Residual-convolutional network deep learning-based photovoltaic power supply and communication anomaly identification method
Through the method based on residual-convolution network deep learning, the problem of distribution network abnormal identification in distributed photovoltaic power generation systems is solved, and abnormal identification with high accuracy and anti-noise performance is achieved, which improves the stability and operating efficiency of the power system.
Patent Information
- Application Number
- CN202411976985.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-31
- Publication Date
- 2025-05-06
AI Technical Summary
In distributed photovoltaic power generation systems, photovoltaic equipment accounts for a high proportion and insufficient monitoring, which makes it difficult to identify fine abnormalities of distribution networks, and traditional monitoring methods cannot meet the current demand for refined management.
Using photovoltaic power supply and communication exception recognition method based on residual-convolution network deep learning, an efficient identification model is constructed through data preprocessing and neural network training, and accurately identify data anomalies and distinguish anomalies.
It achieves an identification accuracy of up to 95% in common power system settings, has noise anti-noise performance, supports real-time monitoring and rapid abnormality detection, and improves the stability and operation efficiency of the distribution network.
Smart Images
Figure CN119939305A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of power system analysis, and in particular to a photovoltaic power supply and communication anomaly identification method based on residual-convolutional network deep learning. Background Art
[0002] Driven by the global energy structure transformation and sustainable development strategy, distributed photovoltaic power generation has developed rapidly and become an important force in the energy revolution. However, the increase in the number of distributed power sources has brought challenges to the operation and control of new distribution networks, and traditional strategies are difficult to adapt. Especially when the proportion of photovoltaic equipment is high and monitoring is insufficient, the refined abnormal identification of distribution networks has become an urgent problem to be solved.
[0003] The stable operation of the distribution network is crucial to the safety of the entire power grid. However, due to measurement errors, communication failures and equipment problems, monitoring data anomalies occur frequently, which not only affects the formulation of control strategies, but also brings new challenges to the operation and management of the distribution network. With the continuous development of distribution network technology and the large-scale access of distributed power sources, traditional monitoring methods can no longer meet the current needs of refined management. Therefore, improving monitoring accuracy and identifying data anomalies in a timely and accurate manner are of great significance to the effective management of the distribution network and the stable operation of the power system.
[0004] In response to the above situation, we proposed a photovoltaic power supply and communication anomaly identification method based on residual-convolutional network deep learning on the basis of existing technologies. Summary of the invention
[0005] The purpose of this application is to provide a photovoltaic power supply and communication anomaly identification method based on residual-convolutional network deep learning.
[0006] The photovoltaic power supply and communication anomaly identification method based on residual-convolutional network deep learning provided in this application adopts the following technical solutions:
[0007] Anomaly identification methods include:
[0008] Data preprocessing, to reduce the noise interference of the original collected data, extract the characteristic data that can reflect the abnormal type, and obtain a data set with known data abnormality type;
[0009] Neural network training first establishes the network model parameter initialization, then performs forward propagation to calculate the error, then performs back propagation to calculate the gradient and update the model parameters, and then trains until the iteration termination condition ends;
[0010] Neural network testing, the test set calculates the model anomaly detection accuracy.
[0011] Furthermore, the data preprocessing includes the generation and preprocessing of a data set. First, the voltage, current, and power information of multiple nodes are collected and labeled, and the data are divided into a training data set and a verification data set in proportion and then saved separately.
[0012] Furthermore, the data set includes signal characteristic information and data label information, and the signal characteristic information and data label information are processed into single-precision floating point data types;
[0013] The Standard Scaler is used to normalize the data and scale each dimension of feature data to a standard normal distribution with a mean of 0 and a variance of 1, so that different features have the same scale and range to eliminate the differences in feature dimensions of different dimensions.
[0014] Furthermore, the two-dimensional vectors in the data set are converted into Pytorch tensors, and their types are converted into floating-point types to represent continuous numerical features.
[0015] Furthermore, in the neural network training, the number of training rounds and learning rate parameters of the neural network are first set, the training set data is forward propagated to calculate the error loss value, and then the training error value is recalculated according to the back propagation training gradient and the weight parameters of each layer of the neural network are updated, and the training is terminated until the iteration termination condition ends.
[0016] Furthermore, the error loss value uses the cross entropy loss function as follows:
[0017]
[0018] Among them, L is the loss value, N is the number of samples, and y i is the true label of the i-th sample, is the predicted label of the i-th sample.
[0019] Furthermore, in the neural network test, the test set data is input into the trained neural network model for calculation and processing, and the network recognition accuracy is calculated according to the abnormal type classification result output by the neural network model.
[0020] Furthermore, the neural network model takes the product neural network model as the basic framework, and is improved in combination with the residual module principle. The basic modules are composed of convolutional layer, maximum pooling layer, batch normalization layer and Relu layer. The overall network structure consists of an input module, an output module and a main module composed of the basic modules.
[0021] Furthermore, the overall network structure of the neural network model is based on a convolutional neural network and combines the structural principles of a residual network to learn the characteristics of the data. Through convolutional layers, pooling layers, and batch normalization layers, the normalization index of the target signal is calculated and the abnormal classification of the target signal is achieved through a classifier, thereby accurately identifying abnormal conditions in the distribution network monitoring data.
[0022] In summary, the present application includes at least one of the following beneficial technical effects:
[0023] 1. The photovoltaic power supply and communication anomaly identification method based on residual-convolutional network deep learning involved in this application learns the historical data features through deep learning, and constructs an efficient recognition model that can accurately identify data anomalies and further distinguish the specific types of anomalies. Experimental results show that compared with traditional deep learning methods, this method exhibits better detection accuracy, with an identification accuracy of up to 95% in common power system setting scenarios. In addition, the model also has anti-noise performance, providing a more solid guarantee for the stable operation of the distribution network. This method has significant advantages in processing complex data and improving the generalization ability of the model, especially in real-time monitoring and anomaly detection, providing a more solid guarantee for the stable operation of the distribution network;
[0024] 2. In this application, the convolution layer uses the convolution kernel to perform sliding convolution operations on the input feature data matrix, extracts data features through convolution calculation operations, obtains local feature data, and superimposes multiple convolution layers to extract high-level features from low-level features. The main function of the maximum pooling layer is to downsample, select the maximum value in the local data as the output, reduce the dimension of the feature data matrix and extract features, which is helpful for subsequent feature learning;
[0025] 3. The residual-convolution architecture of this application surpasses traditional methods in detection accuracy and noise resistance, ensuring high-accuracy recognition in complex environments and improving the stability and reliability of the power system. The method of this application supports real-time data monitoring and rapid anomaly detection, ensuring timely response to distribution network monitoring data anomalies, thereby effectively improving the stability and operation efficiency of the power system. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 It is a schematic diagram of the process of an embodiment of the present application;
[0027] Figure 2 It is a schematic diagram of the basic modules of the embodiment of the present application;
[0028] Figure 3 is a schematic diagram of a residual module in an embodiment of the present application;
[0029] Figure 4 It is a schematic diagram of the overall network structure of an embodiment of the present application. DETAILED DESCRIPTION
[0030] The following is combined with Figure 1 - Attachment Figure 4 , further details of this application are given.
[0031] Example
[0032] A photovoltaic power supply and communication anomaly identification method based on residual-convolutional network deep learning adopts the following technical solutions:
[0033] Anomaly identification methods include:
[0034] Data preprocessing, to reduce the noise interference of the original collected data, extract the characteristic data that can reflect the abnormal type, and obtain a data set with known data abnormality type;
[0035] Neural network training first establishes the network model parameter initialization, then performs forward propagation to calculate the error, then performs back propagation to calculate the gradient and update the model parameters, and then trains until the iteration termination condition ends;
[0036] Neural network testing, the test set calculates the model anomaly detection accuracy.
[0037] Data preprocessing includes the generation and preprocessing of data sets. First, the voltage, current, and power information of multiple nodes are collected and labeled. The data are then divided into training data sets and verification data sets in proportion and saved separately.
[0038] The data set includes signal feature information and data label information, and the signal feature information and data label information are processed into single-precision floating-point data types;
[0039] The Standard Scaler is used to normalize the data and scale each dimension of feature data to a standard normal distribution with a mean of 0 and a variance of 1, so that different features have the same scale and range to eliminate the differences in feature dimensions of different dimensions.
[0040] The two-dimensional vectors in the dataset are converted into Pytorch tensors, and their types are converted into floating-point types to represent continuous numerical features, so that continuous numerical features can be better represented and gradient propagation can be supported during the calculation process.
[0041] In neural network training, we first set the number of training rounds and learning rate parameters of the neural network, perform forward propagation on the training set data to calculate the error loss value, and then use the training error value to recalculate the training gradient according to back propagation and update the weight parameters of each layer of the neural network. The training ends when the iteration termination condition is reached.
[0042] The error loss value uses the cross entropy loss function as:
[0043]
[0044] Among them, L is the loss value, N is the number of samples, and y i is the true label of the i-th sample, is the predicted label of the i-th sample.
[0045] In the neural network test, the test set data is input into the trained neural network model for calculation and processing, and the network recognition accuracy is calculated based on the abnormal type classification results output by the neural network model.
[0046] The neural network model takes the product neural network model as the basic framework and is improved by combining the residual module principle. The basic modules are composed of convolution layer, maximum pooling layer, batch normalization layer and Relu layer. The overall network structure consists of an input module, an output module and a main module composed of basic modules.
[0047] The purpose of building a deep learning network model is to realize the recognition and detection of abnormal power signal conditions by training a high-quality classification model. The anomaly detection network model proposed in this paper takes the convolutional neural network model as the basic framework and combines it with the residual module principle for improvement. Figure 2 and Figure 3 The network consists of a basic module consisting of a convolutional layer, a maximum pooling layer, a batch normalization (BN) layer, and a Relu layer, and a residual module. The overall structure of the network consists of an input module, an output module, and a main module consisting of basic modules. The overall structure is as follows: Figure 4 shown.
[0048] In the basic module, the convolution layer uses the convolution kernel to perform sliding convolution operations on the input feature data matrix, extracts data features through convolution calculation operations, and obtains local feature data. Superimposing multiple convolution layers can extract high-level features from low-level features. The main function of the maximum pooling layer is to downsample, select the maximum value in the local data as the output, reduce the dimension of the feature data matrix and extract features, which is helpful for subsequent feature learning. The role of the BN layer is to normalize the upper layer input during the deep neural network training process, transform the input data into standard normal distribution data with a mean of 0 and a variance of 1, accelerate convergence, and improve the performance of the neural network. Relu is the most commonly used activation function, which allows the neural network to fit nonlinear functions according to the data set.
[0049] The overall network structure of the neural network model is based on the convolutional neural network and combines the structural principles of the residual network to learn the characteristics of the data. Through the convolution layer, pooling layer and batch normalization layer, the normalization index of the target signal is calculated and the abnormal classification of the target signal is realized through the classifier, so as to accurately identify the abnormal situation of the distribution network monitoring data.
[0050] In order to detect the performance of the model in this application, Simulink is used to build a model array, and various working conditions are simulated, and electrical data and environmental data under various working conditions are collected, including the IV characteristic curve obtained by scanning the model array under the corresponding working conditions and the corresponding illumination and irradiance, and the abnormal data in the original simulation data are removed, and the collected original IV curve is downsampled. The sample data is divided into a training set, a validation set, and a test set in equal proportion according to different working conditions. The designed residual-convolutional network structure is used to train on the training set, and the training model is verified on the validation set to obtain the optimal and most generalized training model. The obtained optimal model is further verified for its accuracy and generalization on the test set.
[0051] The implementation principle of the embodiment of the present application is as follows: data preprocessing, reducing the noise interference of the original collected data, extracting feature data that can reflect the abnormal type, and obtaining a data set of known data abnormality type; neural network training, first establishing the network model parameter initialization, then forward propagation to calculate the error, and then back propagation to calculate the gradient and update the model parameters, and then training until the iteration termination condition ends; neural network testing, the test set calculates the model abnormality detection accuracy.
[0052] The embodiments of this specific implementation are all preferred embodiments of the present application, and are not intended to limit the protection scope of the present application. The same components are represented by the same figure marks. Therefore, any equivalent changes made based on the structure, shape, and principle of the present application should be included in the protection scope of the present application.
Claims
1. A photovoltaic power supply and communication anomaly identification method based on residual-convolutional network deep learning, characterized in that: The abnormality identification method comprises: Data preprocessing, to reduce the noise interference of the original collected data, extract the characteristic data that can reflect the abnormal type, and obtain a data set with known data abnormality type; Neural network training first establishes the network model parameter initialization, then performs forward propagation to calculate the error, then performs back propagation to calculate the gradient and update the model parameters, and then trains until the iteration termination condition ends; Neural network testing, the test set calculates the model anomaly detection accuracy.
2. According to claim 1, a photovoltaic power source and communication anomaly identification method based on residual-convolutional network deep learning is characterized in that: The data preprocessing includes the generation and preprocessing of a data set. First, the voltage, current, and power information of multiple nodes are collected and labeled, and the data are divided into a training data set and a verification data set in proportion and then saved separately.
3. According to claim 2, a photovoltaic power source and communication anomaly identification method based on residual-convolutional network deep learning is characterized in that: The data set includes signal feature information and data label information, and the signal feature information and the data label information are processed into single-precision floating-point data types; The Standard Scaler is used to normalize the data and scale each dimension of feature data to a standard normal distribution with a mean of 0 and a variance of 1, so that different features have the same scale and range to eliminate the differences in feature dimensions of different dimensions.
4. According to claim 3, a photovoltaic power source and communication anomaly identification method based on residual-convolutional network deep learning is characterized in that: The two-dimensional vectors in the dataset are converted into Pytorch tensors and their types are converted into floating point types to represent continuous numerical features.
5. According to claim 3, a photovoltaic power source and communication anomaly identification method based on residual-convolutional network deep learning is characterized in that: In the neural network training, the number of training rounds and learning rate parameters of the neural network are first set, the training set data is forward propagated to calculate the error loss value, and then the training error value is recalculated according to the back propagation training gradient and the weight parameters of each layer of the neural network are updated, and the training is completed until the iteration termination condition ends.
6. According to claim 5, a photovoltaic power source and communication anomaly identification method based on residual-convolutional network deep learning is characterized in that: The error loss value uses the cross entropy loss function as: Among them, L is the loss value, N is the number of samples, and y i is the true label of the i-th sample, is the predicted label of the i-th sample.
7. According to claim 1, a photovoltaic power source and communication anomaly identification method based on residual-convolutional network deep learning is characterized in that: In the neural network test, the test set data is input into the trained neural network model for calculation and processing, and the network recognition accuracy is calculated according to the abnormal type classification result output by the neural network model.
8. The photovoltaic power source and communication anomaly identification method based on residual-convolutional network deep learning according to claim 7 is characterized in that: The neural network model uses the product neural network model as the basic framework and is improved by combining the residual module principle. The basic modules are composed of convolutional layer, maximum pooling layer, batch normalization layer and Relu layer. The overall network structure consists of an input module composed of basic modules, an output module and a main module.
9. The photovoltaic power source and communication anomaly identification method based on residual-convolutional network deep learning according to claim 8 is characterized in that: The overall network structure of the neural network model is based on a convolutional neural network and combines the structural principle of a residual network to learn the characteristics of the data. Through convolutional layers, pooling layers, and batch normalization layers, the normalization index of the target signal is calculated and the abnormal classification of the target signal is achieved through a classifier, so as to accurately identify abnormal conditions in the distribution network monitoring data.