Power distribution network internal overvoltage identification method based on deep residual network

Abstract

The application discloses a power distribution network internal overvoltage identification method based on a deep residual network, and the method is as follows: obtaining various different types of temporary overvoltage and operating overvoltage fault samples through PSCAD simulation, pre-processing and grouping the sample data to establish training set samples and test set samples; constructing a residual network model: through the establishment of a plurality of convolution layers, pooling layers and residual blocks, a deep residual network model is formed; after network initialization setting, the training set is input into the deep residual network model, and multi-layer convolution and pooling operations are used to realize abstract representation of original signals; the fault data is input into the model, abstract representation and network training are carried out on the fault data, a Softmax classifier is integrated at the end of the network model, and classification and identification of two types of internal overvoltage samples are realized. The application realizes automatic acquisition and identification of fault features, reduces fault diagnosis workload, and improves power distribution network internal overvoltage monitoring efficiency.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, and specifically relates to a method for identifying overvoltages inside a power distribution network based on a deep residual network. Background Technology

[0002] Operational experience shows that overvoltage is one of the important factors affecting the safe operation of distribution networks. Prolonged internal overvoltages can easily cause insulation damage to equipment, leading to various short-circuit faults and jeopardizing the reliability of power supply. Therefore, timely detection and classification of internal overvoltages in distribution networks, and accurate differentiation of overvoltage types, are of great significance for disaster prevention and fault analysis in distribution networks.

[0003] The commonly used overvoltage identification method both domestically and internationally is to first extract feature quantities and then combine them with classification algorithms such as support vector machines, extreme learning machines, and radial basis function neural networks to obtain the identification result. However, due to the inherent structural characteristics of CNNs, overfitting often occurs when the network depth reaches a certain level, causing the network performance to decline instead of improve.

[0004] Therefore, a new method for identifying internal overvoltages in power distribution networks is needed to improve the efficiency of internal overvoltage monitoring. Summary of the Invention

[0005] In view of the technical problems existing in the background technology, the overvoltage identification method for distribution networks based on deep residual networks provided by the present invention constructs a deep residual network model for the overvoltage classification problem in distribution networks, so as to realize the automatic acquisition and identification of fault features, replacing the two work steps of signal preprocessing and feature set construction in fault diagnosis research, thereby reducing the workload of fault diagnosis and improving the efficiency of overvoltage monitoring in distribution networks.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A method for identifying internal overvoltages in a distribution network based on a deep residual network, comprising the following steps:

[0008] Step 1: Obtain fault samples of various types of temporary overvoltage and switching overvoltage through PSCAD simulation, preprocess the sample data and group them to establish training set samples and test set samples.

[0009] Step 2, Constructing the residual network model: By establishing multiple layers of convolutional layers, pooling layers, and residual blocks, a deep residual network model is formed.

[0010] Step 3: After initializing the network, input the training set into the deep residual network model and use multi-layer convolution and pooling operations to achieve abstract representation of the original signal.

[0011] Step 4: Input the fault data into the model, abstract and represent it, and train the network. Integrate the Softmax classifier at the end of the network model to achieve classification and identification of the two types of internal overvoltage samples.

[0012] Preferably, in step 2, the residual network model includes 5 convolutional layers, 6 pooling layers, and 3 residual blocks; each residual block's main path has two convolutional layers used to extract deep bearing features; simultaneously, there is an identity mapping path used to directly transmit signals from the upper network layer to the lower network layer; the data flows through the residual blocks into the Dropout layer, and finally the extracted features are sent to the Softmax classifier for fault mode classification. The Dropout layer can significantly reduce overfitting and decrease the interaction between feature detectors.

[0013] Preferably, in step 2, the convolutional layer extracts features from the image. As the image passes through the convolutional and pooling layers, the features gradually transform from low-level features to high-level features. These features are then classified through a fully connected layer and an output layer. (Convolutional layer feature map:)

[0014]

[0015] Where: l represents the current layer; x l j x represents the j-th feature map of the current layer; l-1 j 'j' represents the i-th feature map in layer l-1; f(·) represents the activation function; '*' represents the convolution operation; w l ij Represents the convolution kernel; k j 'b' represents the receptive field of the input layer; 'b' is the output bias term.

[0016] Preferably, in step 2, the pooling layer is used to process features with excessively high dimensionality output by the convolutional layer, reducing the resolution of the input image, reducing parameters, improving efficiency, reducing data volume, and preventing overfitting. The max pooling function is:

[0017]

[0018] Where: down(·) represents pooling operation; x l j θ is the j-th feature map on layer l; θ corresponds to the feature map of each input; b is the bias term.

[0019] Preferably, the connection layer is used to achieve dimensionality reduction of the features, converting the two-dimensional signal into a one-dimensional signal, for further feature representation. The fully connected layer outputs:

[0020] y i =f(ω) j x+bj (3)

[0021] Where: y j The output of the fully connected layer is represented by ω; x represents the input feature vector; ω represents the output of the fully connected layer. j This represents the j-th weighted row vector.

[0022] Preferably, the learning process of a multi-layer convolutional neural network is as follows:

[0023]

[0024] Where: z is the medium deceleration reserve, h is the nonlinear activation function, b is the bias, and w is the weight value.

[0025] The residual network directly converts x t The input signal is copied into a deeper layer of the network, where a non-linear activation function is directly used, thus forming a residual block. The input signal is directly passed to the output signal, and the network only needs to learn the differentiated parts, accelerating the training speed. The residual network learning process is as follows:

[0026]

[0027] Preferably, the activation function in the network is the ReLU function, which makes the predicted y values ​​all greater than or equal to 0; the cross-entropy loss function is selected to calculate the difference between the learning model distribution and the training distribution, so that it converges to a better local minimum to improve the network recognition accuracy; the gradient descent adopts the Adam optimization algorithm, which adds first-order momentum and second-order momentum to avoid getting trapped in local optima.

[0028] This patent can achieve the following beneficial effects:

[0029] This invention provides a method for identifying internal overvoltages in distribution networks based on deep residual networks. This method can be used in image recognition systems and distribution network operating systems in the electrical field. Specifically, electronic devices use image recognition technology to collect various internal overvoltage data acquired by the devices and input them into a deep residual network. The deep residual network, which has been trained extensively, is then used to classify the internal overvoltages of each distribution network.

[0030] This invention transforms the identification of overvoltages within a distribution network into an identification method based on deep residual networks. Specifically, it utilizes a model whose parameters have been trained on a large dataset from ImageNet through transfer learning to identify unknown internal overvoltages, thus determining the type of internal overvoltage. Furthermore, this invention achieves autonomous data mining and abstract feature extraction of the raw overvoltage signals from the distribution network by constructing a deep learning network model containing 5 convolutional layers, 6 pooling layers, and 3 residual modules. With the help of fully connected layers and a classifier, it classifies two types of internal overvoltages: temporary overvoltages and operational overvoltages, thereby determining the type of overvoltage within the distribution network to be identified and effectively improving the accuracy of overvoltage classification. Therefore, the deep residual network-based method for identifying internal overvoltages in distribution networks provided by this invention can be used in disaster prevention and fault analysis systems for distribution networks. Attached Figure Description

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0032] Figure 1 This is a flowchart illustrating the identification process of the present invention. Detailed Implementation

[0033] Preferred solutions include Figure 1 As shown, a method for identifying overvoltages within a distribution network based on a deep residual network includes the following steps:

[0034] Step 1: Obtain fault samples of various types of temporary overvoltage and switching overvoltage through PSCAD simulation, preprocess the sample data and group them to establish training set samples and test set samples.

[0035] Step 2, Constructing the residual network model: By establishing multiple layers of convolutional layers, pooling layers, and residual blocks, a deep residual network model is formed.

[0036] Step 3: After initializing the network, input the training set into the deep residual network model and use multi-layer convolution and pooling operations to achieve abstract representation of the original signal.

[0037] Step 4: Input the fault data into the model, abstract and represent it, and train the network. Integrate the Softmax classifier at the end of the network model to achieve classification and identification of the two types of internal overvoltage samples.

[0038] Furthermore, the residual network model includes 5 convolutional layers, 6 pooling layers, and 3 residual blocks; each residual block has two convolutional layers in its main path to extract deep bearing features; at the same time, there is an identity mapping path to directly transmit the upper network signal to the lower network; the data flows through the residual block into the Dropout layer, and finally the extracted features are sent to the Softmax classifier for fault mode classification.

[0039] Furthermore, convolutional layers extract features from the image. As the image passes through convolutional and pooling layers, its features gradually transform from low-level to high-level features. These features are then classified through fully connected layers and the output layer. (Convolutional layer feature map:)

[0040]

[0041] Where: l represents the current layer; x l j x represents the j-th feature map of the current layer; l-1 j 'j' represents the i-th feature map in layer l-1; f(·) represents the activation function; '*' represents the convolution operation; w l ij Represents the convolution kernel; k j 'b' represents the receptive field of the input layer; 'b' is the output bias term.

[0042] Furthermore, pooling layers are used to handle features with excessively high dimensionality in the output of convolutional layers, reducing the resolution of the input image, decreasing parameters, improving efficiency, reducing data volume, and preventing overfitting. The max pooling function is:

[0043]

[0044] Where: down(·) represents pooling operation; x l j θ is the j-th feature map on layer l; θ corresponds to the feature map of each input; b is the bias term.

[0045] Furthermore, the connection layer is used to reduce the dimensionality of the features, converting the two-dimensional signal into a one-dimensional signal, and then performing further feature representation. The fully connected layer outputs:

[0046] y i =f(ω) j x+b j (7)

[0047] Where: y j The output of the fully connected layer is represented by ω; x represents the input feature vector; ω represents the output of the fully connected layer. j This represents the j-th weighted row vector.

[0048] The learning process of a multi-layer convolutional neural network is as follows:

[0049]

[0050] Where: z is the medium deceleration reserve, h is the nonlinear activation function, b is the bias, and w is the weight value.

[0051] The residual network directly converts x t The input signal is copied into a deeper layer of the network, where a non-linear activation function is directly used, thus forming a residual block. The input signal is directly passed to the output signal, and the network only needs to learn the differentiated parts, accelerating the training speed. The residual network learning process is as follows:

[0052]

[0053] The ReLU activation function is used in the network, ensuring that the predicted y values ​​are all greater than or equal to 0. This effectively adds sparsity to the network, reduces the dependence between parameters, prevents overfitting, and improves training speed. The cross-entropy loss function is used to calculate the difference between the learning model distribution and the training distribution, allowing it to converge to a better local minimum to improve the network's recognition accuracy. The gradient descent uses the Adam optimization algorithm, which adds first-order and second-order momentum to avoid getting trapped in local optima.

[0054] Example 1:

[0055] The main application scenario of this invention is: by collecting the overvoltage data extracted from the distribution network by the detection equipment and inputting it into a deep residual network, a simple and efficient identification of the types of overvoltages inside the distribution network can be achieved.

[0056] Step 1:

[0057] A neutral-point ungrounded distribution network model was constructed using PSCAD / EMTDC simulation software. The main components include a 110kV infinite bus power supply G; a 10kV / 10.5kV main transformer T1, and a 10kV / 0.4kV distribution transformer T2; the loads are replaced by inductive loads R+jX=(20+j40)Ω. There are three types of feeders: fully overhead lines (OL), fully cable lines (CL), and hybrid line-cable lines; a time-controlled switch K is used to simulate grounding faults, and R... f This is a transition resistor. The bus PT is an electromagnetic voltage transformer.

[0058] Operational overvoltage: The arc grounding overvoltage multiple obtained from the power frequency arc extinction theory is close to the arc grounding overvoltage multiple that occurs in the distribution network. This theory assumes that the arc of the fault phase will be extinguished and reignited once every power frequency cycle, and assumes that the arc is extinguished when the power frequency current crosses zero and reignited when the maximum recovery voltage is reached. Therefore, based on this theory, the arc grounding overvoltage is simulated by periodically switching the time control switch K1.

[0059] Temporary overvoltage: The disappearance of a single-phase ground fault is used as the "excitation" condition for ferromagnetic resonance. To achieve the parameter matching requirements of ferromagnetic resonance, different matching capacitors are formed by switching feeders of different lengths, thereby changing the three-phase admittance of the resonant circuit to generate fundamental frequency resonance, sub-frequency resonance, and high-frequency resonance. Simulations of sub-frequency resonance are performed with feeders L3 and L6 connected; simulations of fundamental frequency resonance are performed with feeders L5 and L6 connected; and simulations of high-frequency resonance are performed with feeder L6 connected.

[0060] Step Two:

[0061] Each fault category is labeled using one-hot encoding. One-hot encoding extends the values ​​of discrete features to Euclidean space, making each value of a discrete feature correspond to a point in Euclidean space, thus making the distance calculation between features more reasonable. This classification and labeling of the data facilitates subsequent qualitative and quantitative analysis of the signals.

[0062] Step 3:

[0063] To accelerate training, when inputting sample data into the network model, a data generator is used to read the samples in segments, with each batch serving as a training unit. Backpropagation is then performed, and the network model calculates the loss of each training sample in the batch and averages it. This requires shuffling the sample data in the dataset. The shuffled dataset is then used as the training and validation sets in a 7:3 ratio and input into the deep residual network model. The training set is used to extract features of overvoltages within the distribution network and determine model parameters, while the validation set is used to test the model's ability to identify and classify overvoltages, finding the optimal solution in a parameter space and adjusting the model configuration. This study uses a hold-out validation method for model evaluation, i.e., creating a validation set.

[0064] Step Four:

[0065] The study performed mixed identification of all internal overvoltage samples, including two representative types: resonant overvoltage and arc grounding overvoltage. Each label represents the state type of each sample. The data was input into the network, achieving an accuracy of 98.3% on the training set and 97.8% on the validation set. The increased data volume provided richer training data, thus exceeding the performance of purely qualitative and quantitative experiments in the final identification. The deep residual network demonstrates effective qualitative and quantitative identification capabilities for internal overvoltages in distribution networks, with an overall validation set accuracy exceeding 97%, exhibiting high reliability.

[0066] Ultimately, the overvoltage identification method model for distribution networks based on deep residual networks proposed in this invention achieved a maximum identification accuracy of 97.25%. This invention, described in the field of image recognition, is a method for identifying overvoltages within distribution networks based on deep residual networks, with broad application prospects and suitable for widespread promotion.

[0067] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for identifying internal overvoltages in a distribution network based on a deep residual network, characterized in that: Step 1: Obtain fault samples of various types of temporary overvoltage and switching overvoltage through PSCAD simulation, preprocess the sample data and group them to establish training set samples and test set samples. Step 2, Constructing the residual network model: By establishing multiple layers of convolutional layers, pooling layers, and residual blocks, a deep residual network model is formed. Step 3: After initializing the network, input the training set into the deep residual network model and use multi-layer convolution and pooling operations to achieve abstract representation of the original signal. Step 4: Input the fault data into the model, abstract and represent it and train the network. Integrate the Softmax classifier at the end of the network model to achieve classification and identification of the two types of internal overvoltage samples. In step 2, the residual network model includes 5 convolutional layers, 6 pooling layers, and 3 residual blocks; each residual block has two convolutional layers in its main path to extract deep features of the bearing; at the same time, there is an identity mapping path to directly transmit the upper network signal to the lower network; the data flows through the residual block into the Dropout layer. In step 2, convolutional layers extract features from the image. As the image passes through convolutional and pooling layers, its features gradually transform from low-level to high-level features. These features are then classified through fully connected layers and the output layer. (Convolutional layer feature map:) (1)； in: l Represents the current layer; x l j Represents the current layer's first j Each feature map; x l-1 j represent l The i-th feature map in layer -1; f (∙) '*' represents the activation function; '*' represents the convolution operation. w l ij Represents the convolution kernel; k j Represents the receptive field of the input layer; b It is the output bias term; In step 2, the pooling layer is used to process features with excessively high dimensionality in the output of the convolutional layer, reducing the resolution of the input image, decreasing parameters, improving efficiency, reducing data volume, and preventing overfitting. The max pooling function is: (2)； in: down(∙) Represents pooling operations; x l j yes l The first layer j Each feature map; θ Each corresponds to a feature map of its input. b For bias terms; The connection layer is used to reduce the dimensionality of the features, converting the two-dimensional signal into a one-dimensional signal, for further feature representation. The fully connected layer outputs: (3)； in: y j This represents the output of the fully connected layer; x This represents the input feature vector; ω j Indicates the first j Each weighted row vector; The learning process of a multi-layer convolutional neural network is as follows: (4)； in: z For medium-speed deceleration reserves, h It is a non-linear activation function. b For bias, w These are weight values; Residual networks directly x t The input signal is copied into a deeper layer of the network, where a non-linear activation function is directly used, thus forming a residual block. The input signal is directly passed to the output signal, and the network only needs to learn the differentiated parts, accelerating the training speed. The residual network learning process is as follows: （5）。 2. The method for identifying internal overvoltages in a distribution network based on a deep residual network according to claim 1, characterized in that: The activation function used in the network is the ReLU function, which ensures that the predicted y values ​​are all greater than or equal to 0. The cross-entropy loss function is selected to calculate the difference between the learning model distribution and the training distribution, so that it converges to a better local minimum to improve the network's recognition accuracy. The gradient descent uses the Adam optimization algorithm, which adds first-order momentum and second-order momentum to avoid getting trapped in local optima.

Images