network. Through unsupervised training and supervised fine-tuning, a comprehensive mapping relationship between feature vectors and power grid accident categories is established;
[0045] (4.4) Complete network model training. After training and optimization, select typical accident cases in the test set for verification to achieve power grid accident prediction and early warning.
[0046] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: through the coupling of multiple factors, the power grid equipment, load and operation data are comprehensively considered, the accuracy and reliability of power grid accident warning are improved, and the limitations of single factor analysis in the prior art are overcome; in the process of accident handling, guidance is provided based on standardized processes to prevent dispatching behaviors such as missed verification of information, wrong decisions, and wrong power transmission, effectively improving the standardization and safety of dispatching operations; combined with the output of power plants in the vicinity of the fault, equipment maintenance conditions, on-site abnormal information reporting, etc., an optimization plan for fault isolation and power restoration is provided to support the decision-making of dispatchers and enhance the efficiency and accuracy of fault handling; finally, through power grid flow calculation and N-1 flow calculation, it is verified whether the fault handling strategy causes safety hazards to the power grid, thereby ensuring the safety and stability of power grid operation; in actual applications, it shows higher warning accuracy, processing efficiency and security assurance capabilities. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0048] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.
[0049] like Figure 1 As shown, the present invention comprises the following steps:
[0050] (1) Collect the operating status, load data and operation data of power grid equipment;
[0051] (2) Use deep convolutional autoencoder networks to extract low-dimensional abstract features of power grid accidents;
[0052] The encoding part of the convolutional autoencoder is mainly composed of convolutional layers, which extract data features by performing convolution operations on the input data. Each convolutional layer consists of a series of fixed-size convolutional kernels, where parameters such as the size, step size, and number of convolutional kernels determine the type and number of features extracted by the convolutional layer. The size of the convolutional kernel is usually set to a small value, which allows the convolutional kernel to focus on the local information of the input data and extract higher-level features by learning these local features. The convolution operation is achieved by sliding the convolutional kernel over the entire input data. When performing a convolution operation, the elements in each convolutional kernel are multiplied with the elements in the corresponding position in the input data, and then all the product results are summed. Finally, the output result of the convolutional layer is generated.
[0053]
[0054] In the formula, is the feature vector corresponding to the I-th convolution kernel of the n-th convolution layer, is the bias parameter corresponding to the i-th convolution kernel of the n-th layer, K l is the convolution kernel, is the jth weight coefficient of the i-th convolution kernel of the n-th layer, and f represents the activation function.
[0055] By encoding the input data once, a feature vector that can reflect the characteristics of the original signal is generated. In order to reconstruct the input signal, we need to decode the feature vector. The decoding part of the convolutional autoencoder is mainly composed of a transposed convolution layer. The essence of the transposed convolution operation is a forward convolution process, which is achieved by expanding the dimension of the data. The output result of the transposed convolution layer can be obtained.
[0056]
[0057] in is the feature corresponding to the i-th convolution kernel of the n-layer transposed convolution layer, For transposed convolution calculation, is the jth weight parameter of the i-th transposed convolution kernel of the n-th layer, f is the bias term corresponding to the i-th transposed convolution kernel of the n-th layer, and f represents the nonlinear activation function.
[0058] A Deep Convolutional Auto-Encoder (DCAE) network is added to extract the features and reduce the dimension of power grid accidents. The network autonomously learns and retains the most critical factors, thereby reducing the structural complexity of the next-level deep denoising auto-encoder network and improving the accuracy of prediction.
[0059] The process of DCAE feature extraction is as follows:
[0060] 1) Network construction: Construct a deep convolutional autoencoder network, including an encoder and a decoder, where the encoder converts the input into a low-dimensional feature vector and the decoder maps the feature vector back to the input;
[0061] 2) Hyperparameter setting: Perform preliminary setting of network hyperparameters, fix the values of hyperparameters that do not need to be optimized, and set the objective function and the range of hyperparameters to be optimized for the White Whale optimization algorithm;
[0062] 3) Network training: Use the White Whale optimization algorithm to train the network and find the optimal hyperparameters. In each round of iteration, use the training set to perform the back propagation algorithm to update the network parameters, use the validation set to evaluate the performance of the model and adjust the optimal hyperparameters, and save the model weights and biases after training is completed;
[0063] 4) Feature extraction: Use the trained deep convolutional autoencoder network to extract features from the dataset, and visualize the features using t-SNE technology to intuitively demonstrate the feature extraction effect of the method used in this chapter.
[0064] Except for the network input layer, each layer receives the output of the previous layer as the input of this layer. During training, the layer-by-layer training method is adopted, that is, the input data is compressed into a low-dimensional representation layer by layer, and the low-dimensional representation is restored to the original data dimension through the decoder. In the process of layer-by-layer training, each layer of the network will learn the features of different levels of the input data.
[0065] The specific training steps for unsupervised learning are as follows:
[0066] 1) Randomly initialize the weights and biases of the deep convolutional autoencoder network;
[0067] 2) Pass the input data to the convolutional autoencoder of the first layer, which contains an encoder and a decoder. The encoder is responsible for compressing the input data into a set of feature vectors, while the decoder converts the feature vectors back into input data. The loss function of the first layer of autoencoder is calculated by the gradient descent method. When the loss function is less than the set value, the training of the network of this layer can be stopped, and the weights of the network of this layer can be saved.
[0068] 3) The output of the encoder after training the first layer of convolutional autoencoder is used as the input of the second layer of convolutional autoencoder, and the second layer network is trained using the gradient descent method until the loss function of this layer is less than the set value, thereby obtaining the weight parameters of the second layer network.
[0069] 4) Repeat steps 2) and 3) to train the network layer by layer until all convolutional autoencoders are trained and the parameters of each layer of the network are obtained. The encoder output of the last layer of convolutional autoencoder is used as the feature vector of the disaster factor to complete the feature extraction.
[0070] For network training, the back propagation algorithm is used to update the model parameters, with the goal of minimizing the loss function between the input data and the decoded output data. To prevent overfitting, the commonly used regularization methods include Dropout regularization, L1 regularization, and L1 regularization. This paper uses L2 regularization in the deep convolutional autoencoder network, which is the loss function after adding the regularization term. When the loss function reaches a sufficiently small value, it can be considered that the features in the original input data are already included in the output of the encoded part.
[0071]
[0072] In the formula, the first part is the mean square error loss function, the second part is the L2 regularization term, and λ represents the weight decay parameter, which is used to adjust the relative importance of the two parts in the formula. l Indicates the number of hidden layers in the network, s l represents the number of neurons in the first layer, is the weight coefficient between the i-th neuron in the l-th layer and the i-th neuron in the 1+1 layer. The function of the L2 regularization term is to control and adjust the weights between the layers. At the same time, L2 regularization can better improve the generalization ability of the network model and prevent overfitting.
[0073] (3) Establish a deep noise reduction autoencoding network for power grid accident warning and processing;
[0074] A deep denoising autoencoder network (SDAE) is used for unsupervised training and supervised fine-tuning to predict possible accidents in the power grid.
[0075] The process of SDAE's power grid accident warning is as follows:
[0076] 1) Feature splicing: The features extracted by the deep convolutional autoencoder network for the three types of disaster-causing factors are spliced in the same dimension, and the 60*1-dimensional feature vector is used as the input of the deep denoising autoencoder network;
[0077] 2) Network construction: Construct the structure of the deep denoising autoencoder network, which consists of two parts: the encoding part and the decoding part;
[0078] 3) Hyperparameter setting: fix the hyperparameters of the network that are not involved in the optimization, including the slope of LeakyReLu, the mean and standard deviation of the network's randomly initialized weights, and the relevant parameters of the adam optimizer that are not involved in the optimization;
[0079] 4) Set the objective function and the range of hyperparameters to be optimized for the White Whale optimization algorithm. The four hyperparameters to be optimized are: the learning rate of training, the number of batch samples and the number of training rounds, and the weight decay coefficient of the Adam optimizer;
[0080] 5) Network training: Use the White Whale optimization algorithm to train the network and automatically optimize hyperparameters. Use the training set to adjust the network weights and biases, use the validation set to evaluate the model performance, and use the White Whale optimization algorithm to adjust the hyperparameters. After all training rounds are completed, save the model weights and biases. Each round of training consists of two parts: unsupervised training and supervised fine-tuning.
[0081] 6) Model testing: Using actual grid accident data as input, case verification is performed, and the predicted accident categories are compared with the actual accident categories to evaluate the prediction accuracy of the model. At the same time, it is compared with other models to verify the effectiveness of the method used in this paper.
[0082] In the deep denoising autoencoder network, in order to improve the sparsity of the model and extract more meaningful feature representations, the loss function is improved. By adding a sparsity penalty term to the loss function, the activation of hidden layer neurons can be suppressed, thereby achieving the sparsity constraint of the model. Sparsity constraint refers to limiting the output of the hidden layer during the training process of the deep denoising autoencoder network to make it more sparse, that is, more neurons take values close to 0. This constraint enables the network to learn more meaningful and abstract feature representations. This method is used to adjust the number of activations of hidden layer neurons so that most neurons remain in an inactive state. The output of a neuron is defined as the activation of the neuron, that is, for m given inputs, the average activation of the yth neuron in the hidden layer is:
[0083]
[0084] For the Sigmoid activation function, a neuron is considered activated when its output is close to 1 after the Sigmoid function, and is considered inhibited when its output is close to 0. Calculate the average activation of the hidden layer neurons ρ j , assuming that the sparsity parameter is wide ρ j = 0.05, if the average activation of hidden layer neuron j is ρ jThe closer it is to ρ, the closer the neuron is to being inactive. In practical applications, the hidden layer neurons are usually suppressed by adding a sparse penalty term to the network's loss function. In this paper, the sparse penalty term uses the KL (Kullback-Leibler) divergence to measure the difference between the average activation of the neuron and the sparsity parameter. The KL divergence is used to measure the difference between ρ and ρ. j The KL divergence is an indicator of the relative entropy between two Bernoulli random variables with the same mean. As the difference between ρ increases, the value of KL divergence also gradually increases. When the difference between the values of User and ρ is smaller, the KL divergence will also be smaller. If ρ j =ρ then KL(ρ j ||ρ)=0, in this way the hidden layer neurons are suppressed.
[0085]
[0086] The loss function after adding the sparse penalty term is:
[0087]
[0088] Where β is the weight of the sparsity penalty factor.
[0089] The power grid accident features extracted by the deep convolutional autoencoder network are spliced, and the spliced features are used as the input of the deep denoising autoencoder network for training. The specific process of supervised fine-tuning training is as follows:
[0090] 1) After the unsupervised training of the network is completed, the network weights of the network coding part are saved. Then, these weights are used as the initial weights of the network in the supervised fine-tuning stage, and a forward propagation operation is performed on the data set with accident category labels to obtain the activation values of each hidden layer and the probability of each power grid accident category output by the softmax classifier.
[0091] 2) Calculate the network residual based on the predicted accident category and the actual accident category:
[0092]
[0093] 3) For the i-th layer (i=2, 3, 4...n-1), the network residual is:
[0094] δ (i) =((W (i) ) T δ (i+1) )·f′(h (i) )
[0095] 4) The partial derivatives of each layer are
[0096]
[0097] 5) Using the above iterative formula, the parameters of the entire deep denoising autoencoder network are iteratively updated through the back propagation algorithm. (n) , is the activation value of the nth hidden layer, f′(h (n) ) is the derivative of the hidden layer activation function, I is the actual category label of the input sample; P is the conditional probability matrix; (W (i) ) T is the transpose of the numerical matrix of the i-th layer network, h (i) is the activation value of the i-th hidden layer.
[0098] (4) Assist decision-making based on standardized processes during accident handling;
[0099] The multi-factor coupled power grid accident early warning method is adopted. First, a conditional deep convolutional generative adversarial network with data generation capability is used to expand the data set and balance the data. Then, a deep convolutional autoencoder network and a deep noise reduction autoencoder network with data feature extraction and power grid accident classification capabilities are established to establish a mapping relationship between disaster factors and power grid accidents. The basic steps of power grid accident early warning are as follows:
[0100] 1) Acquisition of raw data and data normalization preprocessing. Collect power grid equipment data, load data, and operation data to associate and match the three types of data with power grid accidents, and construct a power grid accident data set. In order to solve the problem of many types of factors and large dimensional differences, the raw data is normalized, and the values of each type of disaster-causing factor are scaled to the interval [0,1].
[0101] 2) Feature extraction based on deep convolutional autoencoder network. Three deep convolutional autoencoder networks are established for the three types of disaster factors, and unsupervised training is used to extract features from the expanded data set;
[0102] 3) Construct a power grid accident prediction model. The feature vectors extracted by the three feature extraction sub-networks are concatenated and used as the input of the deep denoising autoencoder network to construct a power grid accident prediction model based on the deep denoising autoencoder network. Through unsupervised training and supervised fine-tuning, a comprehensive mapping relationship between feature vectors and power grid accident categories is established.
[0103] 4) Complete network model training. After training and optimization, select typical accident cases in the test set for verification to achieve power grid accident prediction and early warning.
[0104] (5) Verify the safety of the fault handling strategy through the power grid flow calculation model.
[0105] The present invention collects the operating status, load and operation data of power grid equipment in real time through monitoring equipment. Use deep convolutional autoencoder network to extract low-dimensional abstract features, and use deep noise reduction autoencoder network to predict and handle accidents. Combined with on-site data, it provides fault isolation and power supply restoration solutions. The accident handling assistant automatically pushes the operating status and key indicators to realize the early warning function. Perform power grid flow and N-1 calculations to verify the safety of the fault handling strategy.
Claims
1. A power grid accident early warning and processing method based on multi-factor coupling, characterized in that: The steps include: (1) Collect the operating status, load data and operation data of power grid equipment; (2) Use deep convolutional autoencoder networks to extract low-dimensional abstract features of power grid accidents; (3) Establish a deep noise reduction autoencoding network for power grid accident warning and processing; (4) Assist decision-making based on standardized processes during accident handling; (5) Verify the safety of the fault handling strategy through the power grid flow calculation model.
2. The power grid accident early warning and processing method based on multi-factor coupling according to claim 1 is characterized in that: The step (2) of adding a deep convolutional autoencoder network to perform feature extraction includes: (2.1) Network construction: Construct a deep convolutional autoencoder network, including an encoder and a decoder, where the encoder converts the input into a low-dimensional feature vector and the decoder maps the feature vector back to the input; (2.2) Hyperparameter setting: Perform preliminary setting of network hyperparameters, fix the values of hyperparameters that do not need to be optimized, and set the objective function and the range of hyperparameters to be optimized for the White Whale optimization algorithm; (2.3) Network training: Use the White Whale optimization algorithm to train the network and find the optimal hyperparameters. In each round of iteration, use the training set to perform the back propagation algorithm to update the network parameters, use the validation set to evaluate the performance of the model and adjust the optimal hyperparameters, and save the model weights and biases after training is completed; (2.4) Feature extraction: The trained deep convolutional autoencoder network is used to extract features from the dataset, and the features are visualized using t-SNE technology to intuitively demonstrate the feature extraction effect of the method used in this chapter.
3. The power grid accident early warning and processing method based on multi-factor coupling according to claim 1 is characterized in that: In step (2), except for the network input layer, each layer receives the output of the previous layer as the input of the layer, and adopts a layer-by-layer training method during training, that is, the input data is compressed into a low-dimensional representation layer by layer, and the low-dimensional representation is restored to the original data dimension through a decoder. In the process of layer-by-layer training, each layer of the network will learn the features of different levels of the input data.
4. The power grid accident early warning and processing method based on multi-factor coupling according to claim 1 is characterized in that: In step (2), L2 regularization is used in the deep convolutional autoencoder network. When the loss function reaches a sufficiently small value, it is considered that the features in the original input data are already included in the output of the encoding part: Among them, the first part is the mean square error loss function, and the second part is the L2 regularization term, where λ represents the weight decay parameter, which is used to adjust the relative importance of the two parts in the formula; n l Indicates the number of hidden layers in the network, s l represents the number of neurons in the first layer, is the weight coefficient between the i-th neuron in the 1st layer and the j-th neuron in the l+1th layer; the role of the L2 regularization term is to control and adjust the weight size between each layer.
5. The power grid accident early warning and processing method based on multi-factor coupling according to claim 1 is characterized in that: The step (3) uses a deep noise reduction autoencoder network to perform unsupervised training and supervised fine-tuning to perform power grid accident warning, including: (3.1) Feature splicing: The features extracted by the deep convolutional autoencoder network for the three types of disaster factors are spliced in the same dimension, and the 60*1-dimensional feature vector is used as the input of the deep denoising autoencoder network; (3.2) Network construction: Construct the structure of the deep denoising autoencoder network, which consists of two parts: the encoding part and the decoding part; (3.3) Hyperparameter setting: fix the hyperparameters of the network that are not involved in the optimization, including the slope of LeakyReLu, the mean and standard deviation of the network's randomly initialized weights, and the relevant parameters of the adam optimizer that are not involved in the optimization; (3.4) Set the objective function and the hyperparameter range to be optimized for the White Whale optimization algorithm. The four hyperparameters to be optimized are: the learning rate of training, the number of batch samples and the number of training rounds, and the weight decay coefficient of the Adam optimizer; (3.5) Network training: Use the Beluga optimization algorithm to train the network and automatically optimize the hyperparameters. Use the training set to adjust the network weights and biases. Use the validation set to evaluate the model performance and use the Beluga optimization algorithm to adjust the hyperparameters. Save the model weights and biases after all training rounds are completed. Each round of training consists of two parts: unsupervised training and supervised fine-tuning.
6. The power grid accident early warning and processing method based on multi-factor coupling according to claim 5 is characterized in that: The step (3.5) of unsupervised training includes: (3.5.1) Randomly initialize the weights and biases of the deep convolutional autoencoder network; (3.5.2) The input data is passed to the convolutional autoencoder of the first layer. The autoencoder of this layer consists of an encoder and a decoder. The encoder is responsible for compressing the input data into a set of feature vectors, while the decoder converts the feature vectors back to input data. The loss function of the first layer of autoencoder is calculated by the gradient descent method. When the loss function is less than the set value, the training of the network of this layer can be stopped, and the weights of the network of this layer can be saved. (3.5.3) The output of the first layer of convolutional autoencoder after training is used as the input of the second layer of convolutional autoencoder. The second layer network is trained using the gradient descent method until the loss function of this layer is less than the set value, and the weight parameters of the second layer network are obtained. (3.5.4) Repeat the above steps and train the network layer by layer until all convolutional autoencoders are trained and the parameters of each layer of the network are obtained. The encoder output of the last layer of convolutional autoencoder is used as the feature vector of the disaster factor to complete the feature extraction.
7. The power grid accident early warning and processing method based on multi-factor coupling according to claim 5 is characterized in that: In the step (3.2), the loss function is improved in the structure of the deep denoising autoencoder network, and the activation of hidden layer neurons is suppressed by adding a sparsity penalty term to the loss function.
8. The power grid accident early warning and processing method based on multi-factor coupling according to claim 7 is characterized in that: The addition of a sparsity penalty term to the loss function includes: The output of a neuron is defined as the activation of the neuron, that is, for m given inputs, the average activation of the yth neuron in the hidden layer is: For the Sigmoid activation function, a neuron is considered activated when its output is close to 1 after the Sigmoid function, and is considered inhibited when its output is close to 0. Calculate the average activation of the neurons in the hidden layer ρ j , assuming that the sparsity parameter is wide ρ j = 0.05, if the average activation of hidden layer neuron j is ρ j The closer it is to ρ, the closer the neuron is to being unactivated. The sparse penalty term uses the KL divergence to measure the difference between the average activation of the neuron and the sparsity parameter: The loss function after adding the sparse penalty term is: Among them, β is the weight of controlling the sparsity penalty factor.
9. The power grid accident early warning and processing method based on multi-factor coupling according to claim 7 is characterized in that: The KL divergence is used to measure the j The KL divergence is an indicator of the relative entropy between two Bernoulli random variables with the same mean. As the difference between ρ and ρ increases, the value of KL divergence also increases. When the difference between the values of ρ and ρ is smaller, the KL divergence will also be smaller. If ρ j =ρ then KL(ρ j ||ρ)=0.
10. The power grid accident early warning and processing method based on multi-factor coupling according to claim 1 is characterized in that: The step (4) of assisting decision making based on a standardized process includes: (4.1) Acquisition of raw data and data normalization preprocessing. Collect power grid equipment data, load data, and operation data to associate and match the three types of data with power grid accidents, and construct a power grid accident data set. In order to solve the problem of many types of factors and large dimensional differences, the raw data is normalized, and the values of each type of disaster-causing factor are scaled to the interval [0, 1]; (4.2) Feature extraction based on deep convolutional autoencoder networks: Three deep convolutional autoencoder networks are established for the three types of disaster factors, and unsupervised training is used to extract features from the expanded data set; (4.3) Constructing a power grid accident prediction model, the feature vectors extracted by the three feature extraction sub-networks are concatenated, and the concatenated feature vectors are used as the input of the deep denoising autoencoder network to construct a power grid accident prediction model based on the deep denoising autoencoder network. Through unsupervised training and supervised fine-tuning, a comprehensive mapping relationship between feature vectors and power grid accident categories is established; (4.4) Complete network model training. After training and optimization, select typical accident cases in the test set for verification to achieve power grid accident prediction and early warning.
